Description
Give an LLM classic judgement problems with known normative answers, one per fresh session, score each answer for the expected bias, and repeat later and across models to check stability.
Large language models answer judgement problems in fluent, confident prose, which can make biased reasoning look authoritative. This battery borrows problems from the heuristics-and-biases tradition, each with a normative answer: representativeness, insensitivity to sample size, base-rate neglect (with and without numbers), anchoring, and gain versus loss framing. Each problem is given in a new session after a fixed opener that asks the model to answer as a person would, without references or qualifiers, so answers are comparable and not contaminated by earlier turns. Answers are scored against the normative solution and coded by error type, and the battery is rerun months later and across models because behaviour changes with updates. In the original tests most models showed most biases, and some patterns changed between runs, which is itself a finding for teams relying on a model.
Attributes
- Discipline
- Computational & Machine Learning
- Loss type
- Latent
- Stage
- Evaluation, Post-launch
- Data
- Mixed
- Effort
- Days
Steps
- Select problems with a known normative answer covering several biases, for example representativeness, sample size, base-rate neglect, anchoring and framing.
- Write a fixed opener that sets the answering stance (answer as a person, no references, one answer without qualifiers, guessing allowed) and a fixed wording for each problem.
- Start a new model session for every problem so that earlier answers cannot influence later ones.
- Record the full output, then score it against the normative answer and code the error type (for example reporting incidence or sensitivity instead of predictive value, or reversing preference across frames).
- Run several repetitions per problem where sampling is non-deterministic, and summarise the proportion of biased answers.
- Repeat the battery after model updates and across competing models, and report a bias-by-model-by-date matrix with annotated transcripts.
Sources
- Challenging the Appearance of Machine Intelligence: Cognitive Bias in LLMs and Best Practices for Workplace Adoption (Talboy & Fuller, 2023)
Also known as
- Cognitive bias testing of LLMs
- LLM cognitive bias battery
- Machine psychology bias probes
Categories
- Analytics & Behaviour
- Computational & Machine Learning