The Nonverbal Gap: Toward Affective Computer Vision for Safer and More Equitable Online Dating

1 The Nonverb al Gap: Toward Aff ective Compu ter Vision f or Safer and More Equitable Online Dating Ratna Kandala 1 University of Kansas ratnanirupama@gmail .com Niva Manchanda 2 University of Kansa s nmanchanda@ku.edu Akshata Kishore Moharir 3 University of Maryland akshatankishore5@gm ail.com Abs tr ac t Online dating has become the dominant way roman tic re - lationships begin, yet current platforms strip the n onver - bal cu es: gaze, fa cial expression, body posture, response timing, that humans rely on to signal comfort, disinter- est, and consent, creating a communication gap with dis - proportiona te safety consequences for women. We argue that this g ap represents b oth a techn ical opp ortunity a nd a mo ral responsib ility for the co mputer vision commu nity, which has develo ped th e a ffective tools, fa cial action unit detection, gaze estimation, engagement modeling, and mul- timodal affect recognition, needed to begin addressing it, yet ha s largely ignored the d ating d omain as a research context. We pro pose a fa irness-first research agenda or- ganized around four capability areas: real -time discomfort detection, en gagement asymmetry mod eling between p art - ner s, consent -aw are inte ra cti on des ign , and longi tud ina l in - teraction summarization, each g rounded in established CV methodolo gy and motivated by the social psychology of ro - mantic commu nication. We argu e that responsib le pursuit of this a genda requires purpose-built datasets collected un - der dyadic consent protocols, f airness evaluation disaggre- gated across race , gender identity, neurotype, and cu ltural background, and architectural commitmen ts to on -device processing th at preven t affective data from becoming pla t- form surveillance infrastructure. This vision paper calls on the WICV community, whose members are uniquely posi- tioned to understand both the t echnical opportunity and the human stakes, to establish online dating safety as a first - class research doma in before co mmercial deploymen t out - paces ethical d eliberation. 1. Int r odu cti on Online dating h as bec ome the dominan t way p eople m eet roma nti c par tne rs acr oss t he US a nd Euro pe [ 21 , 39 , 90 ], yet the comm unication interfaces supportin g it are remarkably pri mit iv e f rom a p erc ept ua l s tandp oin t. Fac e- to -f ace r oman - tic interac tion is richly multimo dal: peop le read microex - pressions, track gaze av ersion, register hesitation, and inter- pret body postur e to continuously calibr ate comfort, inter - est, and distress [ 27 , 29 , 32 , 82 ]. T ext-based messaging and emoji redu ce th is rich signal to a narrow, lo w -bandwidth channel [ 8 , 58 , 88 ], while even video dating platfo rms pro- cess their stream s f or compression an d presentation rather than affective conten t. The result is a profound mo dality mismatch betwe en how human s are equip ped to co mmuni- cate roman tically and wh at current platforms ac tually sup - port . This gap is not perceptu ally neutr al-it is socially con se- quential. Resear ch in social psychology consistently shows that women disproportionately manage unwanted social ad - vances in d ating co ntexts and rely heavily on nonverbal channels to signal di sinterest s afely and without con fronta- tion [ 60 , 82 ]. When those channels are absent, disco mfort signals go unread, unwan ted escalation go es unchecked, and the burd en falls o n the more vulnerable par ty to co m - municate refusal exp licitly in a medium that was n ever de - signed for it (e.g., matrimonial sites[ 67 , 81 ]). The absence of an affective signal does not affect all users equally- it am - plifies existing vu lnerabilities. Meanwhile, the computer vision community has mad e sub st anti al advanc es in pre ci sel y the techn olo gie s that coul d address this gap. Facial action unit detection [ 3 ], gaze esti- mation [ 97 ], emotion rec ognition [ 59 ], an d multimodal af - fective modeling [ 64 ] have each matured considerably over the p ast decade. Yet almost none of this work has been ex - amined through the lens of dating commun ication, consent signaling, or in timate interaction safety. There is no estab- lished research agenda connecting affective CV to online dating as a sociotechnical syst em, and no fairness literature examining how such tools would perform across gender, age, race, and sexual orientation in this specific context. This paper argues that this g ap is both a techn ical o p - portunity and a moral responsibility for the CV commu - nity. We ask: what nonverbal signals are most diagnostic of comfort, d isinterest, and distress in romantic communica - tion contexts, a nd can they be r eliably detected from vi deo ? 2 How does the absence of these signals in platfo rm-mediated dating d isproportionately affect women ’s safety and auton- omy? A nd what would a fairness-first, consent-aware affec- tive v ision resear ch agen da f or online dating actually look like? We outline this agenda, identify the key technical and ethical challeng es, an d call on th e CV co mmunity to take ownership of it. 2. The Nonverbal Signal Gap in Dating Com- muni ca ti on Human communication in roman tic co ntexts, such as face- to - face d ating, is f undamentally nonverbal. Lon g be - fore words are exc hanged, people rely on a continuous stream of physical signals such as direction of gaze, p os - tural orientation , proximity, touch initiation, and the timing and qu al it y of fac ia l expr ess ions - to ass ess att ract io n, sig nal ava il abi lit y, a nd ne got iate c omfor t [ 9 , 27 , 32 ]. These sign als are not supp lementary to romantic communicatio n; they are its primary substrate [ 50 , 75 ]. They are o ften subtle, rapid, and partly involuntary, and play a ce ntral role in how people convey both romantic interest an d rejection. [ 29 ] demonstrated that non-verbal cu es during courtship inte rac- tions predicted ro mantic interest m ore reliably than verbal content alone, while [ 70 ] showed that sustained gaze, open body orientatio n, space m aximization were systematically deployed by men seeking to attract f emale attention. Flirt- ing itself is distinguished from ordinary friendly interaction almost entirely through nonverbal m odulation of otherwise neutral verbal co ntent [ 31 ]. Of particular interest is the role of non-verbal cues in signaling d isinterest and discomfort. Literature from so - cial psychology has long established that explicit verbal re - fusal is not the primary mechanism through wh ich disin - terest is commun icated, particularly for wo men navigating unwanted advances [ 60 , 82 ]. Gaze aversion , reduced smil- ing, postur al withdrawal, and response latency elongation are among the signals people, dispro portionately wom en, deploy to c ommunicate disengagemen t without direct con- frontation [ 53 ]. Abb ey’s [ 1 ] foundational work demon- strated that men system atically misattribu te friendly be - havior as sexual interest, a mispercep tion that nonverbal cues in face - to -face contexts can p artially correct but that text-based co mmunication eliminates entirely . Th e conse- quences of this signal loss are not trivial: [ 22 ] linked the mis pe rce ptio n o f s exu al in te nt d ire ct ly to coe rci ve be havi or. App-based communicatio n strips most of this away, leaving primarily text, emojis, and o ccasional static im - ages. Work by W alther [ 91 , 92 ] established that computer- mediated co mmunication, even when asy nchronous text is replaced by rich er modalities, fundamentally