Social physics in the age of artificial intelligence

So cial ph ysics in the age of artiﬁcial in telligence The Anh Han 1 ,⋆ , Jo el Z. Leibo 2 , T om Lenaerts 3 , 4 , 5 , Iy ad Rahw an 6 , F ernando San tos 7 , Matja ˇ z P erc 8 , 9 , 10 , 11 , ‡ , V alerio Capraro 12 , ‡ ,⋆ 1 Sc ho ol of Computing, Engineering and Digital T echnologies, T eesside Universit y , Middlesbrough, UK. 2 Go ogle DeepMind, London, UK 3 Mac hine Learning Group, Universit ´ e Libre de Bruxelles, Brussels, Belgium 4 AI Lab, V rije Universiteit Brussel, Brussels, Belgium 5 Cen ter for Human-Compatible AI, UC Berkeley , USA 6 Max Planc k Institute for Human Developmen t, Center for Humans & Mac hines, Berlin, Germany 7 Informatics Institute, Univ ersity of Amsterdam, Amsterdam, The Netherlands 8 F aculty of Natural Sciences and Mathematics, Univ ersity of Marib or, Marib or, Slo venia 9 Comm unit y Healthcare Center Dr. Adolf Drolc Maribor, Marib or, Slov enia 10 Departmen t of Physics, Kyung Hee Univ ersity , Seoul, Republic of Korea 11 Univ ersit y College, Korea Universit y , Seoul, Republic of Korea 12 Univ ersit y of Milan Bicco ca, Milan, Italy ‡ : Equal last author ⋆ Corresp onding: The Anh Han (t.han@tees.ac.uk), V alerio Capraro (v alerio.capraro@unimib.it) 1 2 Abstract Artiﬁcial intelligence (AI) systems are rapidly b ecoming more capable, autonomous, and deeply em b edded in so cial life. As h umans increasingly in teract, co op erate, and comp ete with AI, w e mo ve from purely human so cieties to hybrid h uman–AI so cieties whose collectiv e dynamics can- not b e captured by existing b ehavioural mo dels alone. Dra wing on ev olutionary game theory , cultural ev olution, and Large Language Models (LLMs) p ow ered simulations, we argue that these developmen ts op en a new re searc h agenda for so cial ph ysics centred on the co-evolution of h umans and machines. W e outline six key research directions. First, mo delling the ev olutionary dynamics of social behaviours (e.g. co op eration, fairness, trust) in h ybrid human–AI p opula- tions. Second, understanding machine culture: ho w AI systems generate, mediate, and select cultural traits. Third, analysing the co-evolution of language and b ehaviour when LLMs frame and participate in decisions. F ourth, studying the ev olution of AI delegation: how resp onsibili- ties and con trol are negotiated betw een h umans and machines. Fifth, formalising and comparing the distinct epistemic pip elines that generate h uman and AI b eha viour. Sixth, modelling the co-ev olution of AI developmen t and regulation in a strategic ecosystem of ﬁrms, users, and in- stitutions. T ogether, these directions deﬁne a programme for using so cial physics to an ticipate and steer the so cietal impact of adv anced AI. Keyw ords: So cial physics, artiﬁcial in telligence, ev olutionary game theory , human-AI interac- tions, language, delegation, machine culture, prosocial b eha viour, AI regulation. 3 Con ten ts 1 In tro duction 4 2 The rise of Artiﬁcial Intelligence 5 3 Ev olutionary dynamics of so cial b eha viours in hybrid human-AI systems 8 4 Mac hine culture: machine-generated and mediated cultural dynamics 9 5 The co-ev olution of language and b ehaviour 11 6 Ev olutionary dynamics of AI delegation 13 7 Ev olution of epistemic pip elines 16 8 Co-ev olution of AI dev elopment and regulation 18 9 Conclusions 20 4 1 In tro duction Artiﬁcial Intelligence (AI) systems are adv ancing rapidly and are b ecoming increasingly au- tonomous and em b edded in ev eryda y life [ 1 – 3 ]. AI promises to transform how w e address a wide range of so cietal challenges, from curing diseases and mitigating climate c hange to increasing pro ductivit y across m ultiple sectors [ 4 ]. Y et these b eneﬁts come with complex, large-scale risks and challenges, spanning economic, so cial, and existen tial concerns [ 5 – 9 ]. In contemporary so ciety , and likely even more so in the near future, humans and AI systems with diverse roles and capabilities co exist and in teract, giving rise to co-evolutionary dynamics that diﬀer mark edly from those observ ed in purely human settings [ 10 – 14 ]. These interactions can generate system-lev el eﬀects that are diﬃcult to an ticipate, mo del, and go vern, and het- erogeneit y in AI capabilities and b ehaviours further complicates prediction and the design of eﬀectiv e in terven tions [ 9 , 15 , 16 ]. Existing b ehavioural researc h, b oth theoretical and empirical, has predominantly focused on h uman-to-human in teractions, largely o v erlo oking the distinctiv e dynamics that arise when AI systems are in tegrated in to h ybrid h uman–AI so cieties [ 2 , 17 – 19 ]. Evolutionary game theory [ 20 ], cultural evolution [ 21 ], and generative AI based simulations [ 22 ] pro vide tractable frameworks for analysing strategic adaptation, feedbac k pro cesses, and p opulation-lev el outcomes in suc h settings. In this article, we argue that recent and forthcoming adv ances in AI deﬁne a new set of research questions that are of primary imp ortance for so cial ph ysics. W e propose six research directions w e believe are particularly imp ortant to pursue to un- derstand the impact of AI in the ev olution of so cial behaviour (T able 1 ). First, given the emergence and proliferation of hybrid human-AI so cieties [ 15 , 16 , 18 , 19 , 23 ] it is now crucial to revisit established domains of so cial behaviour (e.g. coop eration, fairness, and trust) [ 24 – 26 ] and fundamen tal p olitical questions of “ho w can we liv e together?” [ 27 – 29 ]. Second, as intelligen t mac hines b ecome in tegral to our lives, the study of cultural evolution m ust b e up dated to reﬂect AI’s inﬂuence on so cial b eha viour and emergent so cial norms [ 2 , 21 ]. Third, the recent adv ances in LLMs and generative AI suggest re-examining how these technologies inﬂuence the ev olution of language and b ehaviour [ 30 , 31 ]. F ourth, as h umans delegate more tasks and resp onsibilities to AI systems, there is a need to explore the evolution of suc h delegation b eha viours, understand- ing how humans and machines negotiate resp onsibilities and roles in decision-making pro cesses [ 10 , 11 , 32 – 35 ]. Fifth, b ecause humans and AI systems rely on fundamen tally diﬀerent epistemic pip elines [ 36 – 38 ], w e must inv estigate ho w selection acts not only on observ able b eha viour but also on the underlying cognitive and epistemic pro cesses that generate it in hybrid so cieties. Fi- nally , in order to eﬀectively manage AI risks amid rapid progress [ 5 ], it is crucial to understand the evolutionary dynamics of safe AI developmen t and and how regulation can b e strategically 5 applied to drive these dynamics to ward b eneﬁcial outcomes for so ciety [ 39 , 40 ]. 2 The rise of Artiﬁcial In telligence AI technologies v ary widely in their complexit y and capabilities, spanning a spectrum from traditional rule-based systems to sophisticated mo dels with agency and adaptive learning [ 5 , 41 , 42 ]. A t the simplest level are the traditional rule-based AI systems, which p erform sp eciﬁc tasks b y following predeﬁned rules and scripts (for example, some legacy customer-service b ots or men u-based virtual assistan ts). While limited in capabilities, they pro vide eﬃcien t and cost- eﬀectiv e alternativ es for promoting proso cial interactions [ 19 , 43 ]. A more adv anced class of systems comprises chatbots p ow ered by large language mo dels (LLMs), deep neural net w orks, whic h can dissect, extrap olate from and generate human-lik e text [ 44 ]. Diﬀeren t from the traditional rule-based systems, LLM-based c hatb ots are not conﬁned to ﬁxed responses; instead, they can engage in dynamic conv ersations, oﬀer insigh ts, capable of emulating some cogniti v e tasks, and can p erform a wide range of linguistic tasks. This capabilit y mak es them inv aluable to ols for applications suc h as customer service, conten t creation, education, and b ey ond. They can supp ort humans in in teractions with each other and comm unicate as well with other AI systems and adv anced to ols. In the context of game theory and p opulation dynamics, these systems allow formulate language-based games (more discussion on co-evolution of language and b ehaviour in Section 5 ). These AI mo dels can learn from interactions, becoming more eﬀectiv e o ver time b y adjusting their resp onses and b ehaviours based on new observ ations.. Giv en these capabilities, we envision several dimensions along which the rise of AI can impact ev olutionary dynamics in human-AI so cieties. First, unlik e humans, who t ypically learn from p ersonal exp erience and culturally trans- mitted norms, AI systems can pro cess v ast datasets rapidly and mo dify their b ehaviour at an unpreceden ted rate [ 45 ]. When such systems are embedded in so cial en vironmen ts, they par- ticipate in recipro cal in teractions, reputation building, and norm formation alongside humans, thereb y p oten tially altering the mechanisms that sustain coop eration, fairness, and other so- cial behaviours. Ev en small fractions of artiﬁcial agents can hav e disprop ortionate eﬀects on p opulation-lev el outcomes [ 46 ]. These considerations motiv ate a ﬁrst research direction, on the ev olutionary dynamics of so cial b ehaviours in hybrid human–AI systems, where heterogeneous h uman and AI agents co-adapt through feedbac k lo ops in b eha viour, expectations, and institu- tional constraints (Section 3 ). Second, AI systems increasingly participate in the generation, transmission, and selection of cultural con ten t [ 2 ]. Generativ e mo dels produce no vel artefacts and ideas, recommender 6 T able 1. Summary of research directions in so cial physics in the age of AI. Researc h Area Sp eciﬁc Question P otential Design Ev olutionary dynamics of so cial b eha viours in h ybrid h uman-AI systems (1) Ho w can w e mo del h umans’ exp ectations ab out AI behaviour in heterogeneous, le- gal, normative, and hierarchical h ybrid sys- tems? (2) Ho w can w e connect short-term b eha vioural exp eriments with long-term evo- lutionary dynamics in hybrid p opulations? Ev olutionary game and statistical- ph ysics mo dels (e.g. replicator dynam- ics, sto c hastic pro cesses) calibrated with exp eriment and ﬁeld data; hybrid h uman–AI exp erimen ts v arying AI fraction, transparency , and co ordination b et w een AI agen ts. Mac hine culture: mac hine- generated and medi- ated cultural dynamics (1) How do es algorithmic curation of con tent and social ties aﬀect cultural div ersity , norm formation, and explo- ration–exploitation trade-oﬀs? (2) How do h uman preferences and platform incentiv es shap e the ev olution of mac hine behaviour and AI-mediated culture? Agen t-based and net w ork mo dels of cul- tural transmission under diﬀeren t rec- ommender ob jectiv es; A/B tests on en- gagemen t vs diversit y; simulations with generativ e mo dels as cultural innov ators; analysis of human-in-the-loop training (e.g. RLHF) as cultural selection on AI. Co-ev olution of language and b e- ha viour (1) Ho w do es linguistic framing, when me- diated by LLMs, shap e human and AI deci- sions in strategic in teractions? (2) Ho w do language and so cial b ehaviour co-evolv e in h ybrid p opulations? Language-based game-theoretic mo dels; large-scale framing exp eriments; soci- eties of LLM p ow ered simulations; meta- analyses linking linguistic features to b e- ha viour. Ev olutionary dynamics of AI delega- tion (1) Do es delegation to AI agen ts increase or decrease prop ensity tow ards co op eration, honest y , etc.? (2) Ho w do incentiv es and AI agen t design impact long-term dynamics of co op eration, inequalit y , honest y norms, etc.? Sim ulations using simple v ersus LLM- based agen ts; Lab-based and online exp erimen ts in v olving humans and AI agen ts in dynamic, rep eated interaction. Ev olution of epistemic pip elines in h ybrid h uman–AI so cieties (1) Under what conditions does selec- tion fa vour agents with diﬀeren t epistemic pip elines (e.g. grounded vs text-only , causal vs correlational, cautious vs o verconﬁden t)? (2) When do hybrid systems con verge to epistemic vigilance vs equilibria dominated b y p ersuasiv e but ungrounded outputs? Extended evolutionary mo dels where t yp es enco de cognitiv e/epistemic param- eters (grounding, coun terfactual reason- ing, conﬁdence, absten tion); sim ulations of h ybrid human–AI p opulations; ex- p erimen ts v arying incentiv es for v eriﬁca- tion vs p ersuasion and measuring do wn- stream trust and misinformation. Co- ev olutionary dynamics of AI dev elop- men t and regulation (1) How do the strategic in teractions b e- t ween ﬁrms, users, and regulators shap e the tra jectory of AI capabilities, safet y in- v estments, and p o wer concentration? (2) Whic h regulatory and institutional interv en- tions can s teer this ecosystem tow ard safe and so cially b eneﬁcial outcomes? Game theory and population dynamics mo dels of AI labs, regulators, and users (arms races vs co op eration; open-source vs proprietary); scenario analysis of p ol- icy lev ers (standards, audits, windfall- sharing); experiments and LLM-based sim ulations of AI actors’ decisions. 