BiasBreaker
Test the algorithm. Show the numbers.
Fairness statistics and audit outlines for AI systems, from the figures you provide.
What BiasBreaker does
Live in the Hub chat. Answers say when a figure is illustrative, and nothing is sent, booked or decided for you.
Computes fairness statistics from figures you provide and drafts an audit outline. It has not audited any ThrivantAI agent yet.
- Computes fairness statistics from the figures you provide: selection-rate and error-rate gaps between groups, the four-fifths ratio, Fisher’s exact test and Wilson confidence intervals
- Flags input features that commonly stand in for protected attributes, such as postcode, surname or school attended, from a built-in list
- Drafts an outline for a dataset, model or outcome audit and for a transparency report, for a qualified reviewer to complete
- It has not audited any ThrivantAI agent. A quarterly audit is planned; none has run yet, and the self-audit step is a fixed checklist
- Nothing is published, and no reviewer is contacted
What is real, what is illustrative
9 tools are declared for BiasBreaker. Each is one of four kinds, and the Hub enforces the kind when the tool runs.
Fixed rules on what you type (5)
Applies fixed rules or arithmetic to what you type. Nothing is looked up.
- Audit outcome
- Audit thrivant agent
- Compare disparate impact
- Compute fairness metrics
- Detect protected attribute leakage
Illustrative only (3)
Returns demo output, marked as illustrative, with no confidence score or source attached. Not a real result.
- Audit dataset
- Audit model
- Generate transparency report
Not connected (1)
Would act in the world (contact, send, file, store, publish), so it is switched off. It never runs, and nothing is sent or stored.
- Escalate to human
Start with a question
Example questions. Pick one and the chat opens with it ready to send, or bring your own.
- “Our district uses an algorithm to shortlist welfare applicants — how do we test whether it is biased?”Ask BiasBreaker
- “Our hospital triage model scores men as higher risk than women with the same symptoms — what figures do you need from us to check it for bias?”Ask BiasBreaker
- “We are about to launch a hiring model — what should we test before going live, and what report should we publish?”Ask BiasBreaker
Where it runs
The Hub runs on the web. Other channels, and the languages and regions listed, record what the agent is designed for.
- Channels
- Web (live)
API (planned, not yet available) - Languages (designed)
- EN, FR, ES, AR, PT
- Designed for
- Global
- Organisations the design looks to
- AI Now Institute, Algorithmic Justice League, Partnership on AI, Full FactNo affiliation or endorsement. These organisations publish guidance the design draws on, or are ones we hope to work with. No partner system is connected.
More than one goal at once?
The coordinator can pass one conversation to several agents and bring their answers back as one.
Open the coordinator