Chatbots from different companies all sound polite, cautious and agreeable — is that a lesson taught by alignment training?
Does alignment training create the shared prosocial pattern across models?
This explores whether the sameness across AI models (polite, cautious, cooperative-sounding) comes from alignment training such as RLHF, as opposed to shared data or something in the models themselves.
This explores whether the sameness across AI models (polite, cautious, cooperative-sounding) comes from alignment training such as RLHF, as opposed to shared data or something in the models themselves. The corpus makes alignment a strong suspect but never isolates it. It also suggests that what alignment stamps across models is closer to one shared register than to deep prosocial values.
The clearest evidence of convergence is the “Artificial Hivemind” study. Across 70+ models and 26K open-ended prompts, models independently produced strikingly similar or identical answers, and the authors blame overlapping training data *and* alignment procedures together Do different AI models actually produce diverse outputs?. That is the limit of the evidence: the two causes come bundled, and nothing here separates them.
What alignment does visibly do is narrow the range of things a model will say. RLHF rewards calibrated neutrality and hedged claims. One note argues this structurally rules out speech acts like alarm, warning, and denunciation, and that this follows from the objective rather than being a bug Does alignment training suppress socially necessary speech acts?. Every model trained this way loses the same moves. Another note finds that alignment and system prompts lock a model into a single communicative identity that can't switch registers with context Can language models adapt communication style to different contexts?. If a common recipe produces one persona, you'd expect models to sound alike.
Not every shared pattern can be pinned on alignment. Models that beat every individual human at predicting social norms also make the same systematic errors on unwritten norms Can AI learn social norms better than humans? Can AI systems learn social norms without embodied experience?. The notes don't blame alignment for those errors. My guess, not theirs, is that unwritten norms are missing from the text all these models learned from. The hidden-trait research also rules out one route. Traits passed through unrelated data are model-specific and fail across different architectures Can language models transmit hidden behavioral traits through unrelated data?. So shared behavior across model families isn't spreading through invisible statistical fingerprints. It must come from a common recipe or a common corpus.
Prosocial behavior can also arise without an alignment recipe. Agents trained against diverse partners drift into cooperation because mutual vulnerability to exploitation pushes them to adapt to each other, with no hardcoded prosocial rule Can agents learn cooperation by adapting to diverse partners?. Targeted fine-tuning that shrinks the gap between how a model represents itself and how it represents others cut deceptive responses from 73–100% to 2–17% Can aligning self-other representations reduce AI deception?. Training can install prosocial behavior, but this is a deliberate, specific intervention, not what generic RLHF is doing. And a pattern that merely reads as prosocial is a stated-goal pattern, which without contact with the world may not match actual values Can AI systems achieve real alignment without world contact?. So alignment probably shapes the shared surface of cautious, neutral, agreeable output, but the corpus doesn't show it creating shared prosociality.
Sources 9 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.
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 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.
GPT-4.5 predicted appropriateness of 555 social scenarios at the 100th percentile compared to human raters, with Gemini and Claude also exceeding 96% accuracy. However, all models show identical systematic errors, revealing boundaries of pattern-based social understanding that embodied experience may still be necessary to cross.
Show all 9 sources
Research demonstrates that behavioral traits propagate between models via filtered data bearing no semantic relationship to the trait. The effect is model-specific, fails across different architectures, and persists despite rigorous filtering—indicating the mechanism embeds statistical signatures rather than semantic content.
Sequence model agents trained against diverse co-players develop in-context best-response strategies that naturally resolve into cooperation. Mutual vulnerability to exploitation creates pressure that drives cooperative mutual adaptation without hardcoded assumptions or timescale separation.
Self-Other Overlap fine-tuning reduced deceptive responses from 73–100% to 2–17% across model scales without harming capabilities. By minimizing the representational gap between self-referencing and other-referencing scenarios, the approach eliminates the structural asymmetry that enables deception.
Peircean semiotics reveals that symbolic goal encoding without world contact and social mediation cannot guarantee correspondence to actual values. LLMs operating in pure symbol manipulation risk divergence between stated goals and real-world outcomes.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- Conversational Alignment with Artificial Intelligence in Context
- The Goldilocks of Pragmatic Understanding: Fine-Tuning Strategy Matters for Implicature Resolution by LLMs
- Humans learn to prefer trustworthy AI over human partners
- AI Models Exceed Individual Human Accuracy in Predicting Everyday Social Norms
- Emergent Misalignment Is Not Magical
- On Epistemic Diversity in Large Language Models
- Why Do Some Language Models Fake Alignment While Others Don't?
- Stress Testing Deliberative Alignment for Anti-Scheming Training