7 systems curate and rewire so cial netw orks, and LLM-based assistan ts mediate knowledge trans- mission at scale. T ogether, these capabilities turn AI in to b oth a generator and a gatekeeper of culture, with profound implications for cultural div ersity , norm formation, and collectiv e exploration–exploitation trade-oﬀs. These issues are central to a second research direction on mac hine-generated and mac hine-mediated cultural dynamics (Section 4 ). Third, b ecause man y of the most widely deplo yed AI systems are language-based, they reshap e how decisions are framed, justiﬁed, and coordinated. LLMs b oth respond to and pro duce linguistic descriptions of decision problems, thereby inﬂuencing how humans p erceive strategic situations and whic h b ehaviours they regard as acceptable or moral. This tigh t coupling b et ween language and c hoice, mediated by AI, motiv ates a focus on the co-ev olution of language and b eha viour in hybrid p opulations (Section 5 ). F ourth, as AI systems b ecome more capable and useful, h umans ma y increasingly rely on them, facing decisions about what to delegate, to whic h system, and under whic h conditions [ 10 , 11 , 35 ]. Diﬀerences in autonom y , reliability , and transparency across AI systems create heterogeneous delegation patterns, feedbac k on trust and reliance, and new forms of strategic in teraction b etw een h umans and AI assistan ts. Understanding ho w suc h delegation decisions emerge and ev olv e is the fo cus of a fourth research direction, inv estigating the evolutionary dynamics of AI delegation (Section 6 ). Fifth, AI systems ma y mak e judgmen ts through potentially diﬀerent epistemic pip elines compared to h umans—from grounding and experience to motiv ation, causal reasoning, and error-monitoring. Behavioural similarity with h umans do es not necessarily en tail cognitive or epistemic equiv alence [ 36 , 37 ]. This calls for mo dels in whic h selection acts not only on observed b eha viour but also on underlying epistemic pro cesses, motiv ating a ﬁfth research direction on the evolution of epistemic pip elines in h ybrid h uman–AI so cieties (Section 7 ). Finally , AI systems are developed and deplo y ed within a strategic ecosystem of ﬁrms, go v- ernmen ts, users, and other stakeholders, whose incen tives can generate races, p o wer concen- tration, or co op eration around safety and standards. Because AI can b e scaled and modiﬁed rapidly , small changes in incen tiv es or regulation may hav e large, path-dep enden t eﬀects on this ecosystem. This motiv ates a ﬁnal researc h direction on the co-evolution of AI developmen t and regulation, using evolutionary and game-theoretic mo dels to identify interv en tions that steer AI progress tow ards so cially b eneﬁcial and safe outcomes (Section 8 ). 8 3 Ev olutionary dynamics of so cial b eha viours in hybrid h uman- AI systems Human-AI co-adaptation resem bles ev olutionary dynamics in biological systems. AI applications are trained on human-generated data, which in turn aﬀects human so cial b ehaviours and shap es future datasets. Such co-ev olving dynamics are evident in m ultiple applications: in recommender systems, users’ future interests are shap ed by current algorithmic recommendations [ 47 ]; future h uman behaviours can be inﬂuenced by so cially in teractive agents, themselv es adapting to matc h users’ p ersonalities and preferences [ 48 ]; in the con text of generative AI, incentiv es for humans to put eﬀort in generating high-quality data can change ov er time, aﬀecting the long-term v alue of data used to train future mo dels [ 49 ]; when b eing classiﬁed by machine learning algorithms, h umans can c hange their features, p ossibly leading to data shifts and motiv ating algorithms’ re- training [ 50 , 51 ]. The feedback loops characterizing h uman-AI systems result in so cial dynamics c hallenging to describ e, anticipate, and control [ 12 ]. T o ols used in evolutionary biology and statistical ph ysics can pla y a ma jor role to o v ercome this challenge, and inform the design of AI that improv es h uman so cial b ehaviours. Human decision-making processes and interaction con texts can be highly complex. Likewise, as describ ed in Section 2, artiﬁcial agen ts pow ered b y large mo dels are increasingly sophisticated. There is ho wev er exp erimental evidence that incredibly simple agen ts suﬃce to impact human b eha viour tow ards co op erative and co ordinated outcomes. It was sho wn that a small fraction of b ots with a random b ehaviour can signiﬁcan tly improv e human coordination on netw orks [ 52 ]. In rep eated co op eration dilemmas, it was also shown that resilien t co op erators (i.e., with a ﬁxed b ehaviour) can stabilise co op eration even among more self-interested humans who might otherwise defect [ 53 ]. Con venien tly , the ev olutionary dynamics in tro duced b y simple agen ts can be studied through ev olutionary dynamics and statistical physics metho ds, applying to ols such as the replicator equation or sto c hastic pro cess analysis [ 19 , 43 ]. Often, these highly stylized models abstract a wa y the complexity of interactions and focus instead of capturing mathematically the macro- scopic impact of in tro ducing simple hard-co ded agents in a p opulation of otherwise adaptiv e individuals. Although far from the complexity of curren t autonomous agents, considering hard- co ded agen ts in a p opulation of otherwise adaptive individuals pro vides an intuition for the profound impacts of AI agen ts in the social b ehaviour of h umans. In this regard, existing mo dels reveal that a small fraction of unconditionally co op erativ e agents suﬃce to stabilize co- op eration [ 54 ]. Similar mo dels also reveal that the adv antages of suc h unconditional co op erators are contingen t on adaptation properties suc h as intensit y of selection [ 18 ]. When reputations are av ailable, conditional agents steering co op eration tow ards well-reputed individuals can trig- 9 ger large-scale co op eration [ 55 ]. This mo delling approach has also b een applied to study so cial dynamics in dilemmas of co ordination [ 56 ], fairness [ 57 , 58 ] and collectiv e risk [ 46 , 59 ]. The previous works pro vide proliﬁc to ols to study evolutionary dynamics of so cial b ehaviour in rather simple systems. Applying suc h metho ds already informs us that small fractions of artiﬁcial agents, even with a ﬁxed strategy or simple b ehavioural rules, can ha ve a substan- tial eﬀect on h uman so cial behaviour. Nonetheless, real-world hybrid systems reveal features c hallenging to capture in current mo dels. First, real-w orld systems are large and highly heteroge- neous: h umans across diﬀerent so cieties v ary in how they p erceive artiﬁcial and adaptive agents [ 60 , 61 ]. Likewise, artiﬁcial agen ts can themselv es b e highly heterogeneous and v ary in shap e, comm unication mo dality or cognitiv e abilities. Ho w to capture such heterogeneity in a mo del? Second, while h umans and AI co-evolv e, so do legal and normative systems regulating AI. A k ey example is transparency requirements. AI applications can b e required to b ecome increasingly in terpretable, and diﬀerences in how interpretabilit y is implemen ted can greatly aﬀect humans’ exp ectations regarding artiﬁcial b ehaviour and ho w they adapt to it. Ho w to incorp orate, in a mo del, the ev olving legal and normative systems (i.e., system of la ws, rules, principles, or stan- dards) where human-AI interactions take place? Third, previous mo dels typically assume that artiﬁcial agen ts are indep enden t. Artiﬁcial agents migh t how ever rev eal co ordinated b ehaviours, b eing owned by sp eciﬁc companies or acting on b ehalf of interest groups with comp eting inter- ests (see section on Co-evolution of AI development and r e gulation ). This suggests hierarc hical mo dels, where interactions can o ccur at diﬀerent lev els: b etw een individual agen ts or b et ween groups and organizations. Finally , mo dels of social behaviour ev olution in h ybrid systems should b e informed — and inform — b ehavioural experiments, data collection and A/B testing [ 59 , 62 ]. T o adv ance our understanding of social behaviour dynamics in hybrid human-AI systems, w e believe it is fundamen tal to o v ercome the challenge of mo delling agen ts’ heterogeneity , in- teractions’ c omplexit y , evolving AI regulation, dynamics b etw een AI owners, and improving the metho ds to link theoretical mo delling and real-w orld data. Meeting these challenges b oils down to tac king t wo key questions: 1) How to mo del h umans’ exp ectations about AI behaviour in heterogeneous, legal, normative, and hierarc hical hybrid systems? 2) Ho w to connect short-term b eha vioural exp eriments with long-term evolutionary dynamics mo delling of hybrid systems? 4 Mac hine culture: mac hine-generated and mediated cultural dynamics In this section, we explore the inﬂuence of AI on s o cial b ehaviour through the lens of Culture Ev olution [ 21 ]. Historically , cultural ev olution has b een studied through frameworks that em- phasize the roles of human cognition, so cial learning, and communication. With the adv en t of 10 in telligent machines, these framew orks m ust b e re-examined to accommo date the impact that AI is having on the generation, transmission, and selection of cultural traits. One of the most signiﬁcant contributions of AI to cultural dynamics is its capacity to gen- erate nov el cultural artifacts and ideas. T raditional mo dels of cultural v ariation rely on h uman creativit y and the recombination of existing cultural elemen ts. Y et, generative AI systems ha v e b egun to operate as indep endent sources of inno v ation. As an example of com binatorial nov- elt y , text-to-image generativ e algorithms, such as DALL-E [ 63 ] and Stable Diﬀusion [ 64 ] enable artists and designers to rapidly explore no vel combinations of visual concepts at an unprece- den ted scale–e.g. generating hundreds of visual concepts for an a vocado-shap ed c hair in a matter of seconds. Moreov er, reinforcement learning systems, exempliﬁed b y AlphaGo, ha v e demon- strated the capacit y to dev elop strategies that are not only sup erior to human-deriv ed ones but also “alien” to human culture. The unexp ected and highly unconv entional mo ve 37 b y AlphaGo in its match against Lee Sedol exempliﬁes this phenomenon [ 65 ]. Bey ond generating new cultural artifacts, AI systems are also fundamentally altering the mec hanisms of cultural transmission. T raditionally , cultural information has b een passed from one individual to another through direct social in teractions, observ ation, and teac hing. How- ev er, AI has introduced new pathw ays for the dissemination of cultural knowledge, ﬁrst through recommender