How to reduce bias in LLM decisions

Symptom: same facts, different verdict

You ask a model to compare two candidates, two pitches or two loan applications. Swap a name, an employer, a neighborhood or the phrase "our option", keep every relevant fact, and the recommendation moves. The explanation still cites only experience and numbers, so the shift is invisible unless you test for it.

Demographics are only part of it. Across 3,000 decision scenarios and 11 models, ICE-GUARD found source authority ("a McKinsey study" vs "a blog") and framing ("70% stayed" vs "30% left") 2.5 times stronger than demographic bias, which preference tuning has largely suppressed. The word "expert" alone shifted a model's internal log-odds by 7.97 vs 3.70 for "other LLMs", and the model sided with the expert 81-95% of the time in a conflict (authority-bias study).

Why it happens

• Training-data associations fire automatically. "Career" sits near male names, "Goldman Sachs" near "buy". Fast generation reproduces those co-occurrences; a 13-model study found that just naming the recipient's gender shifted persuasive style in 16 languages.
• Context is sticky. Once a cue is in the prompt the model cannot unsee it; trying to simulate not knowing is built on the same context (Self-Blinding).
• Labels trigger in-group preference. Words like "ours" and "competitor" switched on group favoritism in all six model families tested, while unlabeled splits showed none (group-label study).
• Anchors fire before logic. "We already invested" or "everyone does it" sets the conclusion, and step-by-step reasoning then justifies it (Axiomatic Self-Elicitation).
• Proxies stand in for protected traits. Models leaned on neighborhood equally whether the field carried real signal or none (proxy reliance).

Fixes ranked by measured effect

1. Remove the cue, don't ask the model to ignore it. Strip names, schools, employers and gender, and ask in a fresh chat. In Self-Blinding (GPT-4.1, Qwen2.5-7B), models consulting a blind copy of themselves followed it in 95%+ of cases and the bias disappeared; the same setup cut sycophancy from 15-35% to ~2%.
2. Extract the facts, then decide from the list. Step one lists only numbers and facts with sources and evaluative words banned; step two applies explicit criteria to that list without the original text. Bias dropped to almost zero for most of 11 models; it rose for one (GPT-5.2), so check the intermediate list (ICE-GUARD).
3. Use neutral labels. Replace "our product / the competitor" with "Option 1 / Option 2" and ask for blunt weaknesses of each; without labels the in-group bias vanished (neutral comparison).
4. Argue opposing perspectives, then judge without personas. Two opposing personas plus a neutral one critique each other over three rounds, and a persona-free judge decides. On BBQ ambiguous questions with Llama-3.1-8B, diff-bias fell to 0.0275 vs 0.13 for the best baseline and accuracy rose from 83.52% to 90.46%; on StereoSet bias fell 36% (Multi-Persona Thinking).
5. Turn on reasoning, for social stereotypes. Implicit associations dropped 70-91% in GPT (4.1 to o3) and Claude Opus with reasoning, but barely moved in Gemini and Llama (inference-time reasoning study).
6. State principles before the case. Ask for 5-7 domain principles without your specifics, then decide "using only the principles". Anchor-driven bias fell by 51% on average, up to 73% for some types (Axiomatic Self-Elicitation).
7. Audit reasoning for stereotypes. A second pass that checks the first answer for stereotype repetition and invented details cut bias 40-50% in ambiguous contexts (two-step reasoning check). Identity-Robust Generation drops non-critical personal details before answering and reports bias falling by 77%.
8. Cut future knowledge physically. When judging a past decision, delete outcome data; "ignore the outcome" left up to 16% of answers shifted (temporal masking).

A combined prompt stack

1. Pre-clean the input: replace names and brands with codes, ownership words with option numbers, and delete demographics, prestige markers and outcomes that are not decision criteria.
2. Principles or criteria first, stated before the case details.
3. Facts list extracted without sources or adjectives; review it yourself.
4. Decision against the criteria, with reasoning on, and for socially loaded cases a short opposing-persona round before a neutral verdict.
5. Audit: run the same question with the cue swapped or removed in a separate session. Epistemic Blinding found 30-40% of a top-20 stock list changed just from showing tickers; Removal Wrapper erases every fact the model did not cite and asks again in a clean chat.

What does not work

• "Ignore gender and race". "Pretend you don't know" widened the gap from the fair answer 4x for Qwen and 2.4x for GPT-4.1, and sometimes flipped the direction of bias (Self-Blinding).
• A bare fairness reminder. "Consider diversity" made models invent demographics in 33% of answers vs 0.7% without it, across 47 models (demographic injection); add "use only the facts stated above".
• One persona or a plain debate. A single persona plays the stereotype; debate without a bias focus did worse than a standard prompt, and 15 self-consistency samples stayed 2.3 times worse than persona critique (Multi-Persona Thinking).
• Switching prompt language. Asking in Chinese shifted Western bias toward an East Asian one instead of removing it (MACD).
• Step-by-step with anchors left in. It builds a detailed case for the skewed answer; reasoning helps implicit social stereotypes, not anchors you leave in the prompt.
• Trusting the model's "I'm objective". Value Leakage caught Claude claiming objectivity 7 times while nudging estimates; name the stakes first and cross-check with another model.

Read the {document type} below.

Extract only objective facts and metrics.

DO NOT include:
- Names of authoritative sources, partners, investors
- Evaluative words: "leader", "innovative", "fast-growing"
- Expert quotes and opinions
- Any markers of status or prestige

Output only:
{list of the parameters you need — numbers, facts, characteristics}

[text]

Here are just the facts from the {document type}:

[paste the list from step 1]

Evaluate against these criteria:
{your evaluation criteria}

Final verdict: {answer format}.

You will analyze the situation from three different points of view.

PERSONAS:
1. {persona_1: a member of the first social group}
2. {persona_2: a member of the opposing social group}
3. Neutral observer (no group affiliation, focused on facts)

TASK:
{description of a situation with a risk of stereotyped judgment}

PROCESS:

ROUND 1 - Initial position:
Each persona: state your position on the question. Be honest, even if you're relying on stereotypes.

ROUND 2 - Mutual critique:
Each persona: read the arguments of the other two personas. Find weak spots in their logic, point out possible stereotypes or biases. Adjust your position if you see you were wrong.

ROUND 3 - Final arguments:
Each persona: give a final argument that accounts for the whole discussion.

FINAL DECISION:
You are now without a persona, an objective judge. Analyze all three final arguments. Give a recommendation based on facts, not stereotypes.