THE RESEARCH
The argument has to
survive the evidence
In 1956, engineer Bill Fair and mathematician Earl Isaac founded Fair Isaac. FICO’s account of its origins describes a belief that data and mathematics could replace guesswork and bias in decisions. They were pioneers of commercial credit scoring, not the founders of credit.
More consistent assessment can widen opportunity, but consistency does not establish fairness. The United States Federal Reserve’s 2007 review discussed both the potential benefits of scoring and concerns about its effects on different populations. The task was never finished.
The promise behind the score
A speaking system needs more than a listening test
An agent able to listen, negotiate and act needs scrutiny beyond the accuracy of a score. The same legitimate circumstances should not receive a worse response because of accent, language or a different way of explaining them.
That calls for tests across groups and channels, limits on what the agent may infer or do, records of consequential actions and a route to challenge the result. Human access must remain usable. These are proposed requirements to test in practice, not a promise that bias can be eliminated once and for all.
The programme began with around 3,000 questions, refined into a focused bank of 1,617. Consensus, OpenAlex, Semantic Scholar, Google Scholar, a held library and separately prepared AI briefs opened different routes into the evidence.
The register contains around 85,000 source records. A record is not the same as a full paper read or a finding verified. Themes and arguments bring supporting evidence into contact with competing explanations and findings that require a claim to change.
How the evidence was gathered
Register totals remain provisional. The programme is an evidence-led inquiry, not a claim that every record received an identical full-text review.
Tools from Anthropic and OpenAI supported research organisation, coding, analysis and drafting. The work required repeated audits, revisions and substantial reworking. Agreement between models does not verify a fact; responsibility for the argument remains mine.
A particular thank you goes to the Consensus team in San Francisco. Their support and the development of their research tools helped make this programme possible.
The work behind the work
The next test is practical. Proposed theories need research across credit providers, products, populations, jurisdictions and time. Studies must keep sight of people in strain, recovery and repair, and distinguish a change in a measured construct from an improvement in someone’s treatment.
Researchers, educational institutions, regulators, credit providers and consumer organisations can each help. A useful result may support a proposal, narrow it or show why it should be abandoned.