Ethics, Policy & Safety

The policy and safety layer of AI is not abstract: it determines which systems get deployed in courtrooms, hospitals, and public infrastructure, and what accountability exists when they fail. This cluster covers AI safety research, algorithmic harm cases, regulatory divergence, and the transparency collapse at frontier labs.

The Stanford HAI 2026 AI Index puts the transparency problem in specific numbers: average FMTI transparency scores dropped from 58 to 40 in a single year, a 31% collapse. Frontier labs are simultaneously asking for public trust and disclosing less about training data, evaluation methodology, and safety testing. The gap between claimed safety assurances and auditable evidence is widening, not narrowing.

Regulatory divergence between the EU and US has moved from policy debate to compliance reality. The EU AI Act’s risk-tier framework is now being implemented—requirements for high-risk systems, banned applications, and conformity assessments—while the US moved toward lighter-touch executive guidance. The result is a genuine Balkanization problem for AI products operating across jurisdictions.

Algorithmic harm cases are no longer edge cases. Groundy has covered the wrongful arrest pattern in AI facial recognition—at least eight documented cases, nearly all affecting Black individuals, in a system where NIST studies show 10-100x higher error rates for darker-skinned faces. The issue is not only the technology; it is the institutional trust police departments extend to vendors whose accuracy claims don’t survive demographic auditing.

Constitutional AI, RLHF alignment, and safety via output filtering represent different philosophical approaches to the same problem: how do you make models that reliably refuse harmful requests while remaining useful? Groundy covers the research critically—distinguishing genuine safety gains from sophisticated filtering that degrades at the edges of distribution.

Scientific integrity is an underreported ethics story. AI-enabled paper mills are flooding academic literature with automated fraudulent research at a pace peer review cannot match; the journals most affected are the ones where AI tools also do most of the checking. When AI enables fraud and AI is supposed to detect it, the feedback loop is not self-correcting. Groundy covers the governance implications—for publishers, for researchers, and for AI training pipelines that ingest scientific literature.