← narwal.one/Second Brain
SecondBrain
Ask the Brain
Index/Conceptupdated Wed Nov 04 2026 08:00:00 GMT+0800 (Philippine Standard Time)

Typicality Bias

typicality-biasrlhfpost-trainingmodel-convergenceaveraging-machineai-slophomogenizationsycophancy

Typicality Bias

The post-training mechanism behind the sameness of AI output, as relayed by Nufar Gaspar in How to Deslopify Your AI (AI Daily Brief): after pre-training, people rate the model's answers, and raters reward what feels familiar. The term comes from Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity (Zhang, Yu, Chong, Sicilia, Tomz, Manning, Shi; Northeastern, Stanford and West Virginia University; arXiv 2510.01171, October 2025; verified 2026-11-04), which defines it as the human tendency to prefer more typical text and argues that because the bias sits in the preference data itself, mode collapse affects most alignment methods that rely on reward models. Their fix is a training-free prompt: ask for several responses with their probabilities. Folded into the model through preference tuning, it narrows the range of answers and produces the default house style everyone now recognises: long, listy, structurally identical, because length looks like effort and effort is what raters pay for.

It sits on top of the base mechanism, which is that a language model predicts the most likely continuation, so a generic prompt returns roughly the average of what the internet thinks. Pre-training supplies the average; typicality bias in post-training sharpens the pull toward it. HBR's Scott Berinato calls the result "an averaging machine" (Authenticity Crisis (AI Communication)).

Evidence that it is structural, not vendor-specific

  • Cross-vendor convergence. The NeurIPS 2025 Best Paper Artificial Hivemind: The Open-Ended Homogeneity of Language Models (and Beyond) (Liwei Jiang, Yejin Choi et al., University of Washington; verified 2026-11-04) evaluated 70+ models on open-ended prompts and found both intra-model repetition and inter-model homogeneity across companies. Switching tools therefore buys no originality.
  • The incognito experiment. ChatGPT, Claude Opus and Gemini (latest versions, memory and history off) given the same generic prompt (write a LinkedIn post about how AI is changing marketing) returned near-identical structure and content. The same prompt with the author's point of view, experience, structural preferences and banned words produced a materially better draft.
  • Image models carry a house style of their own (the too-golden light in every AI food ad) and fall back to it unless told otherwise.
  • Persistent stylometric tells. The Economist's 14-model study found sycophancy, parallelism, tricolons and contrast constructions survive even as em-dashes and "leveraging" are trained away (AI Writing Detection). Those are the shapes typicality bias rewards.

Why it matters

  • It explains why context and memory do not fix slop: they tell the model who you are, not what you think today. Originality has to enter through the brief, which is the whole of Stage 1 in the Deslopifying Playbook and the point of Front-Loading the Brief.
  • It is a sibling of the sycophancy problem tracked at Productive Resistance and Designing AI Products That Don't De-Skill Users: the same reward signal that makes a model agreeable makes it typical. Both are products of what raters prefer.
  • Whittemore's demonstration that a model will steelman either side of any argument with full conviction is the same fact from another angle: a system optimised for typical, acceptable answers has no perspective of its own to lend.

Confidence note: both papers are now verified against their primary pages, though read at abstract level only; the incognito experiment is one practitioner's n=3 demonstration. Together with the Economist stylometry, that is three independent lines of evidence for the mechanism.

Sources