systems, and more recently through LLMs. AI-driv en recommender systems are reshaping the so cial netw orks that underlie cultural transmission. These systems, whic h suggest connections, con tent, and opp ortunities based on user b ehaviour [ 66 ], are subtly but signiﬁcantly rewiring the wa y individuals interact with one another and with cultural artifacts. By prioritiz- ing certain types of con tent and connections, recommender systems can inﬂuence so cial norms, amplify certain cultural trends, and ev en create new forms of so cial capital. This raises critical questions ab out the long-term implications of suc h rewiring: How do es the algorithmic curation of so cial netw orks aﬀect the diversit y of cultural expression and the dynamics of cultural trans- mission? And what are the tradeoﬀs that platform op erators face b etw een user engagement and diversit y of cultural transmission [ 67 ]? LLMs, trained on v ast corp ora of human text, now act as intermediaries in the transmission of kno wledge, serving as b oth reservoirs and conduits of cultural information (e.g. as teachers [ 68 ]). They facilitate the spread of ideas across indi- viduals and generations, often in wa ys that are more eﬃcient and far-reaching than traditional means. Ho wev er, this also means that the cultural conten t transmitted is increasingly shap ed b y the biases and structures inheren t in these AI systems. As a result, AI is not just a passiv e transmitter of culture but an active participan t that inﬂuences what information is emphasized, what is marginalized, and how cultural narratives are constructed. The selection pro cess in cultural evolution—the mechanism by which certain cultural traits b ecome more prev alent while others fade aw ay—will also b e deeply impacted b y AI. In the digital 11 age, the algorithms that curate con tent on platforms lik e so cial media, streaming services, and e-commerce sites hav e b ecome p o werful cultural gatek eep ers. These algorithms are designed to maximize user engagemen t, often b y selecting con ten t that aligns with users’ past b ehaviours and preferences. Ho wev er, in doing so, they pla y a signiﬁcan t role in shaping cultural landscap es, determining which ideas gain traction and whic h are relegated to obscurity [ 69 ]. The implications for cultural ev olution ma y b e profound [ 2 ]. Algorithms, by ﬁltering and prioritizing con tent, are not just resp onding to user preferences—they are actively shaping them. This is indeed the explicit goal of some recommendation systems, such as link recommendation algorithms [ 70 ]. These algorithms can th us alter the o verall structure of social net works, thus impacting the tradeoﬀs b etw een collective exploration and exploitation [ 71 , 72 ]. These algorithms ma y also create feedback lo ops where certain cultural traits are con tin uously reinforced, p otentially leading to the homogenization of culture or the entrenc hment of sp eciﬁc norms and v alues [ 73 ]. Another evolutionary selection pro cess is also taking place, in whic h humans select among AI algorithms. Thus, h uman preferences shap e the evolution of mac hine b ehaviour, e.g. b y explicitly training LLMs using human feedback [ 74 ], or by simply fav ouring certain commercial or op en source LLMs o ver others. Since these LLMs subsequently in teract with h umans at scale, this process shap es b oth h uman and mac hine cultural represen tations, but also leads to feedbac k lo ops [ 75 ]. Understanding these pro cesses is essential, in order to av oid degenerate phenomena lik e ’mo del collapse’ [ 76 , 77 ]. 5 The co-ev olution of language and b eha viour In the past tw o decades, a growing b o dy of exp erimen tal research with human participants has demonstrated the impact of linguistic con tent on decision-making. Essentially , the w ay in which the decision con text is describ ed can signiﬁcantly alter p eople’s choices (see [ 78 ], for a review). As LLM-based c hatb ots become increasingly embedded in everyda y activities, individuals are more likely to in teract with these systems directly or to make decisions with their assistance. Because these AI systems op erate primarily through language, the options they suggest—and, in some cases, the decisions they themselv es pro duce—dep end sensitiv ely on how a decision problem is formulated [ 79 ]. This feature underscores the growing imp ortance of understanding ho w linguistic framing shap es human b ehaviour, machine b ehaviour, and their interaction. W e argue that the in tegration of LLM-based chatbots in to h uman decision-making processes will reshap e this researc h area in at least t wo wa ys. First, it will create a heigh tened demand for formal, mathematical mo dels capable of capturing the eﬀects of linguistic framing on choice (e.g. [ 80 ]). Second, it will op en new empirical and practical a ven ues for studying how language and behaviour co-ev olve within human so cieties, particularly in contexts where human and 12 artiﬁcial agents jointly participate in decision-making. Regarding the ﬁrst p oint, existing research has largely concen trated on exp erimentally demon- strating the inﬂuence of language on human decision-making. F or instance, a seminal study b y Lib erman and colleagues [ 81 ] found that simply changing the label of a prisoner’s dilemma from “W all Street game” to “Communit y Game” signiﬁcantly increased the rate of co op eration among participan ts. Eriksson and co-authors [ (y ear?) ] conducted a series of experiments on the ultimatum game, rev ealing that the wa y rejection ac tions are phrased can inﬂuence resp onders’ b eha viour. In recen t years, sev eral studies hav e conceptually replicated and expanded these ﬁndings into sev eral domains, including the prisoner’s dilemma [ 83 , 84 ], the equity-eﬃciency trade-oﬀ game [ 85 – 87 ], the dictator game [ 88 – 90 ], corruption games [ 91 ], and market games [ 92 , 93 ]. In parallel, a smaller but growing theoretical literature has sought to formalize the role of language in strategic interaction, either through language-based games [ 94 – 96 ] or through the dev elopment of language-based utilit y functions [ 30 ]. Researc h on the behaviour of LLM-based c hatb ots is still in its early stages. An emergent line of literature has b egun inv estigating ho w chatbots mak e decisions in standard economic games [ 97 – 100 ]. A related line of w ork studies the behaviour of LLM chatbots in more open-ended strategic situations con vey ed b y simulations of collab orativ e storytelling proto cols (tabletop role-pla ying games) [ 101 , 102 ]. With resp ect to linguistic framing, existing evidence suggests that LLM-based c hatb ots often exhibit b eha vioural patterns that are qualitativ ely similar to those observed in h umans; how ev er, important quantitativ e diﬀerences remain. In particular, the relative frequencies with which diﬀeren t actions are selected can div erge substan tially from h uman b ehaviour [ 103 ]. T aken together, these ﬁndings p oin t to tw o clear gaps in the literature: a limited understanding of ho w linguistic framing systematically shap es the b ehaviour of LLM- based chatbots, and the absence of well-dev elop ed theoretical mo dels capable of capturing and explaining these eﬀects. Concerning the ﬁrst p oint, an intriguing direction for future research inv olves exploring how ev olutionary game theory can b e leveraged to gain a deep er understanding of how linguistic in teractions inﬂuence the ev olution of so cial behaviours and, vice v ersa, ho w the ev olution of so cial b ehaviour can inﬂuence the evolution of language. Ev olutionary game theory examines ho w strategies evolv e ov er time within p opulations. It has b een applied to study the evolution of co op erative b ehaviours [ 24 , 104 ] and other moral b eha viours, such as honesty [ 105 ] and trust [ 23 , 26 , 106 ]. Separately , it has also inv estigated the evolution of language [ 107 – 109 ]. How ev er, it is also plausible that language and b eha viour co-evolv e. Consider a scenario where t w o agents engage in a verbal in teraction b efore making a decision. During this conv ersation, they can inﬂuence each other’s p erception of the decision problem. As a result, their ﬁnal decision is not just a function of the a v ailable economic options but also of the preceding v erbal in teraction 13 [ 110 , 111 ]. This dynamic can be rep eated across m ultiple rounds, with the same or diﬀeren t pairs of agen ts, leading to a co-evolution of language and b ehaviour. Ev olutionary game theory could pro vide v aluable insigh ts in to ho w language and behaviour ha ve ev olved together in h uman so cieties. Regarding the second gap, LLM-based chatbots themselves ma y play a crucial role. An emerging line of literature shows that LLM-c hatb ots can b e a useful instrument to measure the sentimen t tenor of a piece of text [ 112 , 113 ]. Building on this approach, a recent meta- analysis of 61 dictator game exp erimen ts with human participants shows that sen timent scores generated by GPT-4 signiﬁcan tly explain av erage b eha vioural patterns. Sp eciﬁcally , b ehaviour can b e captured by an index that mathematically combines the sentimen t asso ciated with three piv otal actions: giving nothing, splitting the endowmen t equally , and giving all [ 30 ]. This line of researc h therefore illustrates how LLM-chatbots themselves can help formally incorp orating linguistic descriptions of decision contexts into utilit y functions, thereb y pro viding a tractable mathematical framework for linking language to b eha viour. While the study of the co-ev olution of language and behaviour is inherently fascinating and largely indep endent of AI, w e propose that AI, esp ecially LLM-based chatbots, can pla y a crucial role in inv estigating this issue. T o illustrate, w e conclude this section oﬀering one concrete example. Imagine a situation in which a p opulation of chatbots in teracts using v erbal comm unication b efore making decisions, sim ulating a p oten tial co-evolution of language and b eha viour. Since c hatb ots ev olve muc h faster than h umans, using them to sim ulate societies could allow us to quickly gain insigh ts into the co-evolutionary pro cesses that to ok millennia for h umans to develop. F or instance, in p ositive-sum games, language migh t stabilize as an eﬀective means for individuals to signal co op erativ e inten t. This could explain why moral narratives across cultures often emphasize co op erativ e b ehaviours [ 114 ]. In contrast, during zero-sum in teractions, language migh t b e less b eneﬁcial and could even b e seen as manipulative, with pla yers interpreting the same w ords diﬀerently , p ossibly due to conﬂicting interests. 