As I think through AI governance and auditability, particularly trying to anticipate what regulation might look like in the era of agentic AI, the mind naturally reaches for known parallels.
First, my engineering training drew me to Heisenberg’s Uncertainty Principle. Just as the act of measuring a particle’s position disturbs its momentum (and vice versa), the very act of auditing or testing a model can alter it – either directly via fine-tuning, or indirectly as model builders optimise for the test parameters. At present, it’s impossible to capture all aspects of model behaviour at once. Focus on explainability, and I might miss bias. