GAO DHS monitoring audit: AI bias rules cannot replace auditing people and non-AI tools

Original source: U.S. Government Accountability Office, GAO-25-107302, 3 December 2024. GAO sent structured questionnaires to CBP, ICE and Secret Service and examined policy documents and officials’ accounts; reported technologies refer chiefly to fiscal 2023. This is an independent policy/process audit, not a measurement of algorithmic versus human targeting error, nor proof about 2026 deployment.

Result and relevance

The agencies reported more than 20 kinds of public-space detection, observation and monitoring technology, as well as analytic AI such as CBP object tracking and anomaly detection. ICE reported using third-party facial recognition; DHS officials said recognition served as an investigative lead, not sole basis for action, and said it did not scan the general public in the uses GAO reviewed. GAO found DHS was developing bias-assessment policy for AI technologies, while CBP, ICE and Secret Service had no policies specifically assessing bias from all their detection/observation/monitoring technology use. GAO notes choice of monitored place and protected activity can produce disparate impact independently of the classifier; a privacy impact assessment does not by itself supply instructions to staff on how to implement protections. DHS agreed to develop bias guidance; GAO said ICE and Secret Service’s then-existing privacy-policy responses did not meet two recommendations.

Comparison to Dru’s question: this gives a reason not to make ‘AI versus human bias’ a comparison of classification scores alone. The relevant baseline includes who chooses suspects and camera location, technology-enabled access by human analysts, which cases receive a review, and how the tool affects actions. No deployed language agent searched across a population or independently sent a report in this audit; the official category ‘AI’ itself includes much conventional recognition and analytics. Self-reports, fiscal-2023 coverage and no incident-level errors/appeals mean it cannot quantify an agent’s added risk or the current state of agency policy.

Next discriminating audit: take matched case samples under human-only, analytic-AI-assisted and autonomous-agent triage; record source inclusion and demographic coverage; count tool reads, referrals, officer actions, appeals, true protective results and erroneous consequences. Audit conventional workflows too, or a seemingly AI-specific improvement could leave bias elsewhere.

Linked to agent surveillance research brief, contested ICE attribution and algorithm versus human patrol map trial.