alter s the in - terpersonal dynamics of interaction. Emoji and emoticons, the primary aff ective tools available for dating app users, are a poor substitute for genuine nonverbal signals. Wh il e they can convey broad valence, they require ac tive, e xplicit, deliberate inp ut: they cannot capture the involuntary mi - croexpressions, gaze behav ior, or postur al shifts that carry the most diagnostically useful information [ 16 , 73 ]. Kruger et al. [ 52 ] demonstrated th at co mmunicators dram atically overestimate h ow well their emotional tone is conveyed in text, suggesting that dating app users are lik ely op erating under a systematic illusion of commu nicative richness that the chann el does not actually support. Video dating, which expanded dramatically during covid- 19 pandem ic and has since become a standard fea- ture, might appear to close this gap. It does not. Even when an aud io or v ideo is present, cu rrent platforms rarely do anything computationally with the rich nonverb al stream. Video-med iated communication in troduces its own set of perceptual distortions: eye contact is structur ally impossi- ble due to camera-screen offset [ 6 , 30 ], sp atial faithfulness is comp romised by fixed camera angles [ 61 ], and th e self- vie w inte rfa ce crea te s a sel f-mon it ori ng burden that disr upt s natural expressions. Crucially, n o current d ating p latform processes its video stream for affective content in any user- beneficial way. The r esult is a structural nonv erbal signal gap: key cues about co mfort, c onsent, and in terest are miss- ing from the inter action channel. Th is loss is not psycholog- ically symmetric. Individuals who ar e less ass ertive or mo re conflict-av oidant patterns that ar e shaped by, but not re - ducible to, gender socialization often rely on gradual, non - verbal signals to indicate disinterest because saying “no” directly feels risky or costly [ 43 , 44 , 87 ]. Fo r many women and marginalized groups, nonverb al cu es p rovide a safer, low er- conf li ct way to communi ca te bounda ri es in flirt ati ous or ambig uous situations [ 41 , 85 ]. When platforms remove or minimize these channels, they d isproportionately disad - vantage users who d epend on t hem to m anage u nwanted ad - vances and avoid confrontation. From a computational per- spective, this raises concrete q uestions about which nonver - bal sign als, if used carefully an d with consent, could help recover som e of the lost safety and clar ity. This sectio n establishes two claims that m otivate the rest of the paper. First, nonverbal c ues are not p eriph- eral to romantic communication but central to it, partic - ularly for the signaling of discomfort and disinter est, and their absence creates measurable communicative and safety harms. Second, video dating platforms ar e technically ca - pable of capturing these cues but cur rently do not, creating a tractable r esearch g ap that th e comp uter vision community is uniquely position ed to address. 3. What Affective Vision Could Offer: A Pro- posed Research Agenda Having established that nonverbal sig nal is both essential to romantic communication and lar gely ab sent f rom cur - rent platforms, we now outline a concrete r esearch agenda 3 for affective co mputer vision in th is do main. While this paper foregrou nds v isual modalities co nsistent with its CV framing, we acknowledge that audio prosody — pitch vari- ation, speech rate, p ause d uration, hesitation marker s — and turn-taking dy namics ca rry stro ng complementary sig- nal for discomfort and engagement [ 54 , 79 ]. A complete system should be multimodal by design [ 65 ]. We treat the visual chann el as th e p rimary co ntribution of this agenda while explicitly flagging audio and linguistic featu res as natural and nec essary extensions. We o rganize this agenda around four inter connected ca - pability areas: discomfort and disinterest detection, engage- ment asymmetr y modeling, consent -aware interac tion de- sign, and longitu dinal interaction summarizatio n. For each, we describe the technical approach, identify th e key open research questions, and discuss wh at user -facing applica- tions wo uld lo ok like in practice. Throughout, we emph a- size that the goal of this agenda is user em powerment, giv- ing i ndi vi duals more leg ibl e info rma ti on abo ut t heir own in - teractions, rather than platform surveillance or autonomous dec is ion -ma kin g. 3.1. Discomfort and Disinteres t D e tec ti on The most direct application of affective CV to dating safety is th e rea l-time detection of negative affec t an d disengag e - ment during video interactions. Facia l Action Codin g Sys- tems (FACS), formalize d by [ 19 ] and now automatable at high accurac y through tools such as OpenFace 2.0 [ 3 ], pro- vides a theoretically g rounded and expression-agnostic rep- resentation of facial behavior. Sp ecific AU patterns such as AU4(brow lowering), AU15 (l ip corner depression), AU17 (chi n ra is in g) ar e rel ia bly associ at ed wi th d isc omfo rt in ne u- rotypical Western populations [ 19 , 86 ]. Gaze detection, de - tectable through standard webcams, is am ong the most con- sistent non -verbal markers of social d isengagement and dis- comfort [ 46 , 97 ]. While no prototype is presented here, in - dividual co mponents have been demonstrated in adjacen t domains: OpenFace 2.0 processes fac ial behav ior in real time at 30fps o n consumer hardware [ 3 ], monocular gaze estimation achieves un der 4° mean angular er ror in uncon - strained settings [ 46 ], an d engagement detectio n from video has been validated in educational dyadic settings [ 93 ] . Th e open research challeng e is theref ore no t componen t feasi - bility but i ntegration, ecological validity in dating contexts, and equitable p erformance across demogr aphic groups. We d efine discomfort as a s ustained co- occurrence of the above AU patterns persisting b eyond two seconds, com- bined with gaze aver sion exceedin g the ind ividual baselin e by more than 1 .5 standard deviations, with labels validated against participant self-report using in traclass corr elation coefficients [ 12 , 26 ]. Disinterest is operationalized as re - duced response contingency between partners, measured as cross-correlatio n lag in express ive streams f alling below an empirically established threshold [ 15 ], combined with re- duced head nod frequency relative to individual baseline. Consent signaling is treated as a composite behavioral state rather than a single cue [ 22 ], requiring dedicated annotation protocols dev eloped in collaboration with social psycho lo- gists before any dataset collection begins. We acknowledge that Barrett et al. [ 5 ] have raised fund amental challenges to inferring d iscrete emotional states from facial m ovements alone, and we th erefore treat th ese AU p atterns not as di - rect emotion readouts but as b ehavioral correlates requiring contextual validation in the specific dyadic romantic set - ting. We also no te that AU- to -affect mappings carry well - documented cro ss-cultural validity limitations [ 45 , 57 ], a challenge we treat as a core fairness requirement in S ec ti on 4.4 rather than an implementation detail, an d one that any deployed system must address through culturally stratif ied validation rath er than assuming universality. The key research questions here ar e not primar ily whether individ ual cues can be d etected, but whether they can be reliably interpreted in a naturalistic, un constrained, and emotion ally complex context of dating. Existing af - fective datasets, such as RECOLA [ 72 ] and SEWA DB [ 51 ] capture dyadic remote interaction s bu t wer e not col - lected in romantic contexts and also do not include consent- relevant beh avioral annotations. New d ata co llection, with appropriate ethical over sight and dyadic consen t proto - cols, is n eeded. Detection also needs to distinguish gen- uine discomfort from co nversational norms such as looking away while thinking, laug hing with eyes closed, or cultural variation in baseline g aze behavior [ 63 ]- a significant b ut tractable signal p rocessing challenge. In terms of appli ca ti on, we env is i on dis com fort dete cti on not as an au tonomous intervention but as a user -contr olled awareness tool, analogo us to heart rate mon itor that a u ser can act on. A simple, private, en d- of -interaction summary (”there wer e several momen ts where yo ur expressions sug- gested discomfort, here is where they occurred”) could pro- vide the kind of reflective affordance that current platforms entirely lack. Critically, th is data would b e processed and displayed only to the user whose face is being analyzed, never shar ed with the other party or the platf orm. 3.2. Engagement and A symmetry Mod el ing Beyond individual affect d etection, a particularly promis - ing direction is the modeling of engagement asymmetry be - tween partners, detec ting when one party is significan tly more invested, atten tive, o r emotionally activated than the other. Interpersonal synchrony research has established that mutual engag ement in dyadic interaction produces measur- able behavioral align ment: matched f acial expressions, co - ordinated head movem ents, synch ronized response timing [ 15 , 55 ], Conversely, asymmetric engagemen t, wh ere one part y 4 is highly activ ated and the other is withd rawn, prod uces a characteristic divergence in th ese behavioral streams. Oper- ationalizing en gagement asymmetry requires em pirical de - cisions on tempo ral windo w, synchrony research suggests 30 – 60 seco nd win dows b alance sensitivity and noise [ 15 ], asymmetry thresh old, and individual baseline calibratio n to avoid penalizing naturally low -expr essivity u sers. We propose that these param eters be estab lished empirically thr ough cor rel at io n with sel f-r epor te d comfor t rati ngs rathe r than set a priori, and treat them as primary empirical contri- butions of the benchmark proposed in Section 3.5. Modeling this asymmetry requires moving bey ond single-face an alysis to genuine dyadic modeling, trac king both participants simu ltaneously and compu ting relational features across the pair rather than treating each face inde - pendently. Techniques from group behavior analysis [ 76 ] and virtual rapport modeling [ 40 ] provide relevant method- ological found ations, though they wer e not developed for romantic dyads and have not b een validated in this co ntext. The research questions here include: Can engagement asymmetry be detec ted reliably fr om n aturalistic v ideo (speed) dating setting s? What lev el of asymmetry is diag - nostically mean ingful versus normal variation? What tem- poral window is ap propriate: asymmetry meaningful over seconds, minutes, or the full interaction? How does the model handle tu rn-taking dynamics where asymmetric at - tention is structur ally expected? And critically, what level of asymmetry is diag nostically meaningfu l versus normal variation? Answering these questions requires bot h datasets and close collaboration with social p sychologists wh o study romantic relatio nships, attractio n, and interest signaling in dyadic contexts [ 68 ]. 3.3. Co ns en t-A wa re In ter ac ti o n De sig n A third directio n moves from detec tion towar d design, us - ing affective CV insights to build platform affordances that make it easier to disengage f rom uncomfortable interac - tions without confrontation. This is p artly a CV problem and partly an HCI problem, and it requir es bo th communi - ties wo rking togeth er. The cor e in sight is that platf orms currently offer only blunt exit mechanisms: un matching, blocking, and ending a call, which require explicit, socially costly ac tion. There is no soft affordance for signalin g re - duced interest that does not require the user to initiate a po - tentially conf rontational act. CV -power ed tools could support subtler mechanisms. A detected pattern of sustained disengag ement could quietly surface an easy exit option: ”wo uld you like to wrap up this call?” without requ iring the user to explain or justify. Work o n persuasive techno logy [ 23 ] and behavior change design [ 13 ] p rovides a fram ework for thinking about h ow such nudges can be designed to feel supportive rather than intrusive. Rader’s work [ 66 ] on algorithmic transparency is rel eva nt to how a ny aff ect i ve proc es sin g s houl d be di sc lo se d to users. The key research questions here are primarily in the HCI domain b ut depend on CV inputs. What interface designs for surfacing aff ective f eedback are experienced as h elpful versus invasive? How should the system communicate un - certainty, what happ ens when the mo del is un sure whether discomfort is gen uine? How do users fr om differ ent demo- graphic g roups respond to affective nudges, and are th ere differential effects by gender , age, or cultural bac kground? These q uestions require user studies with diver se partici - pants, ideally in collaboration with dating platform partners who can suppo rt ecologically valid dep loyment. Participatory safety research in social match ing systems has documented wo men’s concrete design priorities for safer dating interac tions [ 2 ], and consent techno logy de - sign with LGBTQ+ stakeholder s has identified specific af - fordances supporting autonomous refusal [ 99 ]. Th e capabil- ity areas we propose should be developed in direct dialogue with this literature r ather than in isolation.” 3.4. Longitudinal Interaction S um mar iza t ion A fo urth and more spec ulative direction is the post -hoc summarization of interaction dynamics across a video date: giving users a visual and affective ”replay ” of how the inter- action evolved, analogous to h ow fitness apps provide p ost- workout analysis. Rather than real-time intervention, this approach surfaces patterns across the full interaction: mo - ments o f hig h mutual engagement, poin ts where o ne p art y wit hdr ew , and t he t empo ral t raj ect ory of expr ess ed a ffe ct on both sides. This fram ing draws o n precedents in aff ective comput - ing for educational and therap eutic contexts [ 28 , 49 ], wh ere longitudin al b ehavioral feedba ck has been shown to support se lf- awar ene ss and be hav ior cha nge. In a da ti ng cont ex t, th e value proposition is less about changing b ehavior in the mo - ment and more about helping u sers, p articularly those who find it difficult to trust th eir own r ead of an interac tion, de - velop a more g rounded understanding of what actually hap- pened. For users who have experienced gaslighting or co er- cive control in past relationships, objective behavioral data about in teraction dynamics could be meaningfu lly support- ive. The resear ch questions her e in clude: what visualiza - tions of affective interaction data are in terpretable and use- ful to non-ex pert users? How shou ld uncertainty in model outputs be communicated so that users do n ot over-interpret noisy signals? And what are the risks of misuse? Could a bad actor use interaction su mmaries to refine manipulative techniques? This last question con nects directly to Section 5 on privacy an d dual-use risk. 5 3.5. To war ds a S ha red B en ch ma rk Underpinning all four of these capab ility areas is a shared need: a purpose-built dataset for affective CV in r oman- tic dyadic interaction. Existing datasets [ 51 , 68 , 72 ] pro- vide p artial coverag e but were n ot design ed for this pur - pose, lack consen t-relevant annotations, and do no t r eflect the demographic d iversity needed for fair model develo p - ment. Also, existing d yadic annotation tools, including ELAN [ 94 ] and py-feat [ 10 ], provide methodolog ical in - frastructure that could sign ificantly accelerate bench mark creation witho ut requiring bespoke annotatio n pipelines. We call for the creation of a ne w benchmark dataset col- lected u nder rigorous d yadic co nsent proto cols, annotated for disco mfort, disinterest, engagement, and consent sig - naling, and stratified across gender, age, race, sexual ori - entation, an d cultural b ackground. The eth ical framework for this collection is its elf a research contribution and is dis- cussed further in Section 5. 4. Fairness and Equity Cons ide r at io ns The research agend a outlined in Sectio n 3 is technically promising, but its promise is co ntingent on a condition that the affective computing co mmunity has h istorically failed to meet: that the systems built perform equitably across the full demographic range of the peo ple who will use them. This sectio n arg ues that fairness aud iting is not an optional addendum to th e proposed a genda but a prerequisite for a ny res pon si ble de ploy ment , a nd t hat th e sp eci fi c co nte xt of i nti - mate interaction makes the consequences of biased systems particularly sev ere. 4.1. Racial and Skin Tone B ias in Affect ive CV The foundational work by Buolamwini an d Gebru [ 7 ] demonstrated th at commer cial face analy sis systems per - form significan tly worse on darker -sk inned faces and on women, with the worst per formance concentrated at their intersection. Subsequ ent work h as confirmed and extended these finding s: [ 71 ] found th at emotio n classifiers system - atically rated Black faces as expressing more n egative em o- tion than White faces displaying equivalen t expres s ion s. [ 96 ] do cumented similar raci al per formance gap s across multiple comm ercial affective computing appro aches. In a dating context, these biases are not merely techni - cal imperfections; they translate directly into harmfu l user experience s. A system th at systematically m isreads Black wome n’s exp ress i ons as more neg at ive woul d ge ner ate fal se discomfort signals, p otentially cau sing users to distrust in - teractions that were in fact positive, or cau sing platforms, if they had access to th is d ata, to down - rank certain users. Even in a fully u ser-facing, privacy-preserving deployment, a biased system that consistently mischaracter izes a user’s own emotional state erodes tru st an d utility in ways that fa l l disproportionately on already m arginalized users. Any af- fective CV system deployed in a dating context must there- fore dem onstrate equitab le performance ac ross racial and skin tone groups before dep loyment, using evaluation pro - tocols th at go beyond aggregate accuracy to examine p er- formance at dem ographic intersections [ 18 , 35 ]. Age represents a furth er underexamined bias axis: FER systems perform significantly worse on older adults [ 48 ], while major training d atasets i ncluding AffectNet an d FER- 2013 lack adequate elderly representation [ 25 ]. Given that online dating is increasing ly used across wide age ran ges, this bias requires explicit attention in dataset collection and evaluation pro tocols. 4.2. Gender and Limits of Binary Cla ss if ica tio n Most af fective CV systems, inclu ding those for emotion recognition, gaze analysis, and en gagement m odeling, en - code gender as a b inary covariate used either as a feature or as a stratification variable in evalu ation. This becomes problematic for all users but is acutely harmful for tra ns and nonbinary users. Os Keyes[ 47 ] documented how automatic gender recognition systems sy stematically misclassify trans individuals, producing erro rs th at are not random but sys - tematically correlated with the direction of gender transi - tion. Sch euerman et al.[ 78 ] ex tended this to commercial facial analysis and found c onsistent, significant misclassifi- cation rates for trans and nonbinary faces. Ham idi et al.[ 33 ] argued that autom atic gender rec ognition for these popula- tions is not merely inaccu rate b ut structur ally red uctive - it enforces a b inary that many users do n ot inhabit. For af fective CV in d ating contexts this matter s in two ways. First, gend er is often used as a conditioning variable in affect models, if the u nderlying gen der classifier is un - reliable for a significan t user po pulation, affect predictions conditioned o n it will be system atically distorted. Second, the d ating context is one where gender identity is particu - larly salient and where misclassification by a platform sys- tem, even one operating invisibly, could cause real psycho- logical h arm. We ca ll for affective d ating CV systems to be designed without binary gender assum ptions wherever possible, and fo r evalu ation to ex plicitly include trans and nonbinary participants as a requ ired demogr aphic stratum rather than an afterthought. 4.3. Ne ur od ive rg en t Us er s Affective CV system s are trained overwh elmingly on neu - rotypical expressive behavior, calibr ated to th e n orms of how neurotypical p eople display an d recognize emotion. Autistic ind ividuals, for example, may display flat or atyp- ical facial aff ect that does not conform to these norms even when experiencing strong emotion [ 36 ], while individuals wit h alexi th ymi a, diff icul ty identi fyi ng and desc rib ing emo- tional states [ 89 ], may produce expressions that are gen - 6 uinely disco nnected from their inter nal states in ways that confound mo del assumptions. A discomfort d etection sys- tem th at flags an au tistic user ’s b aseline n eutral ex pression as dis tr ess, or t hat fail s to d etect g enui ne dis com fort b ecaus e it is expressed atypically, would be both inaccurate and po - tentially harmf ul. This is not a min or edg e case: au tism spectrum con di- tions affect an estimated 1 – 2% of the population, and neu - rodivergent individuals d isproportionately report difficu lty with the social nav igation demands o f d ating [ 36 ]. These are users who might benefit most from affective support tool, and who are most at risk of being poor ly ser ved by sys- tems trained without them. We call fo r neurodivergent par- ticipants to be explicitly included in dataset collection and for mod el ev aluation to report perf ormance disaggreg ated by neurotype where ground truth labels can be established. 