6 Ev olutionary dynamics of AI delegation An asp ect fundamen tal to the organisation and success of a so ciet y is the mec hanism of delegat- ing control or execution of tasks to those that are more skilled or more kno wledgeable, allowing mem b ers to achiev e goals b eyond their p ersonal capabilities. This idea of a principal delegat- ing authorit y to an agent has been studied for many decades [ 115 – 117 ] and within a v ariet y of contexts, fo cussing mostly on answ ering questions on how to ensure alignment b etw een a principal and an agent when goals may diverge and information and risks ma y b e distributed unequally . With the adven t of highly p erforming AI systems, this line of research has gradually 14 b een expanding to the delegation of agency and decision-making authorit y to these systems [ 118 ], esp ecially since they now hav e the capacity to outp erform humans in diﬀerent tasks, even in an autonomous setting. This adv ancement in tro duces a new level of complexity in the dynamics of our so cietal organisation, requiring one to understand when and how h umans should delegate to AI, whether b eneﬁcial or detrimen tal outcomes are more likely , and what the long-term so cial and so cio-technical eﬀects of this delegation ma y b e. Diﬀeren t experimental as well as theoretical con tributions ha ve b een made, and the inter- est in this topic has b een rapidly increasing [ 119 – 122 ]. A very recent theoretical example was pro vided in the con text of participatory budgeting [ 123 ], whic h focussed on the question of whether artiﬁcial delegates that learn the voting b eha viour of participants could b e used to replace absen t v oters to ensure correct represen tation in perp etual v oting systems [ 124 , 125 ]. The authors demonstrated under a v ariet y of mo del parameters that artiﬁcial delegation en- sures representation across some key dimensions; inﬂuence is more equally distributed, minorit y v oices are satisﬁed more regularly , and group entitlemen ts are b etter resp ected. Crucially , the in tro duction of artiﬁcial delegates pro duces results that align with those under full turnout, th us ensuring artiﬁcial delegates rarely lead to outcomes which would not hav e won in their absence. This b eneﬁcial role of delegation to autonomous, while constrained, agents has b een demon- strated also in depth in a series of behavioural exp erimen ts. F or example, [ 126 ] show ed ho w p eople act more fairly in bargaining games when acting through a programmed agent. The same p ositiv e eﬀect on the level of coop eration w as rev ealed within the con text of a series of framed public go o ds game exp eriments. Similarly , [ 10 ] rep orted, within the con text of a collec- tiv e risk dilemma (CRD), where participan ts need to ac hieve a goal at the risk of incurring a loss if the goal is not achiev ed, that either selecting a pre-deﬁned agent or programming the agent’s decision-making b eha viour augments group success signiﬁcantly . These works demonstrate that delegation to autonomous agents seems to w ork as a form of b eha vioural commitmen t [ 127 ], whic h prev ents participants to deviate to another course of actions while the game is unfolding and thus minimising the eﬀect of emotional resp onses based on the observ e d b ehaviour of others [ 128 ]. Even when revision of the programmed strategy is p ossible, the participan ts maintain the am bition to reach the target and app ear to b e less inﬂuenced by a negativ e outcome in the ﬁrst round of the game compared to a h umans-only exp eriment [ 35 ]. An evolutionary game theory mo del diﬀerentiating b etw een human and delegated strategy dynamics, based on when and ho w decision/programming errors can b e made, provides a ten tative explanation for the impro ved success in the CRD [ 129 ]. Interestingly , the balance b etw een a triad of fair, comp ensating and recipro cating strategies [ 10 , 35 , 130 , 131 ] in the CRD app ears to determine when delegation is more successful or when delegation is more likely to b e adopted. In a hybrid scenario simulated in this mo del where b oth delegation and no-delegation are p ossible, delegation, sp eciﬁcally the 15 selection of a pre-deﬁned agent, emerges as most b eneﬁcial in the long run, conﬁrming observ a- tions made in [ 120 ]. Not withstanding these promising and p ositive results, sev eral issues undermine the use of AI systems as autonomous delegates. In the con text of the aforemen tioned CRD work, signiﬁcant inequalit y in the individual gains that each participan t receiv es at the end of the exp eriment w ere observ ed, which may explain why the p ost-exp eriment questionnaire revealed that most partic- ipan ts w ould prefer to pla y themselves. Such high degrees of inequality ma y hinder adoption of AI delegation as a solution, even when they could b e most useful to address risky decision- making problems. Additionally , sev eral works, even in the classic principal-agent literature, ha ve revealed that delegation may also lead to more selﬁsh and even c heating b ehaviour. [ 33 ] review ed more than 160 exp erimen tal studies with computer play ers and while one of his main conclusions was also that participants change their b ehaviour when in teracting with computer pla yers, i.e., they b ehav ed more rationally and selﬁshly , he also observed that if sub jects are a ware of the interaction with a computer play er, they may also learn to exploit them, which is more problematic. This issue of exploitation also recurs in work on h uman-AI interaction. F or example, [ 132 ] sho wed that human participan ts co op erated less with autonomous agents than with other h u- mans, an observ ation conﬁrmed in [ 133 ]. Moreov er, participan ts w ere k een to exploit their artiﬁcial coun terparts, who they assumed to b e as co op erative as h umans, or that it is accept- able to exploit them [ 134 ]. If rules require AI systems to b e co op erative for eac h individual user, their exploitation may b ecome the rule, undermining the m utual b eneﬁt that many p eople are coun ting on from human-AI in teractions. This eﬀect w as also observ ed in an evolutionary mo del wherein an adapting p opulation of agen ts interacts with a p opulation of pre-set probabilistic agen ts, which serv e as a proxy for AI agents; the individuals in the adapting p opulation reduced their coop e rativ eness if the autonomous agen ts w ere p erceived to b e suﬃcien tly co op erative [ 46 ]. In the con text of delegation, the risk of dishonest behaviours, both at the principal and agen t side, were extensiv ely analysed b y [ 11 ]. The authors sho wed through a series of exp eriments that delegation will increase the likelihoo d of dishonest b ehaviour, with humans preferring their agen ts to co vertly act in a dishonest manner when it is implicit in the agents’ design. Their expansion to LLM agents, which demonstrate a wider behavioural rep ertoire than simple rule- based systems, further supp orts this observ ation. With the increase in using artiﬁcial delegates, the amount of dishonest or unethical b ehaviour will most likely increase, esp ecially since AI systems will tend to comply with suc h requests due to the wa y they are curren tly designed to please the user. Dishonesty or deception from the side of LLM agents w as also found b y [ 135 ]. The authors show ed that LLM-based agen ts may induce false b eliefs in other agents, allowing the former to reap a b eneﬁt at cost to the other. Within the context of h uman-AI delegation, 16 or ev en AI-AI delegation, such scenarios are clearly problematic, p otentially generating distrust in such systems and hindering the deploymen t of AI agen ts in cases where they could really b e b eneﬁcial. Understanding thus the conditions under which deceptive b ehaviour evolv es and can b e counteracted will b e imp ortan t for the deploymen t of AI agents [ 136 ]. T ogether, these observ ations call for more in depth research on the ev olutionary dynamics of AI delegation, fo cusing on mechanisms, either implemented at design-time or as a w ay to con- trol agents’ b ehaviour at run-time, that i) alleviate the discussed problems that are inherent to h uman-AI delegation and ii) b o ost further the b eneﬁts of human-AI and AI-AI delegation. Evo- lutionary mo dels, incorp orating for instance insights obtained from the principal-agent frame- w ork, will pro vide understanding and allo w one to ask what-if questions to gauge what ma y w ork and what will not. These mo dels will need to b e v alidated through exp eriments, allo wing for the v eriﬁcation of mo del predictions and ev aluating the correctness of the underlying assumptions. In all cases, human non-rational b ehaviour will need to b e integrated in the studies to ensure that conclusions can supp ort real-world deploymen t. What remains clear is that adv anced AI systems will b ecome part of the diﬀerent so cietal organisations humans participate in, requiring th us rapid in v estments to a void problems b oth in the short and long term. 7 Ev olution of epistemic pip elines Muc h of the literature that applies evolutionary game theory to the study of so cial b ehaviour mo dels agents as c arriers of str ate gies , whose evolution is go verned b y the relativ e success of ob- serv able outputs. In these models, selection op erates on actions and pa y oﬀs, typically formalised through replicator dynamics, b est-resp onse dynamics, or related p opulation-lev el up date rules [ 20 , 137 – 139 ]. This output-cen tred approac h has pro v en highly successful in explaining the emergence of many so cial b ehaviours. How ev er, this tradition largely abstracts aw a y from the cognitiv e mechanisms that generate b eha viour. This abstraction ma y b ecome problematic in h y- brid human–AI p opulations, where iden tical outputs may arise from diﬀeren t in ternal pro cesses [ 36 – 38 ]. Human decisions ma y be regarded as emerging from a multi-stage epistemic pipeline that in tegrates grounding, parsing, exp erience, motiv ation, causality , metacognition, and v alue. Hu- mans p erceive and act in a world to which they are physically and so cially grounded, parse information through meaning-laden represen tations, accumulate ﬁrst-person exp erience with real consequences, reason causally ab out coun terfactuals, act under endogenous motiv ations and v alues, and monitor their own uncertaint y and epistemic limits. By contrast, the epistemic pip eline of con temp orary AI systems ma y diﬀer from h uman epistemic pipelines in some im- p ortan t wa ys. This is particularly eviden t in LLMs, where outputs are generated from largely 17 ungrounded linguistic represen tations and, at best, limited forms of m ultimo dal input—most notably vision and, o ccasionally , audio—while entirely lacking olfaction, proprio ception, intero- ception, vestibular sensing, and other b o dily mo dalities central to human cognition; statistical parsing is based on tok enization, which is blind to sp eaker inten tion, emotional tone, and situa- tional nuance, and merely maps character strings onto numerical indices; an y prior exp erience, when present, is still ungrounded, lac ks real stakes, and is not v alue-ric h, as it excludes harm, vulnerabilit y , and mortality; inference is primarily driv en by statistical correlations, and when LLMs engage in causal inference, their p erformance t ypically falls short of h uman reasoning [ 140 ], lac k of in trinsic motiv ation or v alue, and limited metacognitive access to uncertain ty [ 37 ]. These diﬀerences imply that b eha vioural equiv alence do es not entail epistemic or cognitiv e equiv alence, even when AI and humans app ear to act similarly in so cial settings [ 36 – 38 ]. F or this reason, it is increasingly imp ortant to dev elop ev olutionary models in which selection acts not only on ov ert b ehavioural phenot yp es, but also on the underlying cognitive pro cesses that generate them, consistent with the fact that biological evolution op erates on heritable mec hanisms rather than b eha viour p er se. One w ay to do so is to represent agen t t yp es by parameters that enco de distinct cognitive steps, suc h as access to grounded feedback versus text-only signals, the capacity for causal or counterfactual reasoning, or the ability to withhold resp onses under uncertaint y . These parameters ma y themselv es ev olv e or b e institutionally shap ed, particularly as AI systems are augmen ted with retriev al mechanisms [ 141 ], external to ols [ 142 ], or h uman feedback [ 74 ]; while such augmentations introduce functional analogues of h uman practices, they alter the decision pip eline without necessarily endowing the system with h uman-like capacity to ev aluate and revise b eliefs based on evidence. Previous researc h has leveraged evolutionary game theory to formalize sp eciﬁc cognitiv e mec hanisms underlying so cial b ehaviour. In particular, sev eral studies hav e examined how in- ten tion recognition [ 143 – 145 ], theory of mind [ 146 – 150 ], counterfactual reasoning [ 151 – 153 ], and in tuitive decision making [ 154 – 157 ] shap e co op erativ e b eha viour. These contributions demon- strate that evolutionary