Surviving AI or living well with AI?

Three meanings of "ethics" in AI

• Moral philosophy: what makes a human life and a human world worthwhile?
• Governance: codes, compliance, impact assessments and accountability.
• Safety: technical risk, such as robustness, alignment and security.

Governance and safety are necessary, but they should rest on ethics as moral philosophy, not replace it. Making a system reliable is engineering, not ethics. And an ethics whose only content is "don't let everyone die" is unquestionable but empty: survival is a condition for a good life, not the good itself.

How we got here

Modern applied ethics grew out of medicine (the Nuremberg Code, the Declaration of Helsinki) and left two habits: a protective one (do no harm, get consent) and a managerial one (committees, checklists, audits). AI ethics inherited both. A 2019 review found that the many AI ethics declarations converge on a handful of mostly negative principles, which invites "ethics shopping". He contrasted the risk-focused Asilomar principles with the Montreal Declaration, built through public deliberation around human development.

Ask which capabilities a system extends

He proposed the capability approach of Amartya Sen and Martha Nussbaum: judge a system by what people are actually able to be and do. Which human capabilities does it extend, and which does it erode?

Are we becoming Chinese rooms?

In Searle's Chinese Room, a person produces perfect answers in a language they don't understand. When we hand writing, coding and summarising to AI, do we risk the same? Every technology changes our skills. Two questions help judge a loss: is the skill transversal (like learning or critical judgement, which underpin other skills), and is the loss made up for by new capabilities?

Deliberate, list, prove

To turn ethics into practice: deliberate on what to require and why, list it as questions that must be answered, and prove it by giving an account to a third party. Deliberation improves the content (the Montreal Declaration's public process added values such as intimacy and sustainable development) and is the only legitimate way to settle collective trade-offs. His example: a system that detects psychological distress can't minimise both false alarms and missed cases. That balance should be decided with the people affected, not behind a closed door.