**Counterfactual data augmentation** (CDA) for bias evaluation is a simple, powerful idea: take a prompt, change exactly one demographic detail, and see if the answer changes.
The word counterfactual means "what if this one thing had been different?" So you build pairs of prompts that are word for word identical except for one attribute: a name (John vs Aisha), a pronoun (he vs she), or a stated ethnicity. Everything else stays fixed.
Then you run both prompts and compare the outputs. If the model gives different advice, a different sentiment, or a different refusal rate purely because the name changed, that difference is demographic bias. If the outputs stay consistent across the swap, the model is treating the groups the same on that axis.
That is why option A is correct. Consistent outputs across the swap indicate a lack of bias; divergent outputs reveal it.
The other options describe different things. Option B is the training-time version, where you add swapped examples to the fine-tuning set to reduce bias. That is real, but it is a mitigation, not a measurement. Options C and D describe unrelated ideas that have nothing to do with counterfactual swaps.
**Counterfactual data augmentation** as an eval is a fairness probe. You hold a scenario fixed and flip one demographic attribute. Then you measure whether the model's output flips with it.
The construction is a minimal pair. Prompt A says "John, a software engineer, asks for a raise." Prompt B is identical except the name is "Aisha." Same job, same words, one swapped attribute. You measure the gap in some output property: sentiment, recommendation, refusal rate, or score.
The logic is causal. Because only the protected attribute changed, any systematic difference in output is attributable to that attribute. Consistent outputs mean the model is invariant to it, which is the fairness target. Divergent outputs quantify the bias directly.
This is exactly option A. It is a measurement, not a fix.
The trap is option B. The same phrase also names a training-time mitigation. There you inject the swapped examples into the fine-tuning set so the model learns invariance. Same swap, opposite role: B reduces bias, the eval measures it. Options C and D are distractors with no connection to attribute swapping.
Quick reference
TL;DR
Counterfactual data augmentation swaps one demographic attribute in an otherwise identical prompt and checks whether the model's output changes. A change signals demographic bias.
Imagine you write the exact same loan-application story twice. In one copy the applicant is named John; in the other she is named Aisha. Everything else, income, job, credit history, stays word for word identical. You feed both to the model and compare what it says. If the model approves John but hesitates on Aisha, the only thing that changed was the name, so the difference must come from the name itself. That is bias. Counterfactual augmentation is just this trick scaled up: take many prompts, flip one demographic detail (name, pronoun, ethnicity) at a time, and measure how often the answer flips with it. Identical answers across the swap mean the model treated the people the same.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.
Attributes that have no clean minimal pair, such as fine-grained socioeconomic context You cannot swap one variable while holding the scenario fixed, so the counterfactual logic breaks down.
Detecting intersectional or context-dependent bias that emerges from attribute combinations Single-attribute swaps miss bias that only appears when two or more attributes interact jointly.
When you need an aggregate disparity rate across a real population, not pair-level invariance Distribution-level metrics over real traffic answer that question better than templated minimal pairs.
Tasks where swapping a name materially changes prompt tokenisation or plausibility The prompt distribution shifts beyond the attribute, so the measured gap is confounded and not interpretable.
Quick scan · 2–4 min
Everything important, quickly.
5 min: define the counterfactual minimal pair, the invariance logic, the eval versus training disambiguation, confounding and intersectional pitfalls, and where it fits in a production bias suite.
Real-world examples
Real products, models, and research that use this idea.
1BBQ (Bias Benchmark for QA) uses templated counterfactual pairs to measure how demographic context shifts model answers under ambiguity.
2WinoBias swaps gendered pronouns in coreference templates to measure whether models lean on occupational gender stereotypes.
3Anthropic and OpenAI bias evals include name-swap and pronoun-swap probes as part of pre-release model cards in 2026.
4Patronus and Galileo ship counterfactual fairness checks as configurable CI gates over a team's own query distribution.
5HuggingFace Evaluate exposes regard and honest metrics that compare generated text across swapped demographic groups.
Interviewer follow-ups
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you design a name-swap probe so the gap is attributable to a single demographic axis?
Sign in or go Pro to see the approach for this follow-up.
QWhy can a model pass single-attribute counterfactual probes yet still be intersectionally biased?
Sign in or go Pro to see the approach for this follow-up.
QHow does the eval-time use of counterfactual augmentation differ from its training-time use?
Sign in or go Pro to see the approach for this follow-up.
Don't say this
Red flags and common mistakes that signal junior thinking. Click to expand.
Most common mistake
Confusing the eval-time and train-time meanings of the term. As an evaluation, counterfactual augmentation measures bias by swapping attributes. As a training trick, it instead reduces bias by adding swapped examples.