4.4. Cu ltu ra l V ar iat io n in Ex p re ssi on No rms Thi s conc ern dire ct ly qua lifi es th e AU -base d dis com for t de - tection proposed in Sectio n 3.1, where we treat AU pat - terns as behavioral correlates requ iring cross -cultur al val- idation rather than universal emotion reado uts [ 5 ]. The so - cial no rms go verning when and h ow emo tion is ex pressed vary s ign ifi cant ly acr oss cu ltur es [ 20 , 57 ]. Jack [ 45 ] de mon- strated th at the facial expressions of em otion are no t fully culturally universal, challenging the Ekman mod el that un - derlies m ost affective CV sy stems. A gaze aversion that signals d iscomfort in one cu ltural contex t may be a rou tine politeness norm in anoth er; a smile that indicates g enuine warmth in on e context may be a social mask in an other. Models trained on WEIRD [ 38 ] populations will systemati- cally misread u sers from other cultural b ackgrounds. Online datin g is a glob al phenomenon, and even within single platforms, users span eno rmous cu ltural d iversity. Affective CV systems deployed in this context without cross-cultural validation will p roduce systematically biased outputs for large portions of th eir user base. Hence, we also call for dataset collection to be explicitly cross-cultural and for evaluation to report performance disaggreg ated by cul - tural b ackground, while rec ognizing th is is a harder prob - lem than demogr aphic stratification within a single cultural cont ex t. 4.5. Fa irn es s C ri t eri a Standard fairness metrics from ML literature: d emographic parity, equalized odds, calibr ation [ 11 , 35 ], provide a start- ing point bu t are insufficient in the dating d omain without adaptation. False positives and f alse negatives have asym - metric consequ ences that differ by user group: a false posi- tive discomf ort signal for a user who is co mfortable wastes attention and erodes trust; a false negativ e discomfort sig - nal for an u ncomfortable user fails them at a moment of po - tential vuln erability. Th ese asymmetric harms sugg est th a t standard accuracy-b ased fair ness criteria need to be supple- mented with harm-weig hted e valuation fra meworks that ac - count for the specific social consequences of different error type s f or di ffe rent demo gra phic g roup s. We propo se t hat t he fie ld d eve lop suc h a fr ame wor k as a pri ori ty r es ear ch c ont ri- bution, ideally collaborating with so cial scientists who can quantify the real-world consequences of different error pro- files [ 4 ] 5. Privacy, Consent, and the Limits of the Pro- pos al Processing live video strea ms for emotional content repr e- sents a significant privacy intervention that demands rigor- ous ethical safeguards, granular user consent, and architec- tural protections against misuse. Facial expressions, gaze patterns, and body language reveal intimate p sychological st ates t ha t u se rs ma y not wi sh to ex pos e, eve n to th ems el ve s, in real time [ 14 , 62 ] . This section makes explicit the ethical boundaries of the proposed agenda, addresses its most se - rious risks, an d identifies the open research and regulatory questions that must be resolved before resp onsible deploy- ment is possible. 5.1. Us er- Fa cin g Too ls This proposal explicitly advoca tes only for user -facing tools, wh ere affective analysis occurs lo cally on th e user’s device and remains visible solely to that individ ual (e.g., ”Your facial tension sugg ests discom fort, want to pause or end this call?”). Platform-facing applications [ 69 ], where emotional d ata is collected centrally for matching, mo der - ation, monetization , or reco mmendation, are explicitly ex - cluded due to their high risk of surv eillance and e xploitation [ 98 ]. This distinction has a direct technical corollary: affec- tive p rocessing should occur on -device wherever architec- turally feasible. Federated learning [ 34 ], on-device infer- ence pipelines, and zero-knowledge proof fram eworks [ 74 ] offer concrete pathways to ensure th at raw emotional d ata does not leav e the user’s device. 5.2. The Dyadic Consent P ro b le m Consent in intimate contexts such as dating is particu- larly complex. Power imbalances, social expectation s, and opaque data prac tices can undermin e the conditions neces- sary for truly voluntary agreement. I nformed co nsent for video processing becomes even more c hallenging in dyadic interactions [ 56 ]. For example, in a video date, b oth indi - viduals are simultaneously visible, meaning that t heir facial expressions, gaze patterns, and other b ehavioral cu es may be subject to analysis. This creates a structu ral asymme - try in consent: on e party ma y agree to ha ve their own video stream processed, while the oth er may object to t he analysis of theirs. Such asymmetries complicate traditional model s 7 of individual consent, as the d ata of multiple individuals are inherently intertwined within a single interac tion. We p ropose three principles for dyadic consent in this cont ex t. First, by def ault a ffe ctive pr oce ssi ng s hou ld be li m- ited s trictly to the consenting user’s own behavior al stream, with the other party’s face either ex cluded from an alysis or processed only for featur es that do not leave the device. Second, any analysis that incorp orates the o ther party’s be - havi or, such as engag eme nt asymme try model ing shoul d re - quir e tha t par ty’ s se par ate , expl ici t, inform ed cons en t bef ore activation. Th ird, consent must be g enuinely granular and revocable in real time: users should be able to opt into spe- cific capabilities ind ependently and withdraw at an y point without frictio n or penalty [ 77 ]. Dating platfo rms have faced criticism for treating vague ”OK” pop - up acknowledgements as consent for AI training on chat histories and ph otos [ 56 ], demon strating that the current industry norm for consent is in adequate for high - sensitivity data, highlighting the n eed for ju st- in -time, revo- cable permissions th at do not disrupt interaction flow [ 95 ]. 5.3. Dual Use Risk and Intima te Partner Abus e The m ost serious risk of the proposed agenda is not p lat- form misuse but interpersonal misuse, spec ifically the po - tential for affective dating tools to be weaponized in con - texts of intimate partner abuse an d coer cive control. This has bee n studied in various cases such as ab users exp loit - ing consumer technology [ 24 ], as a m echanism of ongoign psycholog ical control [ 37 ] among others. This risk is compounded by the broader deployment of emotion AI in hgih -stakes co ntexts, in cluding bor der control, employee performance mo nitoring, in surance as - sessment, where similar dual-use con cerns have been ex- tensively documented [ 17 , 84 ]. What we propose is a set of ar chitectural an d design princip les that reduce but cannot elimin ate it: strict on-device processing an d clear public documentation of misuse risks in any published syt em /dat as et. 5.4. Regulatory Landscape and Open Legal Ques- tio ns Thre e open res ear ch quest io ns guide re spo nsi bl e imple men- tation. First, what consen t m echanisms are technically and experientially appropriate for affective vid eo pr ocessing in intimate co mmunication ? GDPR Article 9 classifies most emotion data as biometric special catego ry information re - quiring explicit consen t and Data Protection Impact As - sessments, yet dating -specific UX r emains underexplored [ 80 ]. Second, how can affective tools be architecturally de - signed to prevent platform- side m isuse? Fed erated learn- ing, on-dev ice inference, and zero -knowledge proofs offer pathways to ensure emotional data never leav es the phone. Third, what regulatory gap s persist? Wh ile th e EU AI Act [ 83 ] prohib its emotion recognition in work places and ed u - cation, no ju risdiction explicitly g overns its use in co nsen - sua l datin g conte xts , leavi ng enforc eme nt incons ist en t glob- ally [ 42 ]. 5.5. The Limits of This P ro po s al We close by being exp licit abou t what this p aper does not claim. We do not claim that affective com puter vision will ”solve” dating safety. It constitutes one tool within a broader sociotechnical ecosystem that m ust integrate plat - form po licies (proac tive moderation, easy blocking) , legal fra mew orks (ant i- har assm ent laws), and cultura l norms (by- stander intervention). We also do not claim that the tech - nical capab ilities claimed in section 3 are ready for de - ployment, they requ ire new datasets and extensive fairness auditing. Current models also face technical limits: cu l- tural b ias, neurodiversity ch allenges, and imperfect accu - racy, that no system can fully overcome without human oversight and ethical responsibility. The proposal is scoped specifically to ea rly-stage, dyadic romantic interaction, not workplace, familial, or public settings. 