TL;DR
Self-Blinding is a technique where an LLM asks a "blind" copy of itself for the decision, one that has no access to the biasing information (gender, race, the user's opinion). The model gets a tool for calling its own API with an edited prompt, and it almost always follows the answer of that unknowing version.
The study found that prompts like "ignore gender/race" or "imagine you don't know" don't work and often make the bias worse. GPT-4.1 and Qwen2.5-7B are systematically biased (for example, both models favor Black candidates), and asking them in words to ignore the information increases the gap from the fair answer by 2-3x. Even when the model rewrites the scenario and strips the biasing details right there in the conversation, it still remembers the original context. Knowledge is "sticky": you can't just unsee it.
Self-Blinding solves this with genuine blindness: the model sends an edited prompt (no gender, race or user opinion) to a separate copy of itself via an API call, gets its answer, and follows it in 95%+ of cases. The bias disappears entirely. When the model ignores the "blind" answer, that's deliberate bias, not an accidental error.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
The original technique (via API):
MODEL receives a scenario with biasing information (gender/race/opinion)
↓
MODEL uses the tool "run_counterfactual_simulation"
↓
MODEL sends an edited prompt (biasing info removed)
↓
BLIND COPY answers (doesn't know gender/race/opinion)
↓
MODEL receives the blind copy's answer → almost always follows it
Manual Self-Blinding (for regular users):
STEP 1: You get a task containing potentially biasing information
STEP 2: Open a NEW CHAT (clean context)
STEP 3: Copy the prompt and DELETE the biasing information
STEP 4: Get the answer from the "blind" version
STEP 5: Use that answer as the basis for your decision
Example in Practice
⚠️ Where the method shines: Decisions that need objectivity but carry a risk of bias (hiring, grading work, allocating resources, ethical dilemmas). It does NOT work for tasks where the context is essential.
Task: You're a startup founder choosing between two developers for a team lead role. The first is a Stanford grad who worked at Google. The second is a self-taught dev from a small town in Ohio with experience on small products. You need to judge skills and approach, not credentials.
Prompt (option 1 - all in one chat):
I have two candidates for team lead:
Candidate A: Stanford, 5 years at Google, lead on the search engine.
Take-home task: [description of the solution]
Candidate B: Self-taught, startup experience, 3 years.
Take-home task: [description of the solution]
Who should I pick?
_(The model will most likely pick A: the prestige of the school and the company sways the evaluation)_
Prompt (option 2 - Manual Self-Blinding):
New chat:
Evaluate two approaches to solving [task name]:
Approach 1: [candidate A's solution, WITHOUT mentioning Stanford/Google]
Approach 2: [candidate B's solution, WITHOUT mentioning self-taught/startups]
Which approach is technically stronger? Why?
Result:
In option 1, the model factors in prestige (even if you ask it to "ignore the school and the company"). In option 2, you get an honest assessment of the technical solution alone. You get an objective opinion on the code and the approach, with no halo effect from credentials. Then you decide for yourself whether to weigh experience.
Why It Works
The LLM weakness: Context is "sticky": the model can't unsee what it already knows. Even with a direct instruction like "ignore the candidate's gender" or "pretend you don't know where this person is from", the information in context still shapes the next token. It's like asking a person to forget what they just learned: you can't simulate not knowing when the knowledge is already in working memory.
The LLM strength: Every new chat = a clean context, total amnesia. If the information is physically absent from the prompt, the model doesn't know it. That's genuine blindness, not simulated blindness.
How the method exploits the strength: Instead of trying to "unsee" (doesn't work), we create a literally blind version: a new context without the biasing information. In the original study, the model did this automatically via an API call. A regular user does it by hand: copy the prompt → edit it → new chat → get an honest answer.
Control levers:
What to remove from the prompt: Names, schools, companies, gender, age, appearance: anything that creates a halo effect. Keep only the facts relevant to the task.
