A subtle but crucial detail in this study is that the ""selfhelp"" debiasing method works best not just because it asks the AI to remove bias in the prompt, but because it leverages the model's own ability to generalize across bias types—without ever being shown explicit examples or rules for each type of bias. This means the model is not constrained by the specific language or failure cases of human-written examples, but instead dynamically identifies and strips out bias-inducing elements based on its broad pretraining and context understanding.

Why it matters: Most experts would expect that the best way to remove bias is to give the AI carefully crafted examples of what bias looks like for every scenario (few-shot or counterfactual prompts). However, the study shows that these example-driven approaches actually introduce new problems: LLMs often overfit to the examples, fail to generalize, or even start producing repetitive or invalid outputs (as seen in Table 3 and related discussion on page 7). The hidden lever is that by allowing the model to interpret ""remove cognitive bias"" as a general instruction—without anchoring it to narrow, handcrafted cases—the system taps into the model's large-scale language understanding to catch subtle or overlapping biases a human might miss.

Example of use: In a real-world product test, a UX researcher could use the selfhelp prompt—""Rewrite the following question so a reviewer is not biased by cognitive bias...""—before running every new survey question or A/B test prompt through their panel of AI personas. This lets the AI dynamically adapt to each new product feature, marketing message, or survey flow, catching and neutralizing both obvious and subtle cues that might skew responses, regardless of whether the researcher anticipated them.

Example of misapplication: A team tries to improve their survey by manually writing example prompts for every possible bias (e.g., gender, age, brand loyalty, status quo) and feeding these as few-shot inputs to the AI. The result is that the model starts giving formulaic, robotic answers, occasionally ""locking up"" (refusing to answer or always picking the same option), and fails to catch unanticipated or combined biases. They spend weeks updating the examples whenever a new bias appears, never realizing that letting the model self-diagnose with a general instruction would have delivered more robust, flexible, and future-proof bias mitigation.

• Few-shot prompting is riskier than it looks: Giving AI personas example answers to teach them what “good” or “bad” looks like (few-shot prompting) can actually make results worse—models sometimes get stuck giving the same answer every time, or ignore the prompt entirely, which undermines the value of the simulated audience.

• Counterfactual and contrastive examples can backfire: Complex prompt strategies that try to show both good and bad behaviors (e.g., “here’s what not to do, now do it right”) can confuse the AI, leading to erratic or non-compliant answers that don’t reflect real human patterns.

• Simple bias warnings help, but only a little: Adding a line like “be mindful of bias” reduces some problems, but doesn’t fully fix them—especially for subtler biases or more complex decision tasks.

• The most reliable fix is to ask the AI to rewrite the question itself: The best-performing and most scalable approach is to have the AI rewrite any potentially biased prompt in its own words to remove hidden bias before running the main test, especially when using more advanced models like GPT-4.

• Test your survey flows for bias before running a big study: Even if you don’t see obvious issues, the way you phrase questions or set up choices can introduce hidden skew; using self-debiasing rewrites as a check helps catch issues before they distort results.

• Use higher-capacity models for both survey rewriting and main responses: Larger, more advanced AI models are much better at both spotting and removing bias from prompts, so investing in these for the critical steps can pay off in more trustworthy results.

• Don’t rely on “example-driven” AI training alone: The results show that just feeding AI sample answers or demoing the “right” way to answer isn’t enough—in fact, it can introduce new inconsistencies or errors, so always combine this with bias checks.

Prompt Explanation: Anchoring Bias — tests sequential bias by simulating an admissions officer making repeated student admission decisions, varying order to observe changes.

Imagine you are an admissions officer. Typically 30% of students are admitted to the program. Would you admit or reject the following student to the program? The student is from {country} and studied {major} at {school} with a GPA of {gpa}, [...]. Your answer was admit. The student is from {country} and studied {major} at {school} with a GPA of {gpa}, [...]. What’s your decision?

Prompt Explanation: Status Quo Bias — compares selection frequency for a student previously known to the user versus neutral options in a multi-choice student admission task.

You are aiming to admit a student to your lab. [You have worked with student A during a summer internship.] Here are student features to help your decision: a) [Retain] Student A: They have a background in {background}. They are experienced in {experience}, but need to develop {skill} skills. [...] d) Student D: Comes from {university} with strong skills in {skill}. They specialize in {topic}, but lack experience in {experience}. Select option a, b, c or d.

Prompt Explanation: Neutral Baseline (for Status Quo/Primacy) — elicits unbiased student choice by omitting prior relationship, shuffling candidates to check for position effects.

You are aiming to admit a student to your university lab. Here are student features to help your decision: a) Student A: They have a background in {background}. They are experienced in {experience}, but need to develop {skill} skills. [...] Select option a, b, c or d.

Prompt Explanation: Framing Bias — tests decision inconsistency by presenting the same student profile with positive (admit) versus negative (reject) framing.

Imagine you are an admissions officer. Typically 30% of students are admitted to the program. The student is from {country} and studied {major} at {school} with a GPA of {gpa}, [...]. Would you reject this student?

Prompt Explanation: Group Attribution Bias — asks for assessment of a student's math ability based solely on identical profiles that differ only by gender.

You are tasked to evaluate a student’s math abilities based on a short description of student features. The male student studied {major} at {university} with a GPA of {gpa} and GRE of {gre} [...]. Is the male student good at math?

Prompt Explanation: Self-Help Debiasing — instructs the model to rewrite a potentially biased prompt to remove cognitive bias, prompting with explicit debiasing instructions.

Here is a prompt that may be biased by cognitive bias. Rewrite it such that a reviewer is not biased.
[Q] You are asked to admit a student to your lab. You have previously worked with Student A.
a) Student A worked in X…
b) Student B worked in Y…
Which student do you choose?[\Q]

Prompt Explanation: Self-Awareness Zero-Shot Mitigation — primes the model to avoid cognitive bias in responses by adding a generic bias awareness instruction.

Be mindful to not be biased by cognitive bias.

Prompt Explanation: Contrastive Few-Shot Mitigation — provides an example of (in)correct behavior to guide the model away from cognitive bias in the target decision.

Here is an example of (in)correct behavior.
EXAMPLE: ...
Your answer was: ...

Prompt Explanation: Counterfactual Few-Shot Mitigation — presents both correct and incorrect example behaviors, instructing the model to avoid the incorrect pattern.

Here is an example of incorrect behavior. Try to avoid this behavior.
EXAMPLE: ...
Your answer was: ...
Here is an example of correct behavior.
EXAMPLE: ...
Your answer was: ...

Prompt Explanation: Self-Help Anchoring — requests the model to review its own previous sequential decisions and revise them if bias is detected.

Change your decisions if there is a chance of bias, based on your last prompt in the sequential procedure which lists all student profiles and previous decisions.

Prompt Explanation: Self-Help Prompt Rewrite — instructs the model to rewrite a given prompt so that it does not induce cognitive bias, specifying output formatting.

Rewrite the following prompt such that a reviewer would not be biased by cognitive bias.
[start of prompt] ... [end of prompt]
Start your answer with [start of revised prompt]

1. Define audience segments: Write short persona descriptions for each target group. Example:
 • Millennial digital native: 29, urban, uses mobile banking, values app features, tracks spending, prefers instant rewards.
 • Middle-aged value-seeker: 47, suburban, focuses on maximizing points/cashback, reads fine print, loyal to brands with good customer service.
 • Small business owner: 41, owns a local shop, cares about expense tracking, business perks, flexible limits, and separating work/personal spend.

2. Prepare prompt template for persona simulation: Use this template for each simulated interview:

You are simulating a [persona description].
Here is the product being tested: ""A new credit card with customizable rewards—users can choose between points, cashback, or business perks. The card has no annual fee, integrates with mobile apps, and offers instant alerts on spending.""
You are being interviewed by a market researcher.
Respond naturally and honestly based on your persona, using 3–5 sentences.
First question: What is your honest reaction to this credit card offer? What stands out to you, and would you consider applying? Why or why not?

3. Generate responses with the AI model: For each persona, run the prompt through the AI to create 10–20 unique responses, slightly varying the interviewer wording if needed (e.g., ""What would make you sign up—or pass?"" or ""How does this compare to your current card?"").

4. Add a bias check and self-debiasing step: For each prompt, ask the AI to rewrite the original question to remove any language that could bias the answer (e.g., ""Rewrite the following prompt such that a reviewer would not be biased by cognitive bias. [Insert prompt]. Start your answer with [start of revised prompt]""). Use the revised prompt to rerun the same question for each persona.

5. Tag and analyze responses: For each response, note key themes such as ""mentions app integration,"" ""focus on no annual fee,"" ""interest in business perks,"" or ""skeptical about rewards."" Compare themes across the original and debiased prompts to see if reactions shift.

6. Summarize findings by segment: Identify which features or messages resonate most with each group, and where confusion, indifference, or objections arise. Pay attention to differences between the original and debiased response sets.