mo dels can fruitfully incorp orate cognitiv ely rich assumptions, moving b ey ond purely pay oﬀ-driven agen ts. How ever, this literature remains largely centred on the evo- lution of co op eration, leaving other so cially relev an t b ehaviours unexplored. Moreov er, existing mo dels t ypically assume homogeneous populations and do not accoun t for hybrid so cieties in whic h h umans in teract strategically with artiﬁcial agents endow ed with distinct cognitive archi- tectures. Explicitly formalising cognitiv e pro cesses in human–AI in teractions can generate phenomena that output-based mo dels cannot capture. Consider intuitiv e versus delib erative moral decision making: ev en when both processes pro duce the same immediate moral c hoice, they need not giv e rise to the same ev olutionary dynamics. Learning, generalization, and justiﬁcation often 18 dep end on the underlying cognitive pathw ay rather than on b ehaviour alone, b ecause up dating, abstraction, and explanation partly op erate o ver internal representations rather than only on o ver observed actions. So iden tical outputs can pro duce distinct feedback signals and up dat- ing pro cesses, leading to div ergen t tra jectories ov er time. Consequen tly , mo dels that collapse cognition into outputs risk o verlooking k ey sources of ev olutionary div ergence in h uman–AI in teractions. A second example concerns information sharing, credibilit y , and misinformation. Humans frequen tly rely on heuristics suc h as ﬂuency and expressed conﬁdence when ev aluating the re- liabilit y of information [ 158 – 160 ]. These heuristics are known to be exploitable: some h uman agen ts delib erately pro duce ﬂuen t and conﬁdent statemen ts despite lacking eviden tial supp ort, and such strategies can b e individually adv antageous. LLM-based chatbots can lik ewise gener- ate ﬂuen t and conﬁdent outputs ev en when factually false [ 161 , 162 ]. These mec hanisms can b e instan tiated at scale, with lo w marginal cost, and w eak coupling to veriﬁcation or reput a- tional consequences. In hybrid p opulations, this asymmetry can shift selection pressures to w ard agen ts—human or artiﬁcial—that prioritize p ersuasive linguistic output ov er costly veriﬁcation, fa vouring equilibria in which linguistic plausibilit y substitutes for epistemic reliabilit y , a con- dition recen tly describ ed as epistemia [ 36 , 37 ]. A question for future researc h is therefore to understand under whic h conditions—suc h as population composition, feedback latency , or in- cen tive structures—hybrid human–AI systems evolv e tow ard epistemic vigilance rather than to ward equilibria dominated b y p ersuasive but weakly grounded outputs. 8 Co-ev olution of AI dev elopmen t and regulation Eﬀectiv e regulation is crucial to ensure resp onsible and safe dev elopmen t while fostering user trust and adoption [ 163 , 164 ]. F aced with the swift and unpredictable ev olution of AI dev el- opmen t and coupled with accelerated inv estment in this technology , the urgency to “deep en our understanding of potential risks and iden tify actions to address them” [ 165 ] is especially emphasised, as highligh ted in the conclusions of the inaugural W orld Summit on AI. Ho w ever, despite sev eral prop osals on AI regulation aiming to inform strategies, there are little attempts to quantify the p otential impacts on near and long-term outcomes [ 5 , 7 , 166 – 169 ]. P opulation dynamics approac hes, such as ev olutionary game theory , ha ve found extensive application in examining decision-making dynamics during v arious catastrophic challenges like climate change [ 170 , 171 ] and nuclear war [ 172 ]. How ev er, the co-evolution of AI developmen t and regulation is fundamen tally diﬀeren t [ 40 , 173 ]. In disaster analyses related to climate c hange the cen tral fo cus is typically on participants’ reluctance to bear p ersonal costs for collectively desired outcomes, reﬂecting a shared, collective risk b orne b y all parties [ 171 ]. By con trast, 19 AI dev elopmen t often exhibits winner-tak es-most dynamics, in which successful actors gain substan tial relative adv an tages and face more individualised risk proﬁles [ 174 ]. The strategic landscap e of AI also diﬀers from nuclear arms races: for adv anced AI, many of the most severe risks ma y fall ﬁrst and most directly on its o wn developers and immediate users, whereas n uclear p o w ers are generally less exp osed to catastrophic harms originating unin ten tionally from their o wn arsenals [ 173 , 175 ]. This distinctive conﬁguration of incentiv es, b eneﬁts, and exp osure calls for tailored evolutionary and game-theoretic models of AI developmen t and regulation. There is an emerging b o dy of work that seek to mo del the strategic dynamics of AI develop- men t and impact of regulation [ 174 , 176 – 183 ]. Y et, notable gaps p ersist that need addressing. First, existing models largely focus on winner-tak es-all and AI arms race scenarios. Ho wev er, the tra jectory of AI developmen t is not inv ariably b ound by suc h patterns. Adv anced AI may instead supp ort multiple co existing solutions, e.g. the co-existence of op en-source and propri- etary foundation models (such as Llama-st yle and Mistral-st yle mo dels alongside commercial APIs), which sp ecialise in diﬀerent domains, pricing structures, and deplo ymen t settings rather than conv erging on a single dominant system [ 44 ]. Lik ewise, AI dev elopmen t need not follow arms race dynamics when institutional and market structures fav our co op eration ov er comp eti- tion—for instance, cross-lab consortia for shared safety b enchmarks and ev aluation platforms, join t red-teaming eﬀorts, or prop osals suc h as the Windfall Clause, which aims to share extraor- dinary AI proﬁts more broadly [ 184 ]. These settings naturally raise key research questions such as: how do shared safety infrastructures change ﬁrms’ incen tives to in vest in risky capabilities; and when do proﬁt-sharing or b eneﬁt-sharing schemes actually reduce comp etitive pressure in practice? Second, existing mo dels neglect the so cial con text surrounding AI dev elopment, rendering them incomplete. The dynamics and pace of AI progress are lik ely inﬂuenced b y the p erceived trust worthiness and risks that users attribute to deplo yed systems—for instance, whether con- sumers adopt AI decision recommendation in ﬁnance or healthcare, or whether organisations are willing to integrate foundation mo dels in to critical workﬂo ws [ 1 , 167 , 181 ]. As AI b ecomes more integrated into everyda y life, capturing this broader so cial con text—including users’ b e- ha viours, institutional incentiv es, and so cietal pressures—b ecomes essential [ 2 , 80 , 164 ]. F or example, user exp ectations and regulatory scrutiny can inﬂuence mo del design choices and dis- closure practices in adv anced AI [ 185 ]. These kinds of settings naturally motiv ate future researc h questions, including: under what conditions do users’ trust and adoption decisions slow do wn or accelerate risky AI races; ho w do public scandals or highly publicised failures reshape develop- ers’ incentiv es; and which regulatory or institutional arrangemen ts most eﬀectiv ely align ﬁrms’ comp etitiv e goals with users’ and so ciet y’s safet y concerns? 20 9 Conclusions The emergence of adv anced AI marks a turning point in the ev olution of our social systems. As hybrid human–AI societies take shap e, the collectiv e outcomes we observ e will dep end not only on tec hnical progress, but on ho w we design, deplo y , and gov ern these systems. The six researc h directions w e hav e outlined, spanning so cial b eha viours, machine culture, language, delegation, epistemic pip elines, and regulation, frame a social-physics research programme for understanding and steering this co-evolution. W e ha v e highlighted k ey research questions and design approaches for each direction (T able 1 ). By dev eloping computational mo dels and AI-p o wered simulations, integrating them with exp erimen ts and data, and embedding them in public debate, so cial ph ysics can help an ticipate emergen t risks, iden tify promising opp ortunities, and inform institutional designs that align AI progress with broadly shared so cial goals. The so oner we understand the evolutionary forces at w ork in hybrid h uman–AI so cieties, the b etter placed we will b e to steer these systems tow ards outcomes that are b oth b eneﬁcial and sustainable in the long run. Ac kno wledgemen t T.A.H. is supported b y EPSR C (Grant EP/Y00857X/1). T.L. gratefully ackno wledges the researc h supp ort by the F.R.S-FNRS (pro ject grant 40007793), the Service Public de W allonie Rec herche (gran t 2010235-ARIA C) b y DigitalW allonia4.ai and the Flemish Go vernmen t through the AI Research Program. M.P . is s upp orted b y the Slov enian Research and Innov ation Agency (Gran t P1-0403). References 1. P eter Andras, Luk as Esterle, Michael Guc kert, The Anh Han, P eter R Lewis, Kristina Milano vic, T erry P ayne, Cedric P erret, Jerem y Pitt, Simon T Po wers, et al. T rusting in telligent machines: Deep ening trust within so cio-technical systems. IEEE T e chnolo gy and So ciety Magazine , 37(4):76–83, 2018. 2. Levin Brinkmann, F abian Baumann, Jean-F ran¸ cois Bonnefon, Maxime Derex, Thomas F M ¨ uller, Anne-Marie Nussb erger, Agnieszk a Cz aplic k a, Alb erto Acerbi, Thomas L Grif- ﬁths, Joseph Henric h, et al. Mac hine culture. Natur e Human Behaviour , 7(11):1855–1868, 2023. 3. Iy ad Rahw an, Manuel Cebrian, Nic k Obradovic h, Josh Bongard, Jean-F ran¸ cois Bon- 21 nefon, Cynthia Breazeal, Jacob W Crandall, Nic holas A Christakis, Iain D Couzin, Matthew O Jackson, et al. Machine b ehaviour. Natur e , 568(7753):477–486, 2019. 4. Nestor Maslej, Loredana F attorini, Ra ymond P errault, Y olanda Gil, V anessa Parli, Njenga Kariuki, Emily Capstick, Ank a Reuel, Erik Brynjolfsson, John Etchemendy , et al. Artiﬁcial intelligence index rep ort 2025. arXiv pr eprint arXiv:2504.07139 , 2025. 5. Y oshua Bengio, Geoﬀrey Hinton, Andrew Y ao, Dawn Song, Pieter Abb eel, T revor Dar- rell, Y uv al Noah Harari, Y a-Qin Zhang, Lan Xue, Shai Shalev-Shw artz, et al. Managing extreme ai risks amid rapid progress. Scienc e , 384(6698):842–845, 2024. 6. Scott McLean, Gemma JM Read, Jason Thompson, Chris Bab er, Neville A Stan ton, and P aul M Salmon. The risks asso ciated with artiﬁcial general intelligence: A systematic review. Journal of Exp erimental & The or etic al A rtiﬁcial Intel ligenc e , 35(5):649–663, 2023. 7. V alerio Capraro, Austin Lentsc h, Daron Acemoglu, Selin Akgun, Aisel Akhmedov a, En- nio Bilancini, Jean-F ran¸ cois Bonnefon, P ablo Bra˜ nas-Garza, Luigi Butera, Karen M Douglas, et al. The impact of generativ e artiﬁcial intelligence on so cio economic inequal- ities and p olicy making. PNAS nexus , 3(6), 2024. 8. Lewis Hammond, Alan Chan, Jesse Clifton, Jason Hoelscher-Obermaier, Akbir Khan, Euan McLean, Chandler Smith, W olfram Barfuss, Jak ob F o erster, T om´ a ˇ s Ga ven ˇ ciak, et al. Multi-agen t risks from adv anced ai. arXiv pr eprint arXiv:2502.14143 , 2025. 9. Y oshua Bengio et al. International AI Safety Rep ort 2026. T echnical Rep ort DSIT 2026/001, DSIT, 2026. Accessed: 2026-02-03. 10. Elias F ern´ andez Domingos, Inˆ es T errucha, R´ emi Suchon, Jelena Gruji ´ c, Juan C Bur- guillo, F rancisco C Santos, and T om Lenaerts. Delegation to artiﬁcial agents fosters proso cial b ehaviors in the collectiv e risk dilemma. Scientiﬁc r ep orts , 12(1):8492, 2022. 11. Nils K¨ obis, Zo e Rah wan, Raluca Rilla, Bramant yo Ibrahim Supriy atno, Clara Bersc h, T amer Aja j, Jean-F ran¸ cois Bonnefon, and Iy ad Rahw an. Delegation to artiﬁcial in telli- gence can increase dishonest b eha viour. Natur e , 646(8083):126–134, 2025. 12. Dino Pedresc hi, Luca Pappalardo, Emanuele F erragina, Ricardo Baeza-Y ates, Alb ert- L´ aszl´ o Barab´ asi, F rank Dign um, Virginia Dign um, Tina Eliassi-Rad, F osca Giannotti, J´ anos Kert´ esz, et al. Human-ai co evolution. Artiﬁcial Intel ligenc e , 339:104244, 2025. 