Cross-substrate phenomena — in plain English
What humans and language models share, and where they part ways
Confirmation bias shows up in a language model. Loss aversion does not. And that is where it gets interesting. When a trait we thought was deeply human also turns up in a machine with no body, it cannot be about the body. Then it is something basic about choosing under pressure. And what is found only in us, like loss aversion, reveals what a body that can die adds on top.
| Found in both humans and language models | Found only in humans (and other biological substrates) |
|---|---|
| Anchoring · Hysteresis · Confirmation bias · Dunning-Kruger · Mode-shift cost (task-switching) · 1/e timing (37% commitment point) · Expertise reversal · Surprise-weighted encoding · Information overload · Reactance · Inverted-U on challenge | Loss aversion (requires mortality) · Spaced repetition (requires between-session memory) · Field-organised friction by Safety / Meaning / Capability / Effort |
The ones on the left follow simply from having to choose under pressure, so they show up in a language model too. The ones on the right need a body that can die, move, and get tired.
What they share
Here they are, measured directly in language models:
Anchoring — the first answer locks the rest in place
When humans decide on something, the first number or option they consider has disproportionate influence over the final result. For example, if you ask both a human and a language model to estimate the price of a car after first hearing a high number, both tend to guess higher than if they heard a low number first. Language models do exactly the same thing: the first words of a response set the "anchor" for everything that follows. On Qwen2.5-32B, 90% of the response is generated from a plateau established in the first 10%. A corresponding anchor pattern is directly visible in the friction topology.
Hysteresis — traces of the past make reversal expensive
Once a system has "invested" in one path, it costs something to turn back. Everyday analogy: if you start walking down one path in a forest, turning back and choosing a different path feels harder the further you go. This applies to magnetic materials, biological decisions, and language models. I have measured hysteresis across three transformer architectures (Cogito-671B, Qwen3-235B, Llama-3.3-70B) plus partial coverage on Qwen2.5-32B, and compared it with a completely different architecture (LiquidAI LFM2, a state-space model). The behaviour is the same, which indicates that hysteresis is a property of the architecture itself, not of any specific implementation.
Confirmation bias — the cost of admitting you were wrong
Humans tend to maintain their first position because reversing is costly. Language models show the same pattern. More interestingly: a large model (Cogito-671B) actually produces stronger arguments against its own answer than for it, because it was trained to find flaws (the assistant-training step known as RLHF). That is reverse confirmation bias as a side-effect of that training.
Dunning-Kruger — confident before competent
You are most certain when you know least. That curve shows up in a language model as it learns, and because a machine has no ego, the effect cannot be human vanity. It is measured directly in four models. See the plain-English walkthrough →
"1/e timing" — wait until 37% before deciding
Classical mathematics (the secretary problem) says that if you must choose from a sequence of candidates, you should reject the first ~37% and then take the first one better than all those rejected. On the largest matched base-instruct test, Qwen2.5-32B base, the model lands at 39.3%, close to 1/e. Smaller base models typically land between 43-48%, so clean 1/e convergence is most evident on the largest base models. The pattern is consistent with the theory's prediction but not as a universal law; assistant-tuned versions decide about 9 percentage points later.
Mode-shift cost (task-switching)
When humans switch between tasks or modes of thinking, it costs something. Psychologists call it "task-switching cost". Language models show exactly the same effect, and it is localised to the first 5 words of the response. The effect replicates across different model architectures.
Surprised words get more attention
When something unexpected happens, the brain remembers it better. It is part of why surprising advertisements stick. On a small model (Qwen2.5-0.5B) I have measured that the most surprising words get about 34% more attention from the words that follow than the least surprising ones do. The effect is small but statistically solid. It is the first time anyone has measured the mechanism directly in an artificial system. The next step is to test it on other models (Llama, Mistral).
Expertise reversal — help for novices hurts experts
Research by John Sweller and others shows that instructional supports (worked examples) help novices but hurt experts. This is because experts have already internalised the pattern, and additional examples now compete with their internal model. The same effect appears in language models, but only at the largest scales: Llama-3.3-70B shows the classical U-curve (73% → 50% → 61% with 0, 1 and 3 examples; paper on the way), while smaller models cannot show the effect because they lack the capacity.
What they do NOT share
These phenomena exist in humans but NOT in language models, and that is not a bug. It is because they require something language models lack.
Loss aversion — we fear losses twice as much as we enjoy gains
Kahneman and Tversky showed humans are about 2x more averse to losing $100 than they are pleased to gain $100. This is not in language models. Why? Because loss aversion fundamentally comes from the fact that you can die if you choose wrong too often. Humans have mortality; language models do not. My measurement: language models decide later than 1/e, the opposite of what loss aversion predicts. It is a new explanation for an old finding.
Memory between conversations
Every time you start a new conversation with a language model, it starts over. It has no memory of what you talked about last time. This means an entire field of human psychology, the consolidation of memory over time, literally cannot be tested on language models. Such studies require fine-tuning that changes the model's weights across training sessions.
Spaced repetition (Ebbinghaus 1885)
We remember better if we repeat information with breaks in between (this is the basis for language apps like Duolingo). The mechanism requires remembering across sessions, which language models do not. Listed as future work in my papers, testable only via fine-tuning.
Fields: Safety, Meaning, Ability, Effort
My behavioural friction theory describes how living beings organise decisions through four fields. Those fields do NOT exist in language models: they lack mortality (no safety needs), movement (no ability field), and energy cost (no effort field). Language models have friction without field organisation. A statistical analysis of 15 models confirms it: three dimensions (magnitude, distribution, rhythm), but no field structure.
The pattern
What is structurally necessary for any system that must choose under constraints (anchoring, hysteresis, mode-shift, surprise-weighting, expertise reversal) appears in both humans and language models.
What requires specific biological features (mortality, embodiment, between-session memory) appears only in humans. Language models confirm the predictions by their absence.
Friction itself is always there. But how it gets organised depends on the material. Language models let us see friction in raw form for the first time, without the evolutionary layers that make it hard to spot in living beings.
Numbers, references and tables are in the technical version: phenomena (technical).