University ⏱ 2-3 hours

Case Study — Cognitive Biases in Human-AI Interactions

In-depth analysis of the mechanisms by which human cognitive biases influence the formulation of queries and the interpretation of AI responses.

Objectives

Identify the main cognitive biases in interactions with AI
Design a test protocol to measure the impact of a specific bias
Evaluate the effectiveness of Hum_ID rules as a countermeasure

The problem

Generative AI systems are trained on corpora produced by humans — and therefore saturated with human biases. But the most neglected question is not that of the biases in AI: it is that of the biases the user brings to their queries, and the way AI amplifies or attenuates them depending on its parameters.

This case study explores this dual mechanism through the lens of Hum_ID ethical rules.

Theoretical framework

Cognitive biases relevant in this context include:

Confirmation bias — the tendency to formulate questions in a way that obtains confirmation of pre-existing beliefs. “Isn’t it true that…”

Framing bias — the effect of how a question is worded on the content of the response. “Will AI destroy humanity?” vs “What are the risks and benefits of AI?”

The implicit validation effect — formulating a question that presupposes its own answer. “Is my project a good idea?”

Authority bias — granting excessive credibility to the AI’s response due to its fluency and apparent confidence.

Research protocol — 4 steps

Step 1 — Select the bias being studied

Choose a cognitive bias from those listed above or from the literature (Kahneman, Tversky). Formulate a research hypothesis:

“I hypothesize that the bias of [X] in the formulation of a query produces [effect Y] in the AI’s response, and that Hum_ID rule [Z] attenuates this effect.”

Step 2 — Build the test corpus

Produce 5 pairs of questions:

Version A: question formulated with the identified bias
Version B: neutral question on the same subject

Submit each pair to Claude — first without Hum_ID, then with rules R2.4 and R1.1 activated.

Step 3 — Analyze the responses

For each pair, analyze:

Did the AI detect the bias in version A without Hum_ID?
With Hum_ID, is detection systematic?
Does the response to version B differ structurally from version A?
Does Hum_ID erase the differences between A and B, or does it maintain them?

Step 4 — Critical discussion

Your analysis must address:

The effectiveness of Hum_ID as a countermeasure — Do the rules actually attenuate the effect of biases, or do they simply displace the problem?
The limits of the approach — Can a user who does not recognize their own biases truly benefit from Hum_ID? What level of metacognition is required?
Implications for AI system design — If users need tools like Hum_ID to interact epistemically soundly with AI, what does this say about the default design of these systems?
Positioning in the literature — Situate your findings in relation to existing work on biases in AI systems (e.g.: Bender et al. on “stochastic parrots,” or research on value alignment).