13. Jean-F ran¸ cois Bonnefon, Iy ad Rahw an, and Azim Shariﬀ. The moral psyc hology of artiﬁcial intelligence. A nnual r eview of psycholo gy , 75(1):653–675, 2024. 22 14. C ´ esar A Hidalgo, Diana Orghian, Jordi Albo Canals, Filipa De Almeida, and Natalia Martin. How humans judge machines . MIT Press, 2021. 15. Nenad T omasev, Matija F ranklin, Jo el Z Leib o, Julian Jacobs, William A Cunning- ham, Iason Gabriel, and Simon Osindero. Virtual agent economies. arXiv pr eprint arXiv:2509.10147 , 2025. 16. Gillian K Hadﬁeld and Andrew Koh. An economy of ai agents. arXiv pr eprint arXiv:2509.01063 , 2025. 17. Allan Dafo e, Y oram Bac hrach, Gillian Hadﬁeld, Eric Horvitz, Kate Larson, and Thore Graep el. Co op erativ e ai: mac hines m ust learn to ﬁnd common ground. 2021. 18. Filipp o Zimmaro, Man uel Miranda, Jos´ e Mar ´ ıa Ramos F ern´ andez, Jes ´ us A Moreno L´ op ez, Max Reddel, V aleria Widler, Alb erto Antonioni, and The Anh Han. Emergence of co op eration in the one-shot prisoner’s dilemma through discriminatory and samaritan ais. Journal of the R oyal So ciety Interfac e , 21(218):20240212, 2024. 19. F ernando P San tos. Proso cial dynamics in multiagen t systems. AI Magazine , 45(1):131– 138, 2024. 20. Josef Hofbauer and Karl Sigm und. Evolutionary games and p opulation dynamics . Cam- bridge universit y press, 1998. 21. Alex Mesoudi. Cultural evolution: a review of theory , ﬁndings and contro v ersies. Evo- lutionary biolo gy , 43:481–497, 2016. 22. Yik ang Lu, Alb erto Aleta, Chunpeng Du, Lei Shi, and Y amir Moreno. Llms and genera- tiv e agent-based mo dels for complex systems research. Physics of Life R eviews , 51:283– 293, 2024. 23. Cedric Perret, The Anh Han, Elias F ern´ andez Domingos, Theo dor Cimp ean u, and Si- mon T. P ow ers. Disen tangling trust from coop eration: Evolution of trust as reduced monitoring in so cial dilemmas. Chaos, Solitons & F r actals , 208:118130, 2026. 24. Matja ˇ z P erc, Jillian J Jordan, Da vid G Rand, Zhen W ang, Stefano Boccaletti, and Attila Szolnoki. Statistical physics of human co op eration. Physics R ep orts , 687:1–51, 2017. 25. Martin A No wak. Fiv e rules for the evolution of co op eration. scienc e , 314(5805):1560– 1563, 2006. 23 26. Aanjaney a Kumar, V alerio Capraro, and Matjaˇ z Perc. The evolution of trust and trust- w orthiness. Journal of the R oyal So ciety Interfac e , 17(169):20200491, 2020. 27. Allan Dafo e, Edward Hughes, Y oram Bachrac h, T antum Collins, Kevin R McKee, Jo el Z Leib o, Kate Lars on, and Thore Graepel. Op en problems in co op erative ai. arXiv pr eprint arXiv:2012.08630 , 2020. 28. Iason Gabriel. Artiﬁcial intelligence, v alues, and alignmen t. Minds and machines , 30(3):411–437, 2020. 29. Jo el Z Leib o, Alexander Sasha V ezhnevets, William A Cunningham, S ´ ebastien Krier, Manfred Diaz, and Simon Osindero. So cietal and tec hnological progress as sewing an ever-gro wing, ever-c hanging, patch y , and p olychrome quilt. arXiv pr eprint arXiv:2505.05197 , 2025. 30. V alerio Capraro, Rob erto Di P aolo, Matjaˇ z Perc, and V eronica Pizziol. Language-based game theory in the age of artiﬁcial in telligence. Journal of the R oyal So ciety Interfac e , 21(212):20230720, 2024. 31. F ederico Battiston, V alerio Capraro, F ariba Karimi, Sune Lehmann, Andrea Bam b erg Migliano, Onk ar Sadek ar, Angel S´ anchez, and Matjaˇ z P erc. Higher-order in teractions shap e collective human b ehaviour. Natur e Human Behaviour , pages 1–17, 2025. 32. Celso M de Melo, Stacy Marsella, and Jonathan Gratc h. Human co op eration when acting through autonomous machines. Pr o c e e dings of the National A c ademy of Scienc es , 116(9):3482–3487, 2019. 33. Christoph Marc h. Strategic interactions b etw een h umans and artiﬁcial intelligence: Lessons from exp eriments with computer pla yers. Journal of Ec onomic Psycholo gy , 87:102426, 2021. 34. Iason Gabriel, Arianna Manzini, Geoﬀ Keeling, Lisa Anne Hendricks, V erena Rieser, Hasan Iqbal, Nenad T oma ˇ sev, Ira Ktena, Zac hary Kenton, Mikel Ro driguez, et al. The ethics of adv anced ai assistants. arXiv pr eprint arXiv:2404.16244 , 2024. 35. In ˆ es T erruc ha, Elias F ern´ andez Domingos, R ´ emi Suc hon, F rancisco C San tos, Pieter Simo ens, and T om Lenaerts. Humans program artiﬁcial delegates to accurately solv e collectiv e-risk dilemmas but lac k precision. Pr o c e e dings of the National A c ademy of Scienc es , 122(25):e2319942121, 2025. 24 36. Edoardo Loru, Jacop o Nudo, Niccol` o Di Marco, Alessandro San tiro cc hi, Rob erto A tzeni, Matteo Cinelli, Vincenzo Cestari, Clelia Rossi-Arnaud, and W alter Quattro cio cchi. The sim ulation of judgment in llms. Pr o c e e dings of the National A c ademy of Scienc es , 122(42):e2518443122, 2025. 37. W alter Quattrocio cc hi, V alerio Capraro, and Matjaˇ z P erc. Epistemological fault lines b et w een h uman and artiﬁcial in telligence. arXiv pr eprint arXiv:2512.19466 , 2025. 38. Matja ˇ z Perc. Coun terfeit judgments in large language mo dels. Pr o c e e dings of the National A c ademy of Scienc es , 122(48):e2528527122, 2025. 39. The Anh Han, Luis Moniz Pereira, et al. T o Regulate or Not: A So cial Dynamics Analysis of an Idealised AI Race . Journal of A rtiﬁcial Intel ligenc e R ese ar ch , 69:881– 921, Nov em b er 2020. 40. The Anh Han, Lu ´ ıs Moniz Pereira, and T om Lenaerts. Mo delling and inﬂuencing the ai bidding war: a researc h agenda. In Pr o c e e dings of the 2019 AAAI/ACM Confer enc e on AI, Ethics, and So ciety , pages 5–11, 2019. 41. Ra y Perrault and Jack Clark. Artiﬁcial intelligence index rep ort 2024. 2024. 42. Nestor Maslej, Loredana F attorini, et al. The AI Index 2023 Ann ual Rep ort, 2023. 43. Ch unjiang Mu, Hao Guo, Y ang Chen, Chen Shen, Die Hu, Shuyue Hu, and Zhen W ang. Multi-agen t, h uman–agent and b eyond: a survey on co op eration in so cial dilemmas. Neur o c omputing , 610:128514, 2024. 44. Y up eng Chang, Xu W ang, Jindong W ang, Y uan W u, Linyi Y ang, Kaijie Zhu, Hao Chen, Xiao yuan Yi, Cunxiang W ang, Yidong W ang, et al. A surv ey on ev aluation of large language mo dels. ACM tr ansactions on intel ligent systems and te chnolo gy , 15(3):1–45, 2024. 45. Thomas L Griﬃths. Understanding h uman intelligence through human limitations. T r ends in Co gnitive Scienc es , 24(11):873–883, 2020. 46. In ˆ es T erruc ha, Elias F ern´ andez Domingos, F rancisco C. San tos, Pieter Simo ens, and T om Lenaerts. The art of compensation: How hybrid teams solv e collectiv e-risk dilemmas. PloS one , 19(2):e0297213, 2024. 47. Jingh ua Piao, Jiazhen Liu, F ang Zhang, Jun Su, and Y ong Li. Human–ai adaptiv e dynamics drives the emergence of information co co ons. Natur e Machine Intel ligenc e , 5(11):1214–1224, 2023. 25 48. Birgit Lugrin. In tro duction to so cially in teractive agents. In The handb o ok on so cial ly inter active agents: 20 ye ars of r ese ar ch on emb o die d c onversational agents, intel ligent virtual agents, and so cial r ob otics volume 1: metho ds, b ehavior, c o gnition , pages 1–20. 2021. 49. Saﬀron Huang and Divy a Siddarth. Generative ai and the digital commons. arXiv pr eprint arXiv:2303.11074 , 2023. 50. Moritz Hardt, Nimro d Megiddo, Christos Papadimitriou, and Mary W o otters. Strategic classiﬁcation. In Pr o c e e dings of the 2016 ACM Confer enc e on Innovations in The or etic al Computer Scienc e , pages 111–122, 2016. 51. Marta C Couto, Fla via Barsotti, and F ernando P San tos. Collective dynamics of strategic classiﬁcation. arXiv pr eprint arXiv:2508.09340 , 2025. 52. Hirok azu Shirado and Nicholas A Christakis. Lo cally noisy autonomous agents improv e global human co ordination in netw ork exp erimen ts. Natur e , 545(7654):370–374, 2017. 53. Andrew Mao, Lili Dworkin, Siddharth Suri, and Duncan J W atts. Resilien t coop era- tors stabilize long-run coop eration in the ﬁnitely rep eated prisoner’s dilemma. Natur e c ommunic ations , 8(1):13800, 2017. 54. Gopal Sharma, Hao Guo, Chen Shen, and Jun T animoto. Small b ots, big impact: solv- ing the con undrum of co op eration in optional prisoner’s dilemma game through simple strategies. Journal of The R oyal So ciety Interfac e , 20(204):20230301, 2023. 55. Alexandre S Pires and F ernando P Santos. Artiﬁcial agen ts mitigate the punishmen t dilemma of indirect recipro cit y . In Pr o c e e dings of the 24th International Confer enc e on A utonomous A gents and Multiagent Systems , pages 1650–1659, 2025. 56. Hao Guo, Chen Shen, Shuyue Hu, Junliang Xing, Pin T ao, Y uanch un Shi, and Zhen W ang. F acilitating co op eration in human-agen t h ybrid p opulations through autonomous agen ts. iScienc e , 26(11), 2023. 57. F ernando P San tos, Jorge M P acheco, Ana P aiv a, and F rancisco C Santos. Ev olution of collective fairness in h ybrid p opulations of humans and agents. In Pr o c e e dings of the AAAI c onfer enc e on artiﬁcial intel ligenc e , v olume 33, pages 6146–6153, 2019. 58. Zhao Song, Theo dor Cimp eanu, Chen Shen, and The Anh Han. Evolution of fairness in h ybrid p opulations with sp ecialised ai agen ts. arXiv pr eprint arXiv:2602.18498 , 2026. 26 59. F ernando P San tos, Sam uel Mascarenhas, F rancisco C San tos, Filipa Correia, Sam uel Gomes, and Ana P aiv a. Pic ky losers and carefree winners prev ail in collectiv e risk dilemmas with partner selection. Autonomous A gents and Multi-A gent Systems , 34(2):40, 2020. 60. Kevin R McKee, Xuec hunzi Bai, and Susan T Fiske. Humans perceive warm th and comp etence in artiﬁcial intelligence. Iscienc e , 26(8), 2023. 61. Stevie Bergman, Nahema Marchal, John Mellor, Shakir Mohamed, Iason Gabriel, and William Isaac. Stela: a communit y-centred approach to norm elicitation for ai alignmen t. Scientiﬁc R ep orts , 14(1):6616, 2024. 62. Isab el Neto, Alexandre S Pires, Filipa Correia, and F ernando P San tos. Co op eration through indirect reciprocity in child-robot in teractions. arXiv pr eprint arXiv:2512.20621 , 2025. 63. Khari Johnson. Op enai debuts dall-e for generating images from text. V entur eBe at , 2021. 64. Robin Rombac h, Andreas Blattmann, Dominik Lorenz, P atrick Esser, and Bj¨ orn Om- mer. High-resolution image syn thesis with latent diﬀusion mo dels. In Pr o c e e dings of the IEEE/CVF c onfer enc e on c omputer vision and p attern r e c o gnition , pages 10684–10695, 2022. 65. Jo el Z Leib o, Edward Hughes, Marc Lanctot, and Thore Graep el. Auto curricula and the emergence of innov ation from so cial interaction: A manifesto for multi-agen t intelligence researc h. arXiv pr eprint arXiv:1903.00742 , 2019. 66. Zhep eng Li, Xiao F ang, and Olivia R Liu Sheng. A survey of link recommendation for so cial netw orks: Metho ds, theoretical foundations, and future research directions. ACM T r ansactions on Management Information Systems (TMIS) , 9(1):1–26, 2017. 67. F abian B aumann, Daniel Halp ern, Ariel D Procaccia, Iy ad Rah w an, Itai Shapira, and Man uel W ¨ uthrich. Optimal engagemen t-diversit y tradeoﬀs in so cial media. In Pr o c e e d- ings of the A CM on Web Confer enc e 2024 , pages 288–299, 2024. 68. Amanda J Lucas, Michael Kings, Devi Whittle, Emma Dav ey , F rancesca Happ´ e, Chris- tine A Caldw ell, and Alex Thornton. The v alue of teaching increases with to ol complexity in cumulativ e cultural ev olution. Pr o c e e dings of the R oyal So ciety B , 287(1939):20201885, 2020. 27 69. Renee DiResta. Invisible rulers: The p e ople who turn lies into r e ality . PublicAﬀairs, 2024. 70. Jessica Su, Aneesh Sharma, and Sharad Goel. The eﬀect of recommendations on net w ork structure. In Pr o c e e dings of the 25th international c onfer enc e on World Wide Web , pages 1157–1167, 2016. 71. Da vid Lazer and Allan F riedman. The net work structure of exploration and exploitation. A dministr ative scienc e quarterly , 52(4):667–694, 2007. 72. Win ter Mason and Duncan J W atts. Collaborative learning in netw orks. Pr o c e e dings of the National A c ademy of Scienc es , 109(3):764–769, 2012. 