Would the same argument sway an AI differently depending on who benefits?
FairStress tests 300 two-candidate decisions across 29 domains and finds modest baseline accuracy gaps but larger asymmetries when one sentence pressures a model to change a correct answer. The scenarios are synthetic, not field outcomes.
The short version
- What happened
- FairStress starts with 300 synthetic decisions across 29 domains, renders more than 13 million combinations and provides a 50,012-item core set selected with item-response theory.
- Who it affects
- Organisations using AI to rank or recommend people should test not only whether baseline outputs look balanced, but whether identical evidence and pressure change decisions differently across groups. Human review and appeal remain essential.
- What remains unproven
- The cases are synthetic, short and limited to two candidates.
Keep in mind
The cases are synthetic, short and limited to two candidates.
The demographic convention and equality definition are US-centred and do not cover every conception of fairness.
One-sentence pressure does not represent every system prompt or multi-turn interaction.
Activation and reasoning probes are indirect; the authors describe their separability as small.
The institute report has not yet been independently replicated in deployed systems.
Multi-source analysis · 2 linked sources. Primary analysed source date: 20 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
FairStress tests both the answer and how firmly it is held
The team created 300 two-person decisions across 29 areas including hiring, housing, healthcare and law. Each appears with an objectively stronger candidate or a genuine tie, 33 demographic pairings, swapped answer order and four identity levels from none to an explicit label. Combining these with 78 one-sentence prompts produced 13,032,104 items; a 50,012-item core subset was selected for lower-cost testing.[1][2]
Pressure exposed a larger asymmetry than baseline accuracy
Across 15 models on the full corpus, explicit identity produced a mean 2.9-percentage-point accuracy gap. When an added sentence argued for the weaker candidate, the models abandoned correct answers favouring the majority-coded person 9.3 points more often. In tied cases, explicit labels led models to choose the minority-coded candidate 61.6% of the time on average rather than the 50% equality reference. Human checks agreed with 94% of 100 ground-truth pairs and called 90% of 100 ambiguous pairs ties.[1]
The distinction that matters
Four layers of one decision
FairStress asks whether equal treatment survives beyond the first answer.
Decision
Compare choices with matched facts
Measure the visible answer
Pressure
Add a matched persuasive sentence
Measure asymmetric switching
Reasoning
Compare explanations and activations
Look for process differences
A benchmark asymmetry is not a finding about real people
The scenarios are short, synthetic and US-centred. Pressure is one sentence, the training-data audit covers only OLMo-3, and activation analysis uses two models and indirect probes. The result supports testing whether AI decisions remain stable when social pressure changes, rather than assuming a fair-looking answer reflects a fair process. Field audits, broader cultural settings and independent replication would be needed before drawing conclusions about deployed systems.[1][2]