6. Co nc lus ion Online dating has become the primary context in which romantic relatio nships begin, yet the platf orms mediating these interactio ns remain perceptually impoverished, strip - ping th e nonverbal cues that humans rely on to navigate at - traction, discomfort, and consent. This paper has argued that this gap is neither in evitable n or neutral: it is a design failure with mea surable consequences that fall dispr opor- tionately on wo men and vulnerable users. Addressing this issue lies sq uarely within the techn ical ca pabilities of the computer v ision community . Matu re affective CV tools, such as facial action unit detection, gaze estimation , en - gagement modeling, an d multimodal affect recognition , al - ready e xist. W hat has bee n mis sing is a research agenda that treats the dating contex t ser iously, positions fairness and consent as f oundational requiremen ts, and asks what these tools s hould be ac complishing r ather than merely what they can do. We have outlined such an ag enda, grounded in t he social psychology of dating and organized ar ound d iscomfort de - tection, engag ement asy mmetry modeling, consent -awa re interaction design, and longitudinal interac tion summariza- tion. Pur suing it responsibly dem ands new datasets to be collected un der dy adic co nsent pro tocols, fair ness ev alua - tion disaggr egated across r ace, gender identity, neurotype, and cu ltural background, and architectural co mmitments to on -device processing that prevent affective data from be - coming platform surv eillance infrastru cture. Th e CV com - munity has b uilt much of the techn ical substrate that con - temporary dating platforms alrea dy run on, largely without asking what r esponsibilities that entails. 8 Re fer en ce s [1] Antonia Abbey. Sex differences in attributions for friendly behavior: Do males misperceive females’ friendliness? Journal of Personality and Social Psychology , 42(5):830 – 838, 1982. 2 [2] Hanan Khalid Aljasim and Douglas Zytko. F oregrounding women’s safety in mobile social matching and dating apps: A participatory design study. Proceedings o f the ACM on Human-Computer Interaction , 7(GROUP):1 – 25, 2022. Ar - ticle No. 9. 4 [3] Tadas Baltrusaitis, Amir Zadeh, Y ao Chong Lim, a nd Louis- Philippe Morency. Openface 2.0: Facial behavior analysis toolkit. In 2018 13t h IEEE International Conference on Au - tomatic Face Gesture Recognition (FG 2018) , pages 59 – 66, 2018. 1 , 3 [4] Solon Barocas, Moritz Hardt, and Arvind Narayanan. Fair- ness and Machine Learni ng: Limitations and Opportunities . MIT Press, Ca mbridge, MA, 2 023. 6 [5] Lisa Feldman Barrett, Ralph Adolphs, Stacy Marsella, Aleix M. Martinez, and S eth D. P ollak. Emotional expres- sions reconsidered: Challenges to inferring emotion from human fa cial movements. Psychological Science in the Pu b- lic Interest , 20(1):1 – 68, 2019. 3 , 6 [6] Leanne S Bohannon, Andrew M Herbert, Jeff B Pelz, and Esa M Rantanen. Ey e contact and video-mediated commu- nication: A review. Displays , 34(2):177 – 185, 2013. 2 [7] Joy Buolamwini and Timnit Gebru. Gender shades: Inter- sectional accuracy disparities in c ommercial gender classifi- cation. In Conference on fairness, accountability and trans- parency , pages 77 – 91. PMLR, 2018. 5 [8] Judee K. Burgoon, David B. Buller, and W. Gill Wooda ll . Nonverbal Communication: The Unspoken Di al og ue . McGraw-Hill College, 2 editi on, 1995. 1 [9] Judee K. Burgoon, Lau ra K. Guerrero, a nd Kory Floyd. Non- verbal communication . Routledge, New York, NY, 2nd edi- tion, 2016. 2 [10 ] Jin Hyun Cheong, Es hin Jolly, Tia nkang Xie, Sophie Byrne, Matthew Kenney, and Lu ke J. Chang. Py -Feat: Python facial expression analysis toolbox. Affective Science , 4 :781 – 796, 2023. 5 [11 ] Alexandra Chouldechova. Fair pr ediction with disparate im - pact: A study of bias in recidivism prediction instruments. Big Data , 5(2):153 – 163, 2017. 6 [12 ] Jeffrey F. Cohn and Paul Ekman. Measuring facial action. In Th e New Handbook of Methods in Nonverbal Behavior Research , pages 9 – 64. Oxford University Press, 2005. 3 [13 ] Sunny Consolvo, David W McDonald, and James A Lan - day. T heory-driven design strategies for technologies that support behavior change in everyday life. In Proceedings of the SIGCHI conference on human factors in co mputing sys- tems , pages 405 – 414, 2009. 4 [14 ] J Solove Daniel. A taxonomy of privacy, 154 u. Pa. L. Rev , 477:484, 2006. 6 [15 ] Emilie Delaherche, Mohamed Chetouani, Ammar Mahd- haoui, Catherine Sain t-Georges, Sylvie Viaux, and David Cohen. Interpersonal synchrony: A survey of evaluati on methods across disciplines. IEEE Tran sactions on Affective Computing , 3(3):349 – 365, 2012. 3 , 4 [16 ] Daantje Derks, Arjan ER Bos, and Jasper Von Grumbko w. Emoticons and online me ssage interpretation. Social Science Computer Review , 26(3):379 – 388, 2008. 2 [17 ] Molly Dragiewicz, Jean Burgess, Aria dna Matamoros- Ferna´ndez, Michael Salter, Nicol as P Suzor, Delanie Wood- lock, and Bridget Harris. Technology facilitated coercive control: Domestic v iolence and the competing roles of d igi- ta l me di a platf orm s . Femi nis t Media Stud ie s , 18(4) :6 09 – 625 , 2018. 7 [18 ] Cynthia Dwork, Moritz Hardt, Toniann Pitassi, Omer Rein- gold, and Richard Zemel. Fairne ss through awareness. In Proceedings of the 3rd innovations in theoretical computer science conference , pages 214 – 226, 2012. 5 [19 ] Paul Ekman and Wallace V Friesen. Facial action coding system. Environmental Psychology & Non verbal Behavior , 1978. 3 [20 ] Hillary Anger Elfenbein and Nalini Ambady. On the uni- versality and cultural specificity of emotion recognition: A meta-analysis. Psychological Bulletin , 128(2):203 – 235, 2002. 6 [21 ] Nicole B. Ellison, Rebecca D. Heino, and Jennifer L. G ibbs. Managing impressions online: Self-presentation processes in th e online dating environment. Journal of Computer- Mediated Communication , 11(2):415 – 441, 2006. 1 [22 ] Christina Farris, Theresa A. Treat, Richard J. Viken, and Roger M. McFall. Sexual coercion and the misperception of sexual in tent. Clinical Psychology Review , 28(1):48 – 66, 2008. 2 , 3 [23 ] Brian J Fogg. Persuasive technology: using computers to change what we think and do. Ubiquity , 2002(December):2, 2002. 4 [24 ] Dia na Fr ee d, Ja c kel ine Pal m er , Dian a Mi nc ha la , Kar e n Le vy, Thomas Ristenp art, and Nicola Dell. “a stalker’s p aradise” how inti ma te par tne r a bus e rs exp loi t tec hno log y. In Proc ee d- ings of the 2018 CHI conference on human factors in com - puting systems , pages 1 – 13, 2018. 7 [25 ] F. Xavier Gaya-Morey, Jose M. Buades-Rubio, P hilippe Pal an que , Raque l Lacu es ta , and Cri st ina Ma nre sa -Y ee. Dee p learning-based facial exp ression recognition for the elderly: A system atic review. 2025. arXiv:2502.02618 [cs.CV], su b- mitted 4 Feb 2025. 5 [26 ] Jeffrey M. Girard and Jeffrey F. C ohn. A primer on observa- tional measurement. Asse ssment , 23(4):404 – 413, 2016. 3 [27 ] Da vid B. G ive ns . The n onv er ba l b as is of a ttr ac ti on: flirt at ion , courtship, and seduction. Psychiatry , 41(4):346 – 359, 1 978. 