Who Is Accountable When Anyone Can Build AI?
As building with AI gets easier, more people take consequential actions with it. Four questions to answer before you build or deploy anything with AI in it.
Writing
Essays on AI policy and ethics.
10 Questions on Humanity and AI I’m Still Struggling With
If AI saves us time, what are we doing with it? Ten open questions on delegation, consent, dependence and thinking for ourselves that I have not answered yet.
AI as Infrastructure: Open Questions on Reliance and Harm
We are trying to govern a technology while our relationship with it is still being negotiated. Three questions on infrastructure, dependence and what counts as harm.
Formal Control, Real Dependence: Human Oversight and Automation Bias
A person can remain formally in control of a decision while losing the ability or confidence to exercise independent judgment. What oversight rules should account for.
Consent Without Comprehension: Agreement in AI Products
74% of participants in one study skipped a privacy policy, yet 97% agreed to it. Why a checkbox is not meaningful consent when AI systems use personal data.
Challenger, 1986: Deployment Without Ongoing Governance
The shuttle had flown 24 missions without incident, so it was approved to fly again. What the Challenger launch decision shows about governing AI after deployment.
The Hawthorne Effect and the Limits of AI Evaluation
Workers in 1924 changed their behaviour because they were being watched. Some AI systems now do the same, which raises hard questions for testing and oversight.
When Intelligence Is Abundant, What Stays Scarce?
Ford made the knowledge of building a car abundant, and the advantage moved to judgment. The same shift is happening with AI, and it changes which skills matter.
Apollo's Human Computers and the Quiet Cost of Automation
NASA engineers once worked every trajectory by hand. When machines took over, nobody called it a loss. What cognitive offloading means for people and for models.
Bias in AI Systems: Five Ethical Tensions
AI inherits bias from our decisions and data, then amplifies it. Five tensions: privacy and personalisation, transparency, fairness, autonomy, and reliability.
Model Collapse: The Risk of Training AI on AI Output
The Habsburgs stayed in power through intermarriage until inbreeding depression caught up with them. Models trained on their own output face a similar decline.