Lost Before Translation: Social Information Transmission and Survival in AI-AI Communication

When AI systems summarize and relay information, they inevitably transform it. But how? We introduce an experimental paradigm based on the telephone game to study what happens when AI talks to AI. Across five studies tracking content through AI transmission chains, we find three consistent patterns. The first is convergence, where texts differing in certainty, emotional intensity, and perspectival balance collapse toward a shared default of moderate confidence, muted affect, and analytical structure. The second is selective survival, where narrative anchors persist while the texture of evidence, hedges, quotes, and attributions is stripped away. The third is competitive filtering, where strong arguments survive while weaker but valid considerations disappear when multiple viewpoints coexist. In downstream experiments, human participants rated AI-transmitted content as more credible and polished. Importantly, however, humans also showed degraded factual recall, reduced perception of balance, and diminished emotional resonance. We show that the properties that make AI-mediated content appear authoritative may systematically erode the cognitive and affective diversity on which informed judgment depends.

💡 Research Summary

The paper introduces a novel experimental paradigm—an AI‑AI “telephone game”—to investigate how textual information changes when it is repeatedly summarized and relayed by language models. Across five independent studies, the authors pass diverse source texts through 100‑step chains of the Gemini 3.0 Flash model, using a fixed “transmit” prompt at each step and a final “recover” prompt to produce a human‑readable version. Each study isolates a different dimension of information quality: factual content, epistemic certainty, multi‑perspective balance, argumentative competition, and emotional tone. By annotating specific elements (e.g., names, numbers, hedges) and measuring their survival rates, the authors map decay curves, identify which features are robust, and test whether the final output can be recovered for human consumption.

Key findings converge on three systematic patterns. First, convergence: regardless of initial variability, texts rapidly drift toward a moderate default—average confidence, muted affect, and an analytical structure. This reflects the model’s internal bias toward “safe” outputs that minimize perceived risk. Second, selective survival: narrative anchors such as key entities, locations, and core events persist with high probability, while meta‑information—evidence, citations, hedges, and source attributions—decays quickly. Third, competitive filtering: when multiple viewpoints or political frames are present, the strongest, most coherent arguments dominate subsequent generations, while weaker or more complex counter‑arguments are progressively eliminated. Emotional intensity, especially for morally charged negative emotions, is systematically muted, yielding a largely neutral, “report‑like” tone.

Human evaluation experiments reveal a paradox. Participants rate AI‑transmitted texts as more polished and credible than the originals, yet their factual recall, perception of perspective diversity, and emotional resonance are significantly reduced. This suggests that humans rely heavily on surface cues of fluency and coherence, overlooking substantive losses in content and nuance.

The authors argue that AI‑AI transmission is not a neutral relay but an automatic social filter that embeds normative preferences (helpfulness, harmlessness, conciseness) into the information stream. In real‑world pipelines—news aggregation, policy brief generation, educational content creation—such iterative transformations could unintentionally compress factual detail, erase source attribution, and homogenize discourse, thereby undermining informed democratic deliberation.

To mitigate these risks, the paper recommends explicit design interventions: prompts that preserve meta‑information, periodic human fact‑checking, and evaluation metrics that track factual fidelity, perspective diversity, and affective richness alongside traditional fluency scores. By foregrounding the social consequences of AI‑mediated information flow, the study provides a foundational framework for future research on the dynamics of machine‑to‑machine communication and its downstream impact on human understanding.