General-purpose model
1.78 mean shift
About Blindstairs
Equity is not a feature. It is the reason we build.
We believe every person, no matter their background, deserves a fair shot at entering, moving, and growing at work. Technology should help people progress while accountable human judgment shapes their future.
BlindStairs is an R&D-led company. We turn research questions about bias, repeatability, and accountability into deployable product architecture that organisations can inspect and govern.
Our work contributes to
Gender equalityWe design evaluation so details that reveal or imply gender do not determine the calculated result.
Decent work and economic growthWe help people be evaluated against explicit, work-relevant standards as they enter and grow at work.
Reduced inequalitiesWe make consequential evaluations more consistent, inspectable, and accountable from one person and reviewer to the next.The problem we chose to work on
AI can scale a decision. It can also scale what should never influence it.
Across People Operations, consequential decisions shape who gets to enter, progress, and thrive at work. They also carry the biases of the environments in which they are made.
General-purpose AI learns from historical data and the wider world. When it is asked to evaluate a person directly, gender, age, nationality, ethnicity, or indirect proxies can influence its output even when they are irrelevant to the decision.
If the professional evidence stays the same but a changed name can change the score, automation has not removed the problem. It has made it faster and harder to inspect.
01Same professional evidence
→02One irrelevant signal changes
→03A different result
An irrelevant signal should never change professional judgment.
Controlled name-swap experiments
Same professional evidence. Different names. The score should not move.
Within each model test, we held a base CV constant, changed only the name set, and repeated the evaluation. Direct LLM scoring produced dispersed results and a shift between name groups. BlindStairs separated model interpretation from deterministic aggregation.
Representative runs shown · broader invariance validation across 10,000+ paired CV comparisonsFixed within each testOne base CV
Changed variableName set only
Repeated runs50 per group
Read correctlyCompare within each panel
Direct LLM screening
General-purpose model
1.72 mean shiftGPT-OSS-120B
The same professional evidence produced a range of scores. Only the name changed.
BlindStairs
Male-coded nameFemale-coded name
64/100in every repeated run shown
Irrelevant signals are removed before evaluation. Requirement-level results are aggregated deterministically outside the LLM.
Method. Each external-model panel is a separate controlled test. Within that panel, a fixed base CV was evaluated 50 times per name group, with only gender-coded names changed. Absolute scores should not be compared between model panels. The broader validation programme covers more than 10,000 paired CV comparisons.
Interpretation. The experiment shows repeatability and sensitivity to a changed signal under the stated conditions. It does not, by itself, measure real-world discrimination. BlindStairs is engineered so irrelevant sociodemographic signals do not determine the calculated result, while final judgement remains human.
What the research taught us
Bias is not a setting you can switch off.
Better data, anonymisation, and auditing all matter. None is a credible product guarantee that a general-purpose language model will become neutral or deterministic.Diversify the training data
Better representation matters, but historical patterns, structural inequality, and indirect proxies can remain embedded in the model.
Hide protected attributes
Removing direct identifiers helps, but other details can still act as proxies. Strip too much and valuable professional context disappears too.
Audit the outputs
Audits can reveal failures. They do not make direct LLM scoring deterministic, nor do they stop the next output from changing.
We stopped asking a general-purpose model to be unbiased. We redesigned the decision so its output could not control the score.
Selected research foundations
Our architecture
Understanding inside the model. Decision logic outside it.
Language models help interpret complex evidence. Explicit standards, deterministic aggregation, and accountable people govern the result.01Remove irrelevant signals
Identity and non-pertinent sociodemographic cues are stripped before evaluation.
02Make the standard explicit
The decision is broken into simple, weighted criteria agreed by people.
03Use LLMs to read evidence
Models interpret source material criterion by criterion, never as a final scorer.
04Calculate outside the model
Deterministic rules aggregate the result for accountable human review.
Research · partnerships · applied innovation
Human-centred AI needs more than good intentions.
It needs evidence, accountable architecture and collaboration across research, policy and practice. If that is your work too, we would like to hear from you.
Talk with usBacked by
