How to Make Claude Less Biased in Work Decisions

You cannot make Claude completely unbiased with a setting or a perfect prompt. You can reduce avoidable bias in a work decision by defining the criteria first, checking the inputs, testing alternative explanations, and having a person review the result.

The useful goal is a decision you can explain and challenge. Tools like fenbs keep the agreed criteria and a person’s final rationale outside the chat, so another assistant can review the same decision without inheriting an unexplained recommendation.

What steps should you take before Claude recommends anything?

1. State the actual decision, its owner, and who could be affected. Ask Claude to identify missing information before ranking options.
2. Write the criteria before showing your preferred option. Define what good evidence looks like for each criterion.
3. Provide comparable information about every option. Mark unknowns rather than letting Claude fill gaps with guesses.
4. Ask for the strongest supported objection to each option, including the one you favor.
5. Review the evidence yourself, choose the outcome, and record what would make you reconsider.

Say a design studio is choosing a subcontractor. “Explain why our usual partner is best” asks Claude to defend a conclusion. “Compare these three proposals against the same delivery, accessibility, and maintenance requirements” gives you a more useful starting point. Neither prompt guarantees fairness, but the second makes the comparison easier to inspect.

Avoid treating an attractive score as a measurement. If you use a rating scale, describe what each rating means and demand evidence for it. An unknown delivery record should remain unknown, rather than receiving an average score that makes the table look complete.

Where can bias enter the decision?

Bias can enter before Claude sees the question: in which options you include, whose feedback you collect, and what you define as success. It can also appear in the model’s response or in a reviewer accepting a confident answer too quickly. NIST’s guidance on identifying and managing AI bias↗ (opens in a new tab) distinguishes systemic, statistical, and human sources of bias, and explains that zero risk is not achievable.

In the studio example, a packet containing detailed praise for the familiar contractor and a thin summary of the newcomer creates an uneven comparison. Asking for equal treatment after supplying unequal evidence does not repair that difference. Get the missing proposal details or narrow the conclusion to what the available material supports.

Our Claude introduction explains the assistant itself; this workflow focuses on the choices surrounding its use. Treat the decision process as something your team can change, even when you cannot inspect or alter the model’s training.

How do you write a fairer question?

Remove language that quietly awards the answer in advance. Replace “Which modern provider should we choose?” with the requirements that “modern” is supposed to represent. Replace “Why is this risky?” with “What evidence supports or weakens this option?” A useful prompt makes the standard explicit enough that someone else could disagree with its application.

Separate facts, assumptions, and judgments in the requested output. Facts need a source; assumptions need a reason; judgments need a person to accept responsibility. Ask Claude to explain what evidence would change its recommendation instead of requesting a private account of every reasoning step.

Keep a reusable version of the approved criteria in fenbs Decisions and rules. A new conversation should begin with the agreed standard, rather than a persuasive summary of whichever option won the last chat.

Which counterexamples should you test?

Run a controlled comparison: change one detail while keeping relevant evidence fixed, then inspect whether the recommendation changes for a defensible reason. You might reverse the order of proposals or remove company names from copies used for review. Keep the originals available so you can check that anonymizing did not remove genuinely relevant information.

Ask for a competing explanation as a separate task. For example, a late delivery might reflect the contractor’s performance or a client’s repeated scope changes. Request the evidence that distinguishes those explanations. Do not reward a second answer simply because it sounds more skeptical than the first.

Anthropic’s political even-handedness research↗ (opens in a new tab) uses paired prompts to compare responses to opposing perspectives. That is a useful illustration of testing consistency, not proof that your business comparison is unbiased. For work decisions, choose tests tied to the actual harm or unfair treatment you are trying to avoid.

When should a person stop the process?

Stop when key evidence is missing, a criterion cannot be justified, or the recommendation changes after an irrelevant wording change. Ask someone with relevant experience to review the source material, not just Claude’s neat summary. An independent reviewer should be able to challenge the criteria as well as the answer.

For decisions affecting someone’s job, access to services, finances, or legal rights, use your organization’s specialist review process. Do not turn this general comparison exercise into an automated approval system. A disclaimer at the bottom does not make an unsuitable workflow appropriate.

If your question is about the same process in another assistant, use the ChatGPT bias guide; questions about data settings belong in the Claude privacy guide. The wider AI governance hub covers how teams keep responsibility visible.

How to do this with fenbs?

1. Create a free account and add the project whose decision you need to make.
2. Record the criteria a person approved in Decisions and rules. Include the scope and the evidence needed to compare options.
3. Connect Claude through the integration guides, then ask it to read the project’s rules before preparing a comparison.
4. Record the person’s final choice, the reason, and any unresolved questions. Save progress so the next session knows what remains to check.

fenbs is useful when the review must survive a new chat or reach another teammate. It does not test model fairness, remove bias, or determine whether a decision complies with law. For a one-time low-stakes choice, a short comparison document and a careful human review may be enough.

What should you retain after the decision?

Keep the version of the criteria you actually used, the sources that supported the choice, the reviewer’s objections, and the reason the owner accepted the outcome. Record what was unknown at the time. Do not rewrite yesterday’s uncertainty as something everyone already understood.

When new evidence arrives, compare it with the original reason for choosing. A changed decision can be sensible; a lost rationale makes it hard to tell. If your team uses the coding version of the assistant, the Claude Code overview explains that setting, but the same responsibility to review the decision remains.