Seatbelts for the AI era: the precautions individuals, companies and states should take now
Precaution is not pessimism. It is what allows a powerful capability to be used confidently. A practical list, in three layers, that does not require anyone to slow down.
Cars became safe enough for mass use not by driving slower but by adding seatbelts, crumple zones and driving tests. The precautions below are in that spirit: measures that let people and institutions use AI more, not less, because the failure modes are bounded.
For individuals
Keep a human in the loop for anything irreversible — money, health, legal standing, relationships. Treat fluent confidence as a stylistic feature, not a signal of accuracy, and verify claims that matter through a second source. Be sparing with what you share: prompts can be retained, and personal data given to a model is personal data given to a company. Maintain skills you would not want to lose; the capacity to write, reason and calculate unaided is a form of independence.
For companies
Know what you are running. A register of AI systems in production — including the ones inside vendor products — is the precondition for every other control. Separate builders from approvers. Test before deployment against declared thresholds, and monitor against the same thresholds afterwards. Bound what agents may do: allow-lists of tools and actions, spend and scope limits, and logs that a third party can read. Prepare the incident plan before the incident: who can withdraw a model, how fast, and what customers are told.
For governments
Build assessment capacity, not just rules. A regulation nobody can test against is a statement of hope. Sovereign compute and confidential-computing options keep sensitive workloads and national data under domestic control without forgoing frontier capability. Invest in AI literacy across the civil service, because officials who cannot ask the right questions cannot supervise the answers. And establish independent, accredited validators — the equivalent of testing laboratories and auditors — so that conformity claims can be checked by someone other than the claimant.
Precaution is the price of confident use. Paid up front, it is cheap.
The common thread
Every precaution above reduces to the same principle: make the system's behaviour knowable, and make someone accountable for knowing it. Individuals do that by verifying and keeping a hand on the wheel. Companies do it with registers, tests, separation of duties and monitoring. States do it by building the institutions that can check. None of it is heroic. All of it is overdue.