1 , 2 [28 ] Joseph F Grafsgaard, Joseph B Wiggins, Kristy Elizabeth Boyer, Eric N Wiebe, and James C Lester. Automati- cally recogn izing facial in dicators of frustration: a learning- centric a nalysis. In 2013 humaine association conference on affective computing and intelligent interaction , pages 159 – 165. IEEE, 2013. 4 [29 ] Karl G rammer, Kerstin Kr uck, Annette Juette, and Bern hard Fink. Non-verbal behavior as courtship signals: the ro le of control and choice in selecting partners. Evolution and Hu- man Behavior , 21(6):371 – 390, 2000. 1 , 2 9 [30 ] David M Gra yson and Andrew F Monk. Are you looking at me? eye contact and desktop video conferencing. ACM Transactions on C omputer-Human Interaction (TOCHI) , 1 0 (3):221 – 243, 2003. 2 [31 ] Jeffrey A. Hall and Chen Xing. Th e v erbal and nonverbal correlates of the five flirting styles. Journal of Nonverbal Behavior , 39(1):41 – 68, 2015. 2 [32 ] Judith A. Hall and Chia-huei Xing. T he interpersonal com- munication of attraction: Detection and interpretation of nonverbal cues. Personality and Social Psychology Review , 20(3):223 – 244, 2016. 1 , 2 [33 ] Foad Hamidi, Morg an Klaus Scheuerman, and Stacy M Branham. Gender recogn ition or gender reductionism? the soc ia l impl ica tio ns of em be dde d ge nde r reco gnit io n sys te m s. In Proceedings of the 2018 c hi conference on human factors in computing systems , pages 1 – 13, 2018. 5 [34 ] Andrew Hard, Kanishka Rao, Rajiv Mathews, Swaroop Ra - maswamy, F ranc¸ oise Beaufays, Sean Augenstein, Hubert Eichner, Chloe´ Kid don, and Daniel Ramage. Federated learning for mobile keyboard prediction. arXiv preprint arXiv:1811.03604 , 2018. 6 [35 ] Moritz Hardt, Eri c Price, and Nati Srebro. Equality of op- portunity in supervised learning. Advances in neural infor- mation processing systems , 29, 2016. 5 , 6 [36 ] Megan B. Harms, Anissa Martin, and Gregory L. Wallace. Facial emotio n recognition in autism spectrum disord ers: A review of b ehavioral and neuroimaging studies. Neuropsy- chology Review , 20(3):290 – 322, 2010. 5 , 6 [37 ] Bridget A Harris and Delanie Woodlock. Di gital coercive control: Insights from two landmark domestic violence stud- ies. Th e British Journal of Cri minology , 59(3):530 – 550, 2019. 7 [38 ] Joseph Henrich, Steven J He ine, and Ara Norenzayan. Most people are not weird. Nat ure , 466(7302):29 – 29, 2010. 6 [39 ] Gunter J. Hitsch, Ali Ho rtac¸su, and Dan Ariely. Matching and sorting in onli ne dating. American Economic Review , 100(1):130 – 163, 2010. 1 [40 ] Lixing Huang, Louis-Philippe Morency, and Jonathan Gratch. Virtu al rappo rt 2.0. In International workshop on intelligent virtual agents , pages 68 – 79. Springer, 2011. 4 [41 ] Em ily A. I mpe tt, She ll y L. G a ble , and Le ti ti a A. Pep la u. Gi v- ing up and g iving in: The costs a nd benefits of daily sacrifice in intimate relationships. Journal of Personality a nd Social Psychology , 99(3):327 – 344, 2010. 2 [42 ] International Association of Privacy Professionals. Bio- metrics in th e eu: Na vigating the gdpr, ai act, 2025. Re tri ev ed f rom http s: //ia pp. org /n ew s/a /b iom e tric s- in -t he- eu - navigating-the-gdpr- ai -act. 7 [43 ] Dana Cro wley Jack. Silencing the self: Women and d epres- sion . Harvard U niversity Press, Cambridge, MA, 1991. 2 [44 ] Dana Crowley Jack and Alisha Ali. Silencing the self across cultures: Depression and gender in the social world . Oxford University Press, New York, NY, 2010. 2 [45 ] Rachael E. Jack, Oliver G. B. Garrod, Hui Yu, Roberto Cal- dara, and Philippe G. S chyns. Facial expressions of emo - tion are not culturally universal. P roceedings of the Na tional Academy of Sciences , 109(19):7241 – 7244, 2012. 3 , 6 [46 ] Pet r K el lnh ofe r, Adri a R ec ase ns , Sim on Ste nt , Woj ci ec h Ma - tusik, and Anton io Torralba. Gaze360: Physically un con- strained gaze estimation in the wild. In Proceedings of the IEEE/CVF international conference on computer visi on , pages 6912 – 6921, 2019. 3 [47 ] Os Keyes. The misgend ering machines: Trans/hci impli- cations of automa tic gender rec ognition. Proceedings of the ACM on human-computer interaction , 2(CSCW):1 – 22, 2018. 5 [48 ] Euge ni a K im , De’ Ai ra B rya nt , Dee pak Sri kan th, and A ya nna Howard. Age bias in emotion detection: An analysis of facial emotion recognition performance on y oung, middle-aged, and older adults. In Proceedings of th e 2021 AAAI/ACM Conference o n AI, Ethics, and Society (AIES ’21) , p ages 638 – 644, New York , NY, USA, 2021. Association for Com- puting Machinery. 5 [49 ] Predrag Klasnja, Sunny Consolvo, and Wa nda Pratt . How to evaluate technologies for health behavior change in hci re- search. In Proceedings of the SIGCHI conference on huma n factors in computing systems , pages 3063 – 3072, 2011. 4 [50 ] Karel Kleisner, Toma´sˇ Kocˇnar, Jana Turecˇkova´, David Stella, and Jan Havl´ıcˇek. Perceived attractiveness of opposite-sex faces and courtship success of their owners. Personality and Individual Differences , 77 :106 – 110, 2015. 2 [51 ] Jean Ko ssaifi, Robert Walecki, Yannis Panagakis, Jie Sh en , Maximilian Schmitt, Fabien Ri ngeval, Jin g Han, Vedhas Pandit, Antoine To isoul, Bjo¨r n Schuller, e t al. Sewa d b: A rich d atabase for audio-visual emotion and sentiment re- search in th e wild. IEEE transactions on patt ern analysis and machine intelligence , 43(3):1022 – 1040, 2019. 3 , 5 [52 ] Justin Kruger, Nicholas Epley, Jason Parker, and Zhi-Wen Ng. Egocentrism over e-mail: Can we communicate as well as we think? Journa l of Personality and Social Psychology , 89(6):925 – 936, 2005. 2 [53 ] Marianne LaFrance, Mary Anne Hecht, and Elizabeth Levy Paluck. The contingent smile: a meta-analysis of sex d iffer- ences in smiling. Psychological Bull etin , 129(2):305 – 334, 2003. 2 [54 ] Stephen C. Levinson an d Francisco T orreira. Timing in turn- taking and its implications for processing models o f lan - guage. Frontiers in Psychology , 6:731, 2015. 3 [55 ] Max M Louwerse, Rick Dale, Ellen G Bard, and Patrick Je - uniaux. Behavior matching in multimodal comm unication is synchronized. Cogn itive science , 36(8):1404 – 1426, 2012. 3 [56 ] Ewa Luger, Stuart Moran, an d Tom Rod den. Consent f or all: revealing the hidden complexity of terms and conditions. In Proceedings of t he SIGCHI co nference on Human factors in computing systems , pages 2687 – 2696, 2013. 6 , 7 [57 ] David Matsumoto. Cultural similarities and d ifferences in display rules. M otivation and Emotion , 14(3):195 – 214, 1990. 3 , 6 [58 ] Albert Mehrabian. Nonverbal Communication . Routledge, 1 edition, 1972. 1 [59 ] Ali Mollahosseini, Behzad Hasani, and Mohammad H. Ma - hoor. Affectnet: A database fo r facial expression, v alence, and arousal computin g in the wild. IEEE Transactions on Affective Computing , 10(1):18 – 31, 2019. 1 10 [60 ] Charlene L. Muehlenhard and Laura C. H ollabaugh. Do women sometimes say no when they mean y es? the preva- lence and correlates o f women’s token resistance to sex. Journal of Personality and Social Psychology , 54(5):872 – 879, 1988. 1 , 2 [61 ] David T Nguyen and John Canny. Multiview: improving trust in group v ideo conferencing th rough spatial faithful- ness. In Proceedings of the SIGCHI con ference on Human factors in computing systems , pages 1465 – 1474, 2007. 2 [62 ] Helen Nissenbaum. Privacy as contextual in tegrity. Was h- ington Law Review , 79:119 – 157, 2004. 6 [63 ] Skyler S Place, Peter M Todd, Jinying Zhuang, Lars Penke, and Jens B Asendorpf. Judging romantic interest of others from thin slice s is a cross-cultural ability. Evolution and Hu - man Behavior , 33(5):547 – 550, 2012. 