Confirmation Bias
A cognitive psychology concept describing the tendency of individuals to seek out and focus on information that supports their existing beliefs while ignoring contradictory evidence. In AI interaction contexts, this bias can be amplified by large language models, which may reinforce pre-existing user assumptions through their generated responses.
Core Facts
Confirmation bias was first systematically demonstrated by psychologist Peter Wason in the 1960s through the classic 2-4-6 task experiment.
Confirmation bias refers to the tendency of people to seek, interpret, and remember information that is consistent with their existing beliefs while ignoring contradictory evidence.
Timeline (last 90 days)
The formula deliberately omits productivity gains brought by AI tools because such gains are difficult to decouple from the tool versus engineer capability, and including them would introduce a confirmation bias entry point.
When people confide emotional difficulties to AI, they are more likely to accept advice consistent with their own narrative framework, which may reinforce confirmation bias.
The solution is not to skip testing, but to ensure the builder is not the sole inspector.
The reviewer role in multi-agent collaboration is typically implemented as an independent Agent instance, running in isolation from the code-generating Agent to avoid confirmation bias.
Splitting the producers and the critics into two separate groups is a key design decision, as having the same model both generate and review code produces confirmation bias.
Large language models tend to defend their own outputs rather than critically scrutinize them during self-evaluation, a phenomenon researchers call self-consistency bias.
Concepts related to motivated reasoning include confirmation bias and conflict of interest.
The driving forces behind the spread of misinformation include information cascade effects, confirmation bias, and novelty bias.
Confirmation bias was first systematically demonstrated by psychologist Peter Wason in the 1960s through the classic 2-4-6 task experiment.
All Facts (12)
Confirmation bias was first systematically demonstrated by psychologist Peter Wason in the 1960s through the classic 2-4-6 task experiment.
65%VerifiedConfirmation bias refers to the tendency of people to seek, interpret, and remember information that is consistent with their existing beliefs while ignoring contradictory evidence.
65%Unverified谄媚的AI可能充当'确认偏误放大器'，强化用户的认知偏差
75%UnverifiedConfirmation Bias drives users to continuously seek evidence that supports their own negative judgments
60%UnverifiedThe formula deliberately omits productivity gains brought by AI tools because such gains are difficult to decouple from the tool versus engineer capability, and including them would introduce a confirmation bias entry point.
50%UnverifiedWhen people confide emotional difficulties to AI, they are more likely to accept advice consistent with their own narrative framework, which may reinforce confirmation bias.
50%UnverifiedThe solution is not to skip testing, but to ensure the builder is not the sole inspector.
50%UnverifiedThe reviewer role in multi-agent collaboration is typically implemented as an independent Agent instance, running in isolation from the code-generating Agent to avoid confirmation bias.
50%UnverifiedSplitting the producers and the critics into two separate groups is a key design decision, as having the same model both generate and review code produces confirmation bias.
50%UnverifiedLarge language models tend to defend their own outputs rather than critically scrutinize them during self-evaluation, a phenomenon researchers call self-consistency bias.
50%UnverifiedConcepts related to motivated reasoning include confirmation bias and conflict of interest.
50%UnverifiedThe driving forces behind the spread of misinformation include information cascade effects, confirmation bias, and novelty bias.
50%