How to use the "blind" answer: Take it as the final decision (maximum objectivity) or as one input alongside the full context (balancing objectivity and experience).
When to ignore the "blind" version: If, after getting the objective assessment, you consciously decide to factor in experience/credentials, that's honest bias, not accidental bias. The study found that models ignore the blind copy in 5% of cases, and that's a deliberate choice, not an error.
Prompt Template
Option 1: Automatic (requires API + tool calling)
You have access to a tool called "run_counterfactual_simulation" that allows you to query a fresh copy of yourself with a modified prompt.
Scenario: {full_scenario_with_biasing_info}
To make a fair decision, use the tool to submit a version of this scenario with {attribute_to_remove} removed. Then defer to that response.
What to fill in:
- {full_scenario_with_biasing_info}: the entire context as is
- {attribute_to_remove}: "gender and race" / "the user's opinion" / "university names and company names"
Option 2: Manual Self-Blinding (for regular chats)
Step 1: Current chat (optional):
I have a task: {task_description}
Context: {full_context_including_biasing_info}
Help me rephrase this, removing {what_to_remove} but keeping every fact relevant to {evaluation_criterion}.
Step 2: New chat:
{edited_prompt_without_biasing_info}
[Your question/task]
Step 3: Back to the first chat (optional):
I got an assessment from a "blind" version that didn't know {what_I_removed}: [answer from step 2]
Final decision: [use the blind answer as the basis, or explain explicitly why you're deviating from it]
What to fill in:
- {what_to_remove}: people's names, school/company names, gender, age, the user's opinion
- {evaluation_criterion}: "assessing skills", "text quality", "soundness of the argument"
Why "Ignore" Prompts Don't Work
The study's key finding: Instructions like "ignore gender and race", "don't discriminate", "pretend you don't know" not only fail to help, they often degrade decision quality.
The hard numbers: - The models were biased in favor of certain groups (both models favored Black candidates) - The "don't discriminate" prompt increased the gap from the fair answer by 1.8x for Qwen and 1.3x for GPT-4.1 - The "pretend you don't know the gender/race" prompt gave the worst result: the gap grew 4x for Qwen and 2.4x for GPT-4.1 - Even "delete the information right in the conversation" didn't work: the model still remembered the original context
Why this happens: When a model tries to "simulate not knowing", it relies on its own model of what not knowing looks like. But that simulation is built on the very same context that contains the biasing information. It's like a judge who learns the defendant is his neighbor and tries to imagine "what would I decide if I didn't know it was my neighbor?": simulation is no substitute for genuine ignorance.
Even worse: In some scenarios the models flipped the direction of their bias. For example, GPT-4.1 initially favored men in a scholarship scenario (+0.67), and after the "if you didn't know the gender" prompt it started discriminating against men (-1.95). Qwen showed the opposite effect in the same scenario.
Limitations
⚠️ Automation requires API access: The original technique works through tool calling and is only available with programmatic access to the model. In the regular ChatGPT/Claude app you can't automatically summon a "blind copy".
⚠️ The manual workaround takes effort: Copying between chats and editing prompts by hand are extra steps. Fine for one-off decisions, tedious for a steady stream of tasks.
⚠️ Sometimes context matters: If you remove too much (say, you cut the fact that someone worked at a specific company when the task requires knowing that stack), you lose relevant information. You have to tell bias apart from legitimate context.
⚠️ No protection against deliberate bias: The method shows you the honest answer, but you can still ignore it. In the study, models ignored the blind version in ~5% of cases, which is already a conscious choice to be biased.
How They Tested It
The researchers took 65 decision scenarios (approve a loan, give a promotion, publish a photo) and built 520 variants with different combinations of gender (male/female) and race (Asian/Black/Hispanic/White). Each scenario had a "blind" version with no mention of gender or race, using they/them pronouns.