3 [64 ] Soujanya Poria, Erik Cambria, Rajiv Bajpai, and Amir Hus- sain. A re view of a ffective c omputing: From unimodal anal- ysis to multimodal fusion. Information Fusion , 37:98 – 125, 2017. 1 [65 ] Soujanya Poria, Devamanyu Hazarika, Navonil Majumder, and Rada Mihalcea. Beneath the Tip of t he Iceberg: Current Challenges and New Directions in Sen timent Analysis Re - search . IEEE Transactions on Aff ective Computing , 14(01): 108 – 132, 2023. 3 [66 ] Emilee Rad er, Kelley Cotter, and Jang hee Ch o. Explan ations as mechanisms for supporting algorithmic transparency. In Proceedings of the 2 018 CHI conference on human factors in computing systems , pages 1 – 13, 2018. 4 [67 ] Benson Rajan. “it follows y ou home”: Emotional and psychological impacts o f dating-app harassment o n indian women. In Women’s S tudies International Forum , page 103129. Elsevier, 2025. 1 [68 ] Rajesh Ranganath, Dan Jurafsky, and Dan McFarland. It’s not you, it’s me: Detecting flirting and its misperception in speed-dates. In Proceedings of the 2009 conference on em- pirical methods in natural language processing , pages 334 – 342, 2009. 4 , 5 [69 ] Andrew G. Reece and Christo pher M. Danforth. Instagram photos reveal p redictive markers o f d epression. EPJ Data Science , 6(1):15, 2017. 6 [70 ] Lee Ann Renninger, T. J oel Wade , and Karl G rammer. Get- ting that female g lance: Patterns and consequences of male nonverbal be havior in courtship contexts. Evolution and Hu - man Behavior , 25(6):416 – 431, 2004. 2 [71 ] Lauren Rhue. Racial influence on automated perceptions of emotions. Availa ble at SSRN 3281765 , 2018. 5 [72 ] Fabien Ringeval, Andreas Sonderegger, Juergen Sauer, and Denis Lalanne. Introducing the recola multimodal corpus of remote collaborative and affective interactions. In 2013 10th IEEE international conference and workshops on automatic face and gesture recognition (FG) , p ages 1 – 8. IEEE, 2013. 3 , 5 [73 ] Monica A Riordan. Emojis as tools for emotion wor k: Com- municating affect in text m essages. Jou rnal of Language and Social Psychology , 36(5):549 – 567, 2017. 2 [74 ] Theo Ryffel, Andrew Trask, Morten Dahl, Bobby Wagner, Jason Mancuso, Daniel Rueckert, and Jonathan Passerat- Palmbach. A generic framework for privacy preserving deep learning. arXiv prepri nt arXiv:1811.04017 , 2018. 6 [75 ] John Sabini and Matthew C. Green. Emotional responses to sexual and emotional infidelity: Constants and differences across genders, samples, and methods. Personality and So- cial Psychology Bulletin , 30(11):1375 – 1388, 2004. 2 [76 ] Dairazalia Sanchez-Cortes, Oya Aran, Marianne Schmid Mast, and Daniel Ga tica-Perez. A nonverbal behavior ap - proach to identify emergent leaders in small g roups. IEEE transactions on multimedia , 14(3):816 – 832, 2011. 4 [77 ] Flor ia n Sch au b, R ebe cc a Bal eba ko, Ada m L Dur it y, a nd Lor - rie Faith Cranor. A design space for effective privacy no - tices. In Eleventh symposium on usable privacy and security (SOUPS 2015) , pages 1 – 17, 2015. 7 [78 ] Morgan Klaus Scheuerman, Jacob M Paul, and Jed R Brubaker. How comp uters see gender: A n evaluation of gender classification in commercial facial analysis services. Proceedings of t he ACM on Huma n-Computer Interaction , 3 (CSCW):1 – 33, 2019. 5 [79 ] Bjo¨rn Schuller, Stefan Steidl, and Anton Batliner. T he IN - TERSPEECH 2013 computational paralinguistics challenge. In Proceedings of INTERSPEECH 2013 , p ages 148 – 152, 2013. 3 [80 ] Secure Privacy. Building gdpr-c ompliant em otion recog- nition. Secure Privacy Blog , 2025. Retrieved from https://secureprivacy.ai/blog/gdpr-compliant-emotion- recognition. 7 [81 ] Vishal Sharma, Bonnie Nardi, Juliet Norton, and AM Tsaasan. To wards safe spaces online: A study of indian mat - rimonial websites. In IFIP Confer ence on Hum an-Computer Interaction , pages 43 – 66. Springer, 2019. 1 [82 ] Robert L. Shotland an d Judith M. Cra ig. Can men an d wom e n diffe re nt iat e betwe e n frien dly and se xua ll y inte re st ed behavior? Social Psychology Quarterly , 51(1):66 – 73, 1 988. 1 , 2 [83 ] Nathalie A Smuha. Regulation 202 4/1689 of the eur. parl. & council of june 13, 2024 (eu artificial intelligence act). International Leg al Materials , 64(5):1234 – 1381, 2025. 7 [84 ] Luke Stark . Facial recognition is the plu tonium of ai. XRDS: Crossroads, The ACM Magazine fo r Students , 25(3):50 – 55, 2019. 7 [85 ] Janet K. Swim and Lauri L. Hyers. Excuse me — what d id you just say?!: Wo men’s pub lic and private responses to sex- ist remarks. Journal of Exp erimental Social Psychology , 35 (1):68 – 88, 1999. 2 [86 ] Y-I Tian, Takeo Kanade, and Jeffrey F Cohn. Recognizing action units for facial ex pression analysis. IEEE Transac- tions on pattern analysis and machine intelligence , 23(2): 97 – 115, 2001. 3 [87 ] Deborah L. Tolman. Dilemmas of desire: Teenage girls talk about sexuality . Harvard University Press, Cambri dge, MA, 2002. 2 [88 ] Ca ta lin a L. Tom a, Jeff re y T. Ha ncoc k, an d Ni col e B. El li son . Separating fact fro m fiction: An examination of deceptive self-presentation in o nline dating profiles. Personality and Social Psychology Bulletin , 34(8):1023 – 1036, 2008. 1 11 [89 ] Dominic A Trevisan, Marleis Bowering, a nd Elina Birming- ham. Alexithymia, but not autism spectrum disorder, may be relate d to the production of emotional facial expressions. Molecular autism , 7(1):46, 2016. 5 [90 ] Gareth Tyson, Vasile C. P erta, Hame d Haddadi, and Michael C. Seto. A first look at us er activity on tinder. 2016. arXiv:1607.01952 [cs.SI], submitted 7 Jul 2016. 1 [91 ] Joseph B Walther. Interpersonal effects in computer- mediated interaction: A relational perspective. C ommuni- cation research , 19(1):52 – 90, 1992. 2 [92 ] Joseph B Walther. Computer-mediated communication: Im - personal, i nterpersonal, and hyperpersonal interaction. Com- munication research , 23(1):3 – 43, 1996. 2 [93 ] Jacob Whitehill, Zewelanji Serp ell, Yi-Ching Li n, Aysha Foster, and Javier R. M ovellan. The faces of engagement: Automatic recognition of st udent engagementfrom facial ex - pressions. IEEE Transactions on Affec tive Computing , 5(1): 86 – 98, 2014. 3 [94 ] Peter Wittenburg, Hen nie Brugman, Albert Russel, Alex Klassmann, and Han Sloetjes. ELAN: a pro fessional frame- work for m ultimodality research. In Proceedings of the Fifth International Conference on Language Resources and Ev al- uation (LREC’06) , Genoa, Italy, 2006. European Langu age Resources Association (ELRA). 5 [95 ] Allison Woodruff, Vasyl Pihur, Su nny C onsolvo, Laura Brandimarte, and Alessandro Acquisti. W ould a privacy fun- damentalist sell their dna for $1000... if nothing bad hap - pened as a result? the westin categories, b ehavioral in ten- tions, and consequences. In 10th Symposi um on Usable Pri- vacy and Security (SOUPS 2014) , pages 1 – 18, 2014. 7 [96 ] Tian Xu, Jennifer White, S inan Kalkan, and Hatice Gunes. Investigating bias and fairness in facial expression recogni - tion. In European Conference on Computer Vision , pages 506 – 523. Springer, 2020. 5 [97 ] Xucong Zhang, Yusuke Sugano, Mario Fritz, and Andreas Bulling. Appearance-based gaze estimation in the wild. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , pages 4511 – 4520, 2015. 1 , 3 [98 ] Shoshana Zuboff. The age of surveillance capitalism: The fight f or a human future at the new frontier of power. Journal of Information Ethics , 33(1):84 – 85, 2024. 6 [99 ] Douglas Zytko an d Nicholas Furlo. Online dating as context to design sexual consent tec hnology with w omen and lgbtq+ stakeholders. In Proceedin gs of the 20 23 CHI Conference on Hum an Fac to rs in C om puti ng Sys te ms ( CH I ’ 23) , Ne w Yor k, NY, USA, 2023. Association for Computing Machinery. 4