Algorithmic Bias Detection Tool
Readiness
TRL 7 · SRL 5.5 · gap +1.5
Gap = technical readiness − societal readiness. Positive: the technology is ready and society is not. Negative: society wants it more than the technology can deliver.
Not scored. Watchlist entries do not enter any gradient.
Bias detection tools scan a machine-learning pipeline — the training data, the model, the deployed outputs — and measure whether a protected characteristic is influencing results it should not.
The technology is further along than the debate suggests. Fairness metrics exist, adversarial de-biasing works, and the scanning is not hard. What is hard is that fairness is not one quantity. The main metrics are mathematically incompatible with each other except in trivial cases: you can equalise error rates across groups, or equalise the meaning of a positive prediction, but not both. A tool cannot resolve that, because it is not a measurement problem. It is a question about what the institution owes the people it is deciding about, and somebody has to answer it before the tool has anything to check against.
Which is why deployment lags capability here. The instrument is ready and the specification it would test against does not exist.
Common misconception: that bias scanning is a debugging step. Choosing the fairness metric is a policy decision, and running a scan without having made it produces a number that means whatever the reader wants it to.
What it does
mechanism — not yet written. Published as missing rather than hidden.
Candidate megatrends — unreviewed
- Automation of Cognitive Work
- Digital Sovereignty
- Synthetic Media and the Verification Collapse
- The Application Gap