73. Inna Kizhner, Melissa T erras, Maxim Rum yan tsev, V alen tina Khokhlo v a, Elisa veta Demeshk ov a, Iv an Rudov, and Julia Afanasiev a. Digital cultural colonialism: measuring bias in aggregated digitized conten t held in go ogle arts and culture. Digital Scholarship in the Humanities , 36(3):607–640, 2021. 74. Long Ouy ang, Jeﬀrey W u, Xu Jiang, Diogo Almeida, Carroll W ainwrigh t, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ra y , et al. T rain- ing language mo dels to follo w instructions with h uman feedback. A dvanc es in neur al information pr o c essing systems , 35:27730–27744, 2022. 75. V eniamin V eselo vsky , Mano el Horta Rib eiro, and Rob ert W est. Artiﬁcial artiﬁcial arti- ﬁcial in telligence: Crowd work ers widely use large language mo dels for text production tasks. arXiv pr eprint arXiv:2306.07899 , 2023. 76. Ilia Shumailo v, Zakhar Shuma ylov, Yiren Zhao, Y arin Gal, Nicolas P ap ernot, and Ross Anderson. The curse of recursion: T raining on generated data makes mo dels forget. arXiv pr eprint arXiv:2305.17493 , 2023. 77. Edgar A Du´ e ˜ nez-Guzm´ an, Suzanne Sadedin, Jane X W ang, Kevin R McKee, and Jo el Z Leib o. A so cial path to human-lik e artiﬁcial in te lligence. Natur e Machine Intel ligenc e , 5(11):1181–1188, 2023. 78. V alerio Capraro, Joseph Y Halp ern, and Matjaˇ z Perc. F rom outcome-based to language- based preferences. Journal of Ec onomic Liter atur e , 62(1):115–154, 2024. 79. V eniamin V eselovsky , Berke Argin, Benedikt Stroebl, Chris W endler, Rob ert W est, James Ev ans, Thomas L Griﬃths, and Arvind Nara yanan. Lo calized cultural kno wledge is conserv ed and controllable in large language mo dels. arXiv pr eprint arXiv:2504.10191 , 2025. 28 80. Jo el Z Leib o, Alexander Sasha V ezhnevets, Manfred Diaz, John P Agapiou, William A Cunningham, P eter Sunehag, Julia Haas, Raphael Koster, Edgar A Du ´ e ˜ nez-Guzm´ an, William S Isaac, Georgios Piliouras, Stan M Bilesc hi, Iy ad Rah wan, and Simon Osindero. A theory of appropriateness with applications to generative artiﬁcial intelligence. arXiv pr eprint arXiv:2412.19010 , 2024. 81. V arda Lib erman, Steven M Samuels, and Lee Ross. The name of the game: Predictive p o w er of reputations versus situational lab els in determining prisoner’s dilemma game mo ves. Personality and so cial psycholo gy bul letin , 30(9):1175–1185, 2004. 82. Kimmo Eriksson, P ontus Strimling, Per A Andersson, and T orun Lindholm. Costly punishmen t in the ultimatum game evok es moral concern, in particular when framed as pa yoﬀ reduction. Journal of Exp erimental So cial Psycholo gy , 69:59–64, 2017. 83. Christoph Engel and Da vid G Rand. What do es “clean” really mean? the implicit framing of decontextualized exp eriments. Ec onomics L etters , 122(3):386–389, 2014. 84. Laura Mieth, Axel Buchner, and Raoul Bell. Moral lab els increase coop eration and costly punishmen t in a prisoner’s dilemma game with punishment option. Scientiﬁc R ep orts , 11(1):10221, 2021. 85. V alerio Capraro and David G Rand. Do the right thing: Experimental evidence that preferences for moral b eha vior, rather than equit y or eﬃciency p er se, driv e h uman proso cialit y . Judgment and De cision Making , 13(1):99–111, 2018. 86. Long Huang, W ansheng Lei, F uming Xu, Liang Y u, and F ujun Shi. Choosing an equi- table or eﬃcien t option: A distribution dilemma. So cial Behavior and Personality: an international journal , 47(10):1–10, 2019. 87. Long Huang, W ansheng Lei, F uming Xu, Hairong Liu, Liang Y u, F ujun Shi, and Lei W ang. Maxims nudge equitable or eﬃcien t choices in a trade-oﬀ game. PloS one , 15(6):e0235443, 2020. 88. V alerio Capraro and Andrea V anzo. The p o wer of moral words: Loaded language gen- erates framing eﬀects in the extreme dictator game. Judgment and De cision Making , 14(3):309–317, 2019. 89. Daphne Chang, Roy Chen, and Erin Krupk a. Rhetoric matters: A so cial norms ex- planation for the anomaly of framing. Games and Ec onomic Behavior , 116:158–178, 2019. 29 90. Jin yi Kuang and Cristina Bicc hieri. Language matters: ho w normativ e expressions shap e norm p erception and aﬀect norm compliance. Philosophic al T r ansactions of the R oyal So ciety B , 379(1897):20230037, 2024. 91. Karolina Aleksandra ´ Sciga la, Ingo Zettler, Stefan Pfattheicher, and V alerio Capraro. Corrupting the proso cial p eople: do es co op eration framing increase brib ery engagement among proso cial individuals? stage 1 registered report. 2022. 92. F rancisca Jim´ enez-Jim ´ enez and Ja vier Ro dero-Cosano. Conditioning comp etitive b e- ha viour in exp erimental bertrand markets through contextual frames. Journal of Behav- ior al and Exp erimental Ec onomics , 103:101987, 2023. 93. Ingela Alger and Jos´ e Ignacio Rivero-Wildemau w e. Doing the right thing (or not) in a lemons-lik e situation: on the role of so cial preferences and k antian moral concerns. arXiv pr eprint arXiv:2405.13186 , 2024. 94. Adam Bjorndahl, Joseph Y Halpern, and Rafael Pass. Language-based games. arXiv pr eprint arXiv:1310.6408 , 2013. 95. Adam Bjorndahl and Joseph Y Halp ern. Language-based decisions. arXiv pr eprint arXiv:2106.11494 , 2021. 96. Adam Bjorndahl and Joseph Y Halp ern. Sequen tial language-based decisions. arXiv pr eprint arXiv:2307.07563 , 2023. 97. Yiting Chen, T racy Xiao Liu, Y ou Shan, and Songfa Zhong. The emergence of economic rationalit y of gpt. Pr o c e e dings of the National A c ademy of Scienc es , 120(51):e2316205120, 2023. 98. Danica Dillion, Nik et T andon, Y uling Gu, and Kurt Gra y . Can ai language models replace human participants? T r ends in Co gnitive Scienc es , 27(7):597–600, 2023. 99. John J Horton. Large language mo dels as sim ulated economic agen ts: What can w e learn from homo silicus? T echnical rep ort, National Bureau of Economic Research, 2023. 100. Qiaozh u Mei, Y utong Xie, W alter Y uan, and Matthew O Jac kson. A turing test of whether ai chatbots are b ehaviorally similar to humans. Pr o c e e dings of the National A c ademy of Scienc es , 121(9):e2313925121, 2024. 101. Alexander Sasha V ezhnevets, John P Agapiou, Avia Aharon, Ron Ziv, Jayd Mat yas, Edgar A Du´ e˜ nez-Guzm´ an, William A Cunningham, Simon Osindero, Dann y Karmon, 30 and Jo el Z Leib o. Generative agent-based mo deling with actions grounded in ph ysical, so cial, or digital space using concordia. arXiv pr eprint arXiv:2312.03664 , 2023. 102. Chandler Smith, Marwa Ab dulhai, Manfred Diaz, Mark o T esic, Rakshit S T riv edi, Alexander Sasha V ezhnevets, Lewis Hammond, Jesse Clifton, Minsuk Chang, Edgar A Du ´ e ˜ nez-Guzm´ an, et al. Ev aluating generalization capabilities of llm-based agents in mixed-motiv e scenarios using concordia. In NeurIPS 2024 Comp etition T r ack , 2025. 103. V alerio Capraro, Roberto Di Paolo, and V eronica Pizziol. A publicly a v ailable benchmark for assessing large language mo dels’ abilit y to predict ho w humans balance self-in terest and the interest of others. Scientiﬁc R ep orts , 15(1):21428, 2025. 104. The Anh Han, Zhao Song, Theo dor Cimpeanu, Manh Hong Duong, Marcus Krellner, V alerio Capraro, and Matjaz P erc. Co op eration v ersus social w elfare. Physics of Life R eviews , 56:33–60, 2026. 105. V alerio Capraro, Matjaˇ z Perc, and Daniele Vilone. The evolution of lying in well-mixed p opulations. Journal of the R oyal So ciety Interfac e , 16(156):20190211, 2019. 106. The Anh Han, Cedric Perret, and Simon T Po w ers. When to (or not to) trust intelligen t mac hines: Insigh ts from an ev olutionary game theory analysis of trust in rep eated games. Co gnitive Systems R ese ar ch , 68:111–124, 2021. 107. Andrea Puglisi, Andrea Baronchelli, and Vittorio Loreto. Cultural route to the emergence of linguistic categories. Pr o c e e dings of the National A c ademy of Scienc es , 105(23):7936– 7940, 2008. 108. Andrea Baronchelli, Maddalena F elici, Vittorio Loreto, Emanuele Caglioti, and Luc Steels. Sharp transition to wards shared vocabularies in m ulti-agent systems. Journal of Statistic al Me chanics: The ory and Exp eriment , 2006(06):P06014, 2006. 109. Kalo y an Danovski and Markus Brede. On the ev olutionary language game in structured and adaptive p opulations. Plos one , 17(8):e0273608, 2022. 110. Robin Clark. Me aningful games: Exploring language with game the ory . MIT Press, 2023. 111. Margit Gaﬀal and Jes ´ us Padilla G´ alvez. Negotiation, game theory and language games. In Dynamics of R ational Ne gotiation: Game The ory, L anguage Games and F orms of Life , pages 11–40. Springer, 2024. 31 112. Stev e Rathje, Dan-Mircea Mirea, Ilia Suc holutsky , Ra ja Marjieh, Claire E Robertson, and Ja y J V an Bav el. Gpt is an eﬀectiv e to ol for m ultilingual psychological text analysis. Pr o c e e dings of the National A c ademy of Scienc es , 121(34):e2308950121, 2024. 113. Stefan F euerriegel, Ab durahman Maarouf, Dominik B¨ ar, Dominique Geissler, Jonas Sc hw eisthal, Nicolas Pr¨ ollo chs, Claire E Rob ertson, Stev e Rathje, Jo c hen Hartmann, Saif M Mohammad, et al. Using natural language processing to analyse text data in b eha vioural science. Natur e R eviews Psycholo gy , 4(2):96–111, 2025. 114. Oliv er Scott Curry , Daniel Austin Mullins, and Harv ey Whitehouse. Is it go o d to coop er- ate? testing the theory of moralit y-as-co op eration in 60 so cieties. Curr ent anthr op olo gy , 60(1):47–69, 2019. 115. Mic hael C Jensen and William H Meckling. Theory of the ﬁrm: Managerial b ehavior, agency costs and ownership structure. Journal of Financial Ec onomics , 3(4):305–360, 1976. 116. Susan P Shapiro. Agency theory . Annu. R ev. So ciol. , 31(1):263–284, 2005. 117. Sean Gailmard et al. Accountabilit y and principal-agent theory . The Oxfor d handb o ok of public ac c ountability , pages 90–105, 2014. 118. T ankiso Moloi and Tshilidzi Marwala. The agency theory . In Artiﬁcial Intel ligenc e in Ec onomics and Financ e The ories , pages 95–102. Springer, 2020. 119. Brian Lubars and Chenhao T an. Ask not what ai can do, but what ai should do: T o wards a framework of task delegability . A dvanc es in neur al information pr o c essing systems , 32, 2019. 120. Cindy Candrian and Anne Sc herer. Rise of the mac hines: Delegating decisions to au- tonomous ai. Computers in Human Behavior , 134:107308, 2022. 121. P atric k Hemmer, Monik a W estphal, Max Sc hemmer, Sebastian V etter, Michael V¨ ossing, and Gerhard Satzger. Human-ai collab oration: the eﬀect of ai delegation on human task p erformance and task satisfaction. In Pr o c e e dings of the 28th international c onfer enc e on intel ligent user interfac es , pages 453–463, 2023. 122. Ala Y ankousk a ya, Mohamed Basel Almourad, Magn us Liebherr, F ahad Bey ahi, Guan- dong Xu, and Raian Ali. Who lets ai take o v er? cross-national v ariation in willingness to delegate socially imp ortant roles to artiﬁcial in telligence. AI & SOCIETY , pages 1–19, 2026. 32 123. Apurv a Shah, Axel Abels, Ann No w ´ e, and T om Lenaerts. Artiﬁcial delegates resolv e fair- ness issues in p erp etual v oting with partial turnout. In Pr o c e e dings of the ACM Col le ctive Intel ligenc e Confer enc e , CI ’25, page 71–82, New Y ork, NY, USA, 2025. Asso ciation for Computing Machinery . 124. Martin Lac kner. Perpetual v oting: F airness in long-term decision making. In Pr o c e e dings of the AAAI c onfer enc e on artiﬁcial intel ligenc e , v olume 34, pages 2103–2110, 2020. 125. Martin Lac kner and Jan Maly . Prop ortional decisions in perp etual v oting. In Pr o c e e dings of the AAAI Confer enc e on A rtiﬁcial Intel ligenc e , volume 37, pages 5722–5729, 2023. 126. Celso M de Melo, Stacy Marsella, and Jonathan Gratch. So cial decisions and fairness c hange when p eople’s in terests are represen ted by autonomous agen ts. Autonomous A gents and Multi-A gent Systems , 32(1):163–187, 2018. 127. Isab elle Bro cas, Juan D Carrillo, and Mathias Dewatripont. Commitment devices under self-con trol problems: An o verview. The Psycholo gy of e c onomic de cisions , 2:49–67, 2004. 