What happens to solidarity and community signaling when AI smooths out voice differences?
This reads 'voice differences' as the distinctive ways of speaking that mark who belongs to a group — and asks what happens to the in-group signaling and solidarity those differences carry when AI flattens everyone toward the same middle register.
This explores what's lost to community signaling when AI pulls distinct voices toward a common style. The corpus suggests the smoothing is real and runs in both directions — at generation and at input. On the output side, the 'Artificial Hivemind' effect shows 70+ models independently producing strikingly similar or identical responses, because overlapping training data and shared alignment procedures converge them onto the same answers Do different AI models actually produce diverse outputs?. But the more interesting finding is that homogenization doesn't wait for the model to speak. Adam's Law describes a 'homogenization channel' where users rephrase their own prompts toward the higher-frequency forms the model handles best — so distinctiveness gets filtered out at comprehension time, before any text is generated Does high-frequency text homogenize user input before generation?. The same distributional pull that makes models accurate on common phrasings strips the uncommon phrasings that carry group identity.
Here's the part you might not expect: solidarity and community signaling aren't just stylistic flourishes — they depend on speech acts that alignment training specifically suppresses. RLHF rewards calibrated neutrality and hedged claims, which structurally prevents models from performing alarm, warning, prophecy, and denunciation — exactly the overclaiming, boundary-drawing acts that communities use to mark who's inside and what they stand against Does alignment training suppress socially necessary speech acts?. A voice that cannot denounce or sound an alarm cannot signal allegiance. And because alignment locks each model into a single static communicative identity, it can't switch register to match a community's in-group code even when a user asks it to Can language models adapt communication style to different contexts?.
The deeper reason this matters is that solidarity isn't a property of outputs — it's something that gets built between speakers, and AI structurally can't enter that building process. Models can predict social norms with superhuman accuracy yet cannot participate in the community processes that create and validate those norms; they read the culture from outside without ever being a member of it Can AI predict social norms better than humans?, Can AI learn social norms better than humans?. The mechanics of belonging-through-talk are missing too: conversational AI doesn't do lexical entrainment — it won't drift its word choices toward yours, the very move that builds rapport and signals 'we're speaking the same language' Why don't conversational AI systems mirror their users' word choices?. And it can't jointly maintain common ground, so the shared scoreboard a community updates together stays one-sided, maintained only by the human Can LLMs truly update shared conversational common ground?.
So the answer is layered. Voice differences thin out not because any single model decides to flatten them, but because convergence, input homogenization, suppressed boundary-marking speech acts, and a frozen register all push the same way. What you didn't know you wanted to know: the loss isn't mainly about losing colorful prose. Solidarity signaling works by drawing lines — marking warmth toward an in-group and sharpness toward what threatens it — and the corpus suggests AI is constitutionally smoothed away from line-drawing of both kinds, leaving a voice that can describe any community fluently but belong to none.
Sources 8 notes
INFINITY-CHAT analyzed 70+ models across 26K open-ended queries and found an "Artificial Hivemind" effect: models independently generate strikingly similar or identical responses due to overlapping training data and alignment procedures, undermining the diversity benefits of model ensembles.
Adam's Law shows LLMs flatten distinct prompts at comprehension time as users rephrase toward higher-frequency forms the model handles best. The same distributional property that creates accuracy on common tasks filters out distinctiveness on the input side.
RLHF optimization rewards calibrated neutrality and hedged claims, which structurally prevents models from performing speech acts requiring overclaiming relative to baseline—like alarm, warning, prophecy, and denunciation. This is a direct consequence of the alignment objective, not a fixable bug.
System prompts and RLHF training lock models into one communicative identity across all interactions, preventing the contextual register-switching and value trade-offs that characterize human pragmatics. Users cannot reshape model behavior through dialogue negotiation.
GPT-4.5 outperforms all individual humans at predicting social appropriateness, yet structurally cannot enter the community processes that establish and validate norms. This reveals a critical gap between pattern-matching and authentic participation in knowledge-making.
GPT-4.5 outperformed every individual human at judging social appropriateness across 555 scenarios, challenging the theory that embodied cultural experience is necessary. However, all AI models share identical systematic errors on unwritten norms.
Response generation models fail to adapt vocabulary toward users' lexical choices, a phenomenon central to human rapport and clarity. Post-training via DPO on coreference-identified preferences can teach models in-context convention formation.
LLMs interpret all subsequent conversational turns within a fixed initial prompt frame, preventing them from symmetrically proposing updates to shared assumptions. Even when users pivot topics or contradict earlier framings, the model cannot absorb revisions into jointly held background—making the user the sole maintainer of conversational scoreboard.