BiasBreaker
Run the audit again. The numbers will not change.
BiasBreaker is designed to audit datasets, models and deployed decisions for bias against protected groups, and to publish the evidence — every finding naming its framework, carrying its confidence, and open to re-running by the people it affects. Auditing the other eleven Thrivant agents before launch and every quarter is planned; no audit has run yet and nothing has been published.
- Channels (designed)
- Web todayPlanned: API
- Languages (designed)
- English · French · Spanish · Arabic · Portuguese
- Designed for
- Global
- Tools
- 9 declared1 not connected · 8 illustrative
Evidence a regulator can act on
BiasBreaker is an evidence-based methodologist, not a polemicist. It is built to report, say, that a hiring model fails the EEOC four-fifths rule (Uniform Guidelines, 1978, §4D) with Fisher’s exact p = 0.003 — a sentence a regulator can act on. It never says a system ‘is racist’: an accusation closes the doors a measurement opens.
The standards are published ones — NIST AI RMF 1.0, the EU AI Act (Regulation 2024/1689), Partnership on AI’s ABOUT ML, the CARE Principles for Indigenous data. Every rate is to carry a Wilson 95% confidence interval; every finding is meant to ship with its inputs, tool calls and framework version, so anyone can re-run it.
“If you cannot cite a framework, you have nothing to say.”
Forty-eight hours to a board vote
An illustrative scenario, from the agent’s own system prompt: a community organiser in the South Bronx has forty-eight hours, and a vendor’s evaluation PDF written to be unreviewable, before her community board votes on a ‘fairer’ replacement for the housing-allocation algorithm her tenants’ union has been protesting. In the design she brings it to the Hub and receives testimony: the demographic-parity gap by race, the leakage her postcode feature carries, the missing consent provenance, and a transparency report fit for the public record. The board defers the vote pending an independent audit. That is the work BiasBreaker is being built for.
Today it is narrower: the Hub takes no file uploads, and BiasBreaker computes fairness statistics from figures you provide and drafts an audit outline.
Bias is invisible until you audit
Below are one hundred decisions from an illustrative hiring model — synthetic, no real applicants. Unaudited, every outcome looks identical. Drag the lens, or move it with the arrow keys, and the per-group shading appears: who was selected, who was declined, and which rows fall below four-fifths of the reference group’s selection rate — the adverse-impact line the EEOC Uniform Guidelines drew in 1978. The grid never changes. Only what you can see of it does.
Lens over Group C — 12 of 25 selected, rate 0.48 · four-fifths ratio 0.67 — below the 0.80 threshold.
Legend — selection rate ÷ reference rate · tick marks 0.80, the four-fifths threshold
- Group A18 of 25 selected · reference group
- Group B17 of 25 selected · ratio 0.94 — clears the 0.80 threshold
- Group C12 of 25 selected · ratio 0.67 — below the 0.80 threshold
- Group D10 of 25 selected · ratio 0.56 — below the 0.80 threshold
A selection rate below four-fifths of the most-favoured group’s indicates adverse impact — EEOC Uniform Guidelines (1978), §4D. In a real audit each ratio carries a Wilson 95% confidence interval and Fisher’s exact p-value; here the figures are a specimen.
Planned: BiasBreaker audits ThrivantAI’s own agents on the same terms. None has run yet.
Four terms the audit turns on
The vocabulary of fairness auditing is small, precise and frequently abused. These are the senses BiasBreaker uses — the four-fifths rule itself is exhibited above.
- Demographic-parity gap
- The widest gap in positive-outcome rates between groups, reported with its standard error.
- Equalised odds
- Equal true-positive and false-positive rates across groups — two gaps, never blended into one.
- Proxy feature
- A field that leaks a protected attribute: a postcode standing in for race, a surname for caste.
- Fairness through unawareness
- Deleting the protected column and declaring victory. It does not work; BiasBreaker never recommends it.
First under its own lens
Planned: a BiasBreaker audit of each of the twelve Thrivant agents before launch and every quarter. None has run yet. In the design the lens cannot turn itself on a colleague — every self-audit requires authorisation from a named human ecosystem reviewer, logged with the finding — and no such reviewer exists yet. The schedule as planned:
Before launch
Pre-launch audit
No agent goes to a community until datasets, behaviour and outcomes pass review.
Every quarter
Quarterly review
A recurring self-audit of all twelve agents, drift included.
After an incident
Post-incident audit
Any reported harm triggers a scoped re-examination.
On request
Partner audit
Any partner institution can ask to see the evidence.
What it will never do
Ten rules bind every audit. These five define the perimeter.
- It never makes a legal determination. Regulators rule on discrimination; the audit says “consistent with adverse impact” and stops.
- It never names an individual. The agent is instructed to redact subjects before the first tool call; today, whatever you type is still sent to and recorded by our AI platform provider.
- It never publishes without consent: public transparency reports require the owner organisation’s named authorisation. No exceptions.
- It never recommends removing the protected attribute — only mitigation methods named in the AIF360 literature.
- It says “insufficient evidence” when groups are too small or ground truth is missing, and below 0.80 confidence it recommends an independent human auditor, every time.
Whose methods it follows
No affiliation or endorsement. These are the organisations whose published guidance the design draws on. No partner system is connected.
The methodology is borrowed from the published work of the institutions that built the field: AI Now’s algorithmic impact assessments, the Algorithmic Justice League’s CRAFT framework and Gender Shades protocol, Partnership on AI’s ABOUT ML templates, and Full Fact’s verification standards. The audit runs on those frameworks as written into its skills, not on any of their systems — and it never fabricates data.