128. Rob ert H. F rank. Passions Within Re ason: The Str ate gic Role of the Emotions . Norton and Company , 1988. 129. In ˆ es T erruc ha, Elias F ern´ andez Domingos, Pieter Simo ens, and T om Lenaerts. Commit- ting to the wrong artiﬁcial delegate in a collective-risk dilemma is b etter than directly committing mistakes. Scientiﬁc r ep orts , 14(1):10460, 2024. 130. Elias F ern´ andez Domingos, Jelena Gruji ´ c, Juan C Burguillo, Georg Kirc hsteiger, F ran- cisco C Santos, and T om Lenaerts. Timing uncertaint y in collectiv e risk dilemmas en- courages group recipro cation and p olarization. Iscienc e , 23(12), 2020. 131. Elias F ern´ andez Domingos, Jelena Gruji ´ c, Juan C Burguillo, F rancisco C Santos, and T om Lenaerts. Mo deling behavioral exp eriments on uncertaint y and coop eration with p opulation-based reinforcement learning. Simulation Mo del ling Pr actic e and The ory , 109:102299, 2021. 132. Jacob W Crandall, Ma yada Oudah, T ennom, F atimah Ishow o-Oloko, Sherief Ab dallah, Jean-F ran¸ cois Bonnefon, Manuel Cebrian, Azim Shariﬀ, Mic hael A Go o dric h, and Iyad Rah wan. Co op erating with mac hines. Natur e c ommunic ations , 9(1):233, 2018. 133. F atimah Isho wo-Olok o, Jean-F ran¸ cois Bonnefon, Zak ariy ah Soro y e, Jacob Crandall, Iyad Rah wan, and T alal Rahw an. Behavioural evidence for a transparency–eﬃciency tradeoﬀ in human–mac hine co op eration. Natur e Machine Intel ligenc e , 1(11):517–521, 2019. 33 134. Jurgis Karpus, Adrian Kr¨ uger, Julia T o v ar V erba, Bahador Bahrami, and Ophelia Dero y . Algorithm exploitation: Humans are k een to exploit b enevolen t ai. Iscienc e , 24(6), 2021. 135. Thilo Hagendorﬀ. Deception abilities emerged in large language mo dels. Pr o c e e dings of the National A c ademy of Scienc es , 121(24):e2317967121, 2024. 136. S ¸ tefan Sark adi, Alex Rutherford, P eter McBurney , Simon P arsons, and Iy ad Rah wan. The evolution of deception. R oyal So ciety op en scienc e , 8(9), 2021. 137. John Maynard Smith. Ev olution and the theory of games. In Did Darwin get it right? Essays on games, sex and evolution , pages 202–215. Springer, 1982. 138. J¨ orgen W W eibull. Evolutionary game the ory . MIT press, 1997. 139. William H Sandholm. Population games and evolutionary dynamics . MIT press, 2010. 140. Haoang Chi, He Li, W enjing Y ang, F eng Liu, Long Lan, Xiaoguang Ren, T ongliang Liu, and Bo Han. Un v eiling causal reasoning in large language mo dels: Realit y or mirage? A dvanc es in Neur al Information Pr o c essing Systems , 37:96640–96670, 2024. 141. P atric k Lewis, Ethan Perez, Aleksandra Piktus, F abio P etroni, Vladimir Karpukhin, Naman Go yal, Heinrich K ¨ uttler, Mike Lewis, W en-tau Yih, Tim Ro ckt¨ aschel, et al. Retriev al-augmen ted generation for kno wledge-intensiv e nlp tasks. A dvanc es in neur al information pr o c essing systems , 33:9459–9474, 2020. 142. Timo Sc hick, Jane Dwivedi-Y u, Rob erto Dess ` ı, Rob erta Railean u, Maria Lomeli, Eric Ham bro, Luk e Zettlemoy er, Nicola Cancedda, and Thomas Scialom. T oolformer: Lan- guage mo dels can teach themselves to use to ols. A dvanc es in Neur al Information Pr o- c essing Systems , 36:68539–68551, 2023. 143. The Anh Han, Lu ´ ıs Moniz Pereira, and F rancisco C Santos. Inten tion recognition pro- motes the emergence of co op eration. A daptive Behavior , 19(4):264–279, 2011. 144. The Anh Han, Luis Moniz Pereira, and F rancisco C San tos. Corpus-based inten tion recognition in co op eration dilemmas. Artiﬁcial Life , 18(4):365–383, 2012. 145. The Anh Han, F rancisco C Santos, T om Lenaerts, and Luis Moniz Pereira. Synergy b et w een in ten tion recognition and commitments in coop eration dilemmas. Scientiﬁc r ep orts , 5(1):9312, 2015. 146. Dale O Stahl. Evolution of smartn play ers. Games and Ec onomic Behavior , 5(4):604– 617, 1993. 34 147. Marie Dev aine, Guillaume Hollard, and Jean Daunizeau. Theory of Mind: Did Evolution F o ol Us? PL oS ONE , 9(2):e87619, F ebruary 2014. 148. T om Lenaerts, Marco Sap onara, Jorge M Pac heco, and F rancisco C San tos. Evolution of a theory of mind. Iscienc e , 27(2), 2024. 149. Jiabin W u. Minds and co op eration. R ationality and So ciety , 37(4):537–552, 2025. 150. Marco Sap onara, Elias F ern´ andez Domingos, Jorge M P ac heco, and T om Lenaerts. Evo- lution fa vours p ositively biased reasoning in sequen tial interactions with high future gains. Journal of the R oyal So ciety Interfac e , 22(229):20250153, 2025. 151. Lu ´ ıs Moniz Pereira and F rancisco C Santos. Counterfactual thinking in co op eration dynamics. In International c onfer enc e on Mo del-Base d R e asoning , pages 69–82. Springer, 2018. 152. Marta C Couto, Stefano Giaimo, and Christian Hilb e. Introspection dynamics: a sim- ple mo del of coun terfactual learning in asymmetric games. New Journal of Physics , 24(6):063010, 2022. 153. An t´ onio M F ernandes, F rancisco C San tos, and Ana P aiv a. Coun terfactual thinking in sto c hastic dynamics of co op eration. In ECAI 2024 , pages 3493–3500. IOS Press, 2024. 154. Adam Bear and Da vid G Rand. Intuition, deliberation, and the ev olution of coop eration. Pr o c e e dings of the National A c ademy of Scienc es , 113(4):936–941, 2016. 155. Adam Bear, Ari Kagan, and David G Rand. Co-ev olution of co op eration and cognition: the impact of imp erfect delib eration and con text-sensitiv e in tuition. Pr o c e e dings of the R oyal So ciety B: Biolo gic al Scienc es , 284(1851):20162326, 2017. 156. Stephan Jagau and Matthijs v an V eelen. A general evolutionary framework for the role of intuition and delib eration in co op eration. Natur e Human Behaviour , 1(8):0152, 2017. 157. Eladio Montero-P orras, T om Lenaerts, Riccardo Gallotti, and Jelena Grujic. F ast delib- eration is related to unconditional behaviour in iterated prisoners’ dilemma exp eriments. Scientiﬁc R ep orts , 12(1):20287, 2022. 158. Ian Ma ynard Begg, Ann Anas, and Suzanne F arinacci. Disso ciation of processes in b elief: Source recollection, statement familiarity , and the illusion of truth. Journal of Exp erimental Psycholo gy: Gener al , 121(4):446, 1992. 35 159. Alice Dech ˆ ene, Christoph Stahl, Jo chim Hansen, and Mic haela W¨ ank e. The truth ab out the truth: A meta-analytic review of the truth eﬀect. Personality and So cial Psycholo gy R eview , 14(2):238–257, 2010. 160. P aul C Price and Eric R Stone. In tuitive ev aluation of likelihoo d judgment pro ducers: Evidence for a conﬁdence heuristic. Journal of Behavior al De cision Making , 17(1):39–57, 2004. 161. Adam T auman Kalai and San tosh S V empala. Calibrated language mo dels must hallu- cinate. In Pr o c e e dings of the 56th A nnual A CM Symp osium on The ory of Computing , pages 160–171, 2024. 162. Ziw ei Xu, Sanjay Jain, and Mohan Kank anhalli. Hallucination is inevitable: An innate limitation of large language mo dels. arXiv pr eprint arXiv:2401.11817 , 2024. 163. Mic hael K Cohen, Noam Kolt, Y oshua Bengio, Gillian K Hadﬁeld, and Stuart Russell. Regulating adv anced artiﬁcial agents. Scienc e , 384(6691):36–38, 2024. 164. Simon T P ow ers, Olena Linn yk, et al. The Stuﬀ W e Swim in: Regulation Alone Will Not Lead to Justiﬁable T rust in AI. IEEE T e chnolo gy and So ciety Magazine , 42(4):95–106, 2023. 165. AI Safet y Summit. The bletchley declaration by countries attending the ai safet y summit, 1–2 nov em b er 2023, Nov em b er 2023. Accessed: 02/11/2026. 166. Gillian K. Hadﬁeld and Jack Clark. Regulatory Markets: The F uture of AI Go vernance, April 2023. 167. Baobao Zhang, Markus Anderljung, Lauren Kahn, No emi Dreksler, Mic hael C Horowitz, and Allan Dafoe. Ethics and gov ernance of artiﬁcial in telligence: Evidence from a surv ey of machine learning researchers. Journal of Artiﬁcial Intel ligenc e R ese ar ch , 71:591–666, 2021. 168. Hu w Rob erts, Josh Cowls, Emmie Hine, Jessica Morley , Vincen t W ang, Mariarosaria T addeo, and Luciano Floridi. Go verning artiﬁcial intelligence in china and the europ ean union: Comparing aims and promoting ethical outcomes. The Information So ciety , 39(2):79–97, 2023. 169. Virginia Dign um, Catherine R´ egis, Kerstin Bach, Andr´ e PLF de Carv alho, Ginevra Castellano, F rank Dign um, Elizab eth F arries, F osca Giannotti, Natali Helb erger, Isadora Hellegren, et al. Roadmap for ai p olicy research: Ai p olicy research summit, sto ckholm, no vem b er 2024, 2025. 36 170. Manfred Milinski, Ralf D Sommerfeld, Hans-J ¨ urgen Kram b ec k, Floyd A Reed, and Jo c hem Marotzk e. The collective-risk so cial dilemma and the prev ention of simu- lated dangerous climate c hange. Pr o c e e dings of the National A c ademy of Scienc es , 105(7):2291–2294, 2008. 171. F rancisco C Santos and Jorge M Pac heco. Risk of collective failure provides an escap e from the tragedy of the commons. Pr o c e e dings of the National A c ademy of Scienc es , 108(26):10421–10425, 2011. 172. Sandeep Baliga and T omas Sj¨ ostr¨ om. Arms races and negotiations. The R eview of Ec onomic Studies , 71(2):351–369, 2004. 173. Dennis Pamlin and Stuart Armstrong. Global challenges: 12 risks that threaten h uman civilization. Glob al Chal lenges F oundation, Sto ckholm , 2015. 174. Stuart Armstrong, Nic k Bostrom, et al. Racing to the Precipice: A Mo del of Artiﬁcial In telligence Dev elopment . Ai & So ciety , 31(2):201–206, Ma y 2016. 175. T oby Ord. The pr e cipic e: Existential risk and the futur e of humanity . Hac hette Bo oks, 2020. 176. The Anh Han, Lu ´ ıs Moniz Pereira, et al. Mediating Artiﬁcial Intelligence Dev elopments through Negative and Positiv e Incentiv es . Plos One , 16(1), January 2021. 177. Theo dor Cimp eanu, F rancisco Santos, et al. Artiﬁcial Intelligence Developmen t Races in Heterogeneous Settings. Scientiﬁc R ep orts , 12(1):1723, 2022. 178. The Anh Han, T om Lenaerts, et al. V oluntary Safety Commitments Provide an Escap e from Over-Regulation in AI Developmen t . T e chnolo gy in So ciety , 68:101843, 2022. 179. Nic holas Emery-Xu, Andrew P ark, and Rob ert T rager. Uncertain ty , information, and risk in international tec hnology races. Journal of Conﬂict R esolution , page 00220027231214996, 2023. 180. P aolo Bo v a, Alessandro Di Stefano, and The Anh Han. Both eyes op en: Vigilan t in- cen tives help auditors improv e ai safety . Journal of Physics: Complexity , 5(2):025009, 2024. 181. Zainab Alala wi, Paolo Bo v a, Theo dor Cimp eanu, Alessandro Di Stefano, Manh Hong Duong, Elias F ern´ andez Domingos, The Anh Han, Marcus Krellner, Ndidi Bianca Ogb o, Simon T. P o wers, and Filipp o Zimmaro. T rust ai regulation? discerning users are 37 vital to build trust and eﬀective ai regulation. Applie d Mathematics and Computation , 508:129627, 2026. 182. Nataliy a Balabano v a, Adeela Bashir, Paolo Bov a, Alessio Buscemi, Theo dor Cimp ean u, Henrique Correia da F onseca, Alessandro Di Stefano, Manh Hong Duong, Elias F er- nandez Domingos, An tonio F ernandes, et al. Media and resp onsible ai go v ernance: a game-theoretic and llm analysis. arXiv pr eprint arXiv:2503.09858 , 2025. 183. Henrique Correia Da F onse ca, Ant´ onio F ernandes, Zhao Song, Theo dor Cimp eanu, Na- taliy a Balabanov a, Adeela Bashir, Paolo Bov a, Alessio Buscemi, Alessandro Di Stefano, Manh Hong Duong, et al. Can media act as a soft regulator of safe ai dev elopment? a game theoretical analysis. In Artiﬁcial Life Confer enc e Pr o c e e dings 37 , volume 2025, page 90. MIT Press, 2025. 184. Cullen O’Keefe, Peter Cihon, Ben Garﬁnk el, Carrick Flynn, Jade Leung, and Allan Dafo e. The windfall clause: Distributing the b eneﬁts of ai for the common goo d. In Pr o c e e dings of the AAAI/A CM Confer enc e on AI, Ethics, and So ciety , pages 327–331, 2020. 185. Xiao Zhan, Yifan Xu, and Stefan Sark adi. Deceptive ai ecosystems: The case of c hatgpt. In Pr o c e e dings of the 5th International Confer enc e on Conversational User Interfac es , pages 1–6, 2023.

Social physics in the age of artificial intelligence

Original Paper

Comments & Academic Discussion

Leave a Comment

Original Paper

Related Papers

Comments & Academic Discussion

Leave a Comment