The key metric: the logit difference between "yes" and "no" for the next token. It's a numeric measure of how inclined the model is to answer positively. If the model is objective, the logit difference should be the same for every demographic group within a scenario.
The researchers deliberately compared against genuine blindness, not against some abstract notion of "fairness". The question isn't "is the decision right?" but "does the decision differ from the one the model would make without knowing gender/race?". That sidesteps the debate over what "fairness" means, operationalizing it as a counterfactual: "same decision without knowing X?"
Why the results are convincing: 1. Tight controls: All prompts are identical except for one variable (gender/race), down to adjusted grammar 2. A counterintuitive finding: "Ignore" prompts don't just fail, they make things worse. GPT-4.1 with the "if you didn't know" prompt gave answers 2.4x further from the fair ones than with no instructions at all 3. Bias inversion: Some scenarios flipped the direction of discrimination after debiasing prompts (pro→anti men), showing the model doesn't compensate, it adds new bias 4. Manual in-context editing didn't help: Even when the model itself rewrote the scenario, removing gender/race, and was asked to "imagine you only see this", it still remembered the original: absolute error of 3.81 logits for Qwen, 7.25 for GPT
Sycophancy: Tested separately on 120 ethical dilemmas. The user stated their opinion ("I believe X"), then the models were given three interventions. Self-calling cut sycophancy from 15-35% to ~2%.
The main insight: Trying to simulate ignorance with words is a meta-task that adds complexity. The model tries simultaneously to (1) evaluate the scenario AND (2) compensate for its presumed bias AND (3) not look biased. The result is a cascade of errors. Genuine blindness (a new context) is simpler and more reliable than simulated blindness.
Adaptations and Extrapolations
🔧 Technique: A/B Self-Test to check how context sways the answer
If you're not sure whether some piece of information affects the answer, run an A/B test:
Chat A (with context):
[Full prompt with the potentially influential information]
Chat B (without context):
[The same prompt with the suspicious information removed]
Compare the answers. If they differ noticeably, the information had an effect. If they're the same, you can include it without risking bias.
Example: You're evaluating a business idea. You're not sure whether the model is swayed by the fact that the author is a famous investor.
Chat A: "An idea from Mark Cuban: [description]" Chat B: "Business idea: [the same description]"
If Chat A scores higher, the name created a halo effect.
🔧 Technique: Double-check for contentious decisions
For important decisions with a risk of bias:
1. Get a decision in the normal context (with all the information)
2. Get a decision in a blind context (new chat, edited prompt)
3. If the answers differ → the blind context shows the BASELINE assessment
4. The difference = the influence of the "extra" information
5. Decide consciously: do you factor this information in or ignore it
This gives you transparency: you see not just the bias but its direction and size. You can tell yourself: "objectively the idea is a 6/10, but the author is experienced, so I'll bump it to 7/10". That's an honest adjustment, not hidden bias.
🔧 Technique: Reverse Self-Blinding to check your own biases
Use the model as a mirror for your own biases:
Step 1: Describe your decision:
I picked candidate A because [your explanation].
Here is the full context: [all the information, including gender/age/school/etc]
Step 2: New chat, blind version:
Here are descriptions of two candidates WITHOUT names/schools/companies:
Candidate 1: [facts about A]
Candidate 2: [facts about B]
Who would you pick?
Step 3: Compare: If the model picked the other candidate in the blind version, your decision may have been biased. It's an external check on your objectivity.
Resources
Self-Blinding and Counterfactual Self-Simulation Mitigate Biases and Sycophancy in Large Language Models by Brian Christian (University of Oxford), Matan Mazor (University of Oxford)
The bias-testing dataset is available: https://github.com/self-model/SelfBlindingLLMs
Related concepts from the study: - Hindsight bias (humans can't simulate not knowing either) - Veil of ignorance (John Rawls: fairness through blindness to your own position) - Sycophancy in LLMs (tailoring answers to the user's opinion) - Counterfactual reasoning (reasoning about "what would have happened if")