Principles are the easy part
"Be helpful, honest, and harmless." "Avoid bias." "Respect user privacy." Almost every AI company's public principles converge on similar, largely uncontroversial language — because stating a principle costs nothing. The genuinely hard part is operationalizing it: translating "avoid bias" into a specific, measurable testing process applied before every model release, with a real mechanism to delay a launch if the tests fail.
Red-teaming: ethics as an adversarial exercise
The most concrete form AI ethics takes inside companies today is red-teaming — deliberately trying to make a system produce harmful, biased, or dangerous output before real users can, then fixing what breaks. It's a practical, empirical approach to a philosophical problem: instead of debating what "harm" means in the abstract, teams generate concrete failure cases and work backward.
The incentive problem
A genuine tension runs through corporate AI ethics: the team responsible for shipping a product faster is often a different team, with different incentives, than the team responsible for flagging ethical risk in that same product. Structural solutions — giving safety teams real veto power, external audits, regulatory requirements with teeth — matter more than any individual engineer's good intentions, because good intentions don't scale against a quarterly deadline.
Where regulation is actually landing
Rather than one global AI law, 2026 looks like a patchwork: disclosure requirements for AI-generated content in specific contexts (political ads, synthetic performers), sector-specific rules (healthcare, hiring), and security-focused guidance for autonomous "agentic" systems acting in critical infrastructure. This mirrors how other technologies got regulated historically — not with one sweeping law, but piece by piece as specific harms became visible.
What individual users can actually do
You can't audit a frontier model's training data, but you can apply the same skepticism to AI output that you'd apply to any single source: verify factual claims independently, especially for anything consequential (medical, legal, financial), and remember that an AI system optimized to sound confident is not the same thing as an AI system that is correct.