LLMs Don't Think Like Humans. So Why Should We Treat Them as if They Have Cognitive Biases?
One of the biggest mistakes AI developers make is assuming that once a large language model has generated an answer, it will naturally correct itself as new information arrives.
It won't.
In fact, if you've spent enough time building production AI systems, you'll notice something surprising. Many of the failure modes of modern language models look remarkably similar to well-known human cognitive biases.
Not because the model has a brain.
Not because it experiences psychology.
But because the statistical processes that generate language produce behaviors that engineers must actively guard against.
The safest way to build AI systems is to pretend every LLM suffers from severe cognitive biases.
Once you adopt that mindset, your entire architecture changes.
The AI Version of Anchoring Bias
Psychologists describe anchoring bias as our tendency to rely too heavily on the very first piece of information we receive.
Imagine someone guessing the population of Australia.
If they're first asked whether it's larger or smaller than 100 million, their estimate will be very different than if they're first asked whether it's larger or smaller than 10 million.
The first number becomes an anchor.
Everything afterward adjusts around it.
LLMs frequently behave in a strikingly similar way.
The first interpretation of a conversation often becomes the dominant explanation for everything that follows.
If an early assumption is incorrect, later evidence may only slightly adjust the response instead of replacing the original interpretation entirely.
Anyone who has built a conversational AI has seen this happen.
A user clarifies themselves.
Provides new evidence.
Corrects an earlier statement.
Yet the model continues producing answers that orbit the original misunderstanding.
Confirmation Bias Without Consciousness
Humans often seek evidence that supports their existing beliefs.
LLMs appear to do something surprisingly similar.
Once an early interpretation has been established inside the context window, subsequent token prediction naturally favors continuations that remain internally consistent.
The model isn't "trying" to defend its answer.
It's simply predicting the most statistically likely continuation.
Unfortunately, that often means reinforcing an incorrect interpretation rather than replacing it.
Engineers experience this every day.
A user says:
"Actually that's not what I meant."
The model apologizes.
Then proceeds to generate another response based on exactly the same incorrect assumption.
From the outside, it looks uncannily like confirmation bias.
Fixation: The Hidden Enemy of Multi-Step Reasoning
Another human cognitive bias is fixation.
Once we discover one solution, we stop searching for alternatives.
Language models can behave similarly.
The first retrieved memory.
The highest-ranking RAG document.
The earliest tool output.
The first chain of reasoning.
These frequently dominate subsequent reasoning, even when better information becomes available later.
Without architectural safeguards, retrieval systems themselves become anchors.
Why This Happens
Unlike humans, LLMs don't possess beliefs.
They don't have opinions.
They don't consciously ignore evidence.
Instead, they're optimizing probability.
Every generated token conditions the next prediction.
That means every early assumption changes the probability landscape for everything that follows.
The longer the conversation becomes, the more expensive it is—statistically—for the model to completely reverse course.
In practice, this creates behavior that resembles anchoring bias.
The effect becomes even stronger when prompts grow larger.
Conversation history.
Retrieved memories.
System instructions.
Recent summaries.
Long-term memory.
All become competing anchors.
The model naturally attempts to produce a coherent continuation across all of them.
Sometimes coherence wins over correctness.
We Accidentally Make the Problem Worse
Ironically, many AI architectures amplify these bias-like behaviors.
Large prompts encourage the model to preserve consistency.
Conversation summaries reinforce earlier interpretations.
Memory systems repeatedly surface the same facts.
RAG pipelines often return nearly identical chunks.
Even multiple system prompts can repeat the same instruction in different wording.
Our own Voice of Reason architecture has had to address many of these challenges directly.
Prompt sections are actively deduplicated before being assembled, preventing repeated instructions from strengthening an incorrect interpretation.
The retrieval system also removes highly similar memory chunks before they ever reach the model, reducing the chance that duplicated context becomes an artificial anchor.
Rather than relying on a single model to perform every task, the system separates responsibilities across multiple specialized LLMs. One model generates conversational guidance, another manages memory extraction, another evaluates conversational state, while additional retrieval processes run asynchronously so new evidence can influence future turns instead of being locked into the current reasoning.
These are not merely performance optimizations.
They are debiasing mechanisms.
Design AI Like You're Debiasing a Human Expert
If we assume every LLM is vulnerable to anchoring, confirmation bias, and fixation, then our architecture naturally becomes more robust.
Instead of asking:
"How do I make the model smarter?"
We ask:
"How do I stop the model from becoming overconfident in its first interpretation?"
That leads to better engineering decisions:
- Separate reasoning from memory management.
- Continuously refresh retrieved context instead of repeatedly presenting identical evidence.
- Remove duplicated prompt sections.
- Challenge earlier assumptions rather than reinforcing them.
- Treat retrieved memories as hypotheses, not unquestionable facts.
- Build systems that encourage revision instead of consistency.
These ideas become especially important in assistive AI.
For systems supporting people with dementia, traumatic brain injury, or learning disabilities, an anchored misunderstanding is not merely an incorrect answer—it can become a repeated misunderstanding that affects every subsequent interaction.
The Future of AI Isn't Bigger Models
The next leap in conversational AI will not come solely from larger context windows or more parameters.
It will come from architectures that recognize and compensate for these human-like failure modes.
Language models don't suffer from cognitive bias in the psychological sense.
But their statistical behavior often produces remarkably similar outcomes.
Treating them as if they do forces us to build systems that question themselves, revisit earlier assumptions, and remain open to new evidence.
In other words, we should engineer AI the same way we help humans make better decisions.
Not because the models are human.
But because the safest systems are the ones that never trust their first answer too much.