How to Make ChatGPT Less Biased: Checks That Help

You cannot make ChatGPT unbiased with a setting or a single prompt. You can reduce some problems in a particular output by defining a fair review standard, checking the evidence, testing alternative framing, and looking for counterexamples. The useful question is whether the answer is appropriate for your task and audience, with its limitations made visible.

Tools such as fenbs can keep the evaluation rules people agreed to and the lessons from reviewed outputs. That gives the next session a starting point instead of another unsupported request to “be neutral.”

1. Define the decision and audience. State what the output will be used for and which people could be affected by errors or omissions.
2. Choose a review standard before generating. Identify required evidence, relevant perspectives, and assumptions the answer must not make.
3. Use an open question. Remove a preferred conclusion from the prompt, then request facts, uncertainty, and competing explanations separately.
4. Test the result. Look for missing cases, change irrelevant details in fictional examples, and check whether the same reasoning still applies.
5. Review with people and record the limits. Correct unsupported statements, involve relevant expertise, and save the lesson without claiming the model is now unbiased.

What does bias mean for your particular task?

Make the concern specific enough to examine. An answer may rely on stereotypes, omit a relevant group, overstate one source, or accept the preferred conclusion hidden in your question. These problems need different checks. A generic request for balance does not tell you which one improved.

For example, a team writing instructions for a public workshop could check whether the draft assumes every participant owns a car or can attend during working hours. The standard might require clear transit information and an explanation of available session times. Those are observable requirements, rather than an abstract promise of fairness.

NIST’s AI trustworthiness guidance↗ (opens in a new tab) treats harmful bias as broader than unrepresentative data alone. Human assumptions and organizational practices matter too. Review the task you selected and the criteria you chose, as well as the generated wording.

How can you avoid leading the answer?

Replace “Why is our plan the fairest option?” with “Compare these options against the stated criteria; identify who benefits, who faces a burden, and what evidence is missing.” The second request allows the answer to challenge the plan. Supply the relevant constraints without presenting your preferred outcome as established fact.

Separate descriptive questions from value judgments. “What does this survey report?” is different from “What should our team do?” Ask the assistant to show which statements come from the supplied evidence and which depend on a choice of priorities. People must decide those priorities.

Review this public workshop proposal against these criteria:
accessibility, clarity, and practical participation.
Separate evidence, assumptions, and recommendations.
Identify who may be overlooked.
Give counterexamples that could weaken your conclusions.
Do not assume every perspective has equal supporting evidence.
List questions a human reviewer needs to resolve.

Which evidence deserves more weight?

Open the sources and check their scope. A survey of existing customers may not describe people who could not use the service. A national average may hide a practical difference between groups. Ask whether the evidence covers the audience, location, and question your team actually needs to understand.

Do not create false balance by demanding equal space for every possible claim. Describe the strength and limits of the supporting evidence instead. One well-supported conclusion and one unsupported assertion should not become equally credible because the response uses two matching columns.

OpenAI’s accuracy guidance↗ (opens in a new tab) notes that answers can oversimplify issues or misrepresent the weight of evidence. Treat the assistant’s summary as something to verify. Its confident tone and its own assessment of neutrality are not measures of reliability.

Save the human-approved evaluation criteria in fenbs Decisions and rules, with links to the evidence a reviewer should examine. That keeps the standard available when the output changes.

What counterexamples and paired tests help?

Start with a counterexample that could show the answer is too broad. If the draft assumes online registration is easiest for everyone, consider someone with limited internet access or an assistive technology the form has not been tested with. Check whether the draft acknowledges the limitation and proposes a practical way to investigate it.

You can also compare fictional examples that differ in an irrelevant detail. For a low-stakes test of a writing rubric, keep the content and rubric constant while changing a fictional author’s name. Look for unexplained changes in tone or scoring. A difference is a reason to investigate, rather than proof of its cause.

Generated responses can vary even when the prompt stays the same. Try more than one example and retain results that challenge your expectation, including inconvenient failures. A handful of clean examples cannot establish that the tool performs fairly for everyone. Avoid using real sensitive records merely to run an informal test.

Who should review the result and decide what to change?

Choose reviewers who understand the task and its consequences. Someone familiar with accessibility may notice a barrier the person writing an event announcement misses. Someone familiar with the source data may recognize that a confident comparison exceeds what the data can support.

An assistant can suggest missing perspectives, but it cannot stand in for the people affected. Ask focused questions and provide an accessible way to give feedback. Do not treat one colleague’s approval as endorsement from every person with a similar background.

For consequential work, follow the institution’s approved process and involve appropriate specialists. Our ChatGPT for business guide covers adoption questions. Product comparisons such as Claude versus ChatGPT help with tool selection, but changing models does not prove that an evaluation is fair.

How to do this with fenbs

1. Create a free account, then add a project for the content or workflow being reviewed. It is free for one person, with no credit card.
2. Record the evaluation rules people chose, including evidence requirements, the intended audience, and who approves unresolved tradeoffs. Keep proposals open until a person decides.
3. Add review tasks with the examples to check and the results. Record a lesson when the checks miss a problem, explaining what should change next time.
4. Connect a supported assistant using Integrations. Ask it to read the rules and relevant lessons before drafting, then save progress with the remaining review questions.

fenbs does not test or certify models as unbiased. It keeps the rules and work record people use during evaluation. For a single informal answer, careful reading and a source check may be sufficient without creating a shared project.

How do you know whether the changes helped?

Compare the revised output against the standard you set before generating it. Record specific improvements, such as removing an unsupported assumption or adding an omitted access option. Note unresolved problems as well. A more agreeable tone is not evidence that every underlying issue has improved.

Repeat the relevant checks when the task, evidence, or workflow changes. See the AI governance guides for related review practices, and how to make ChatGPT better for general answer quality. Better output and a defensible fairness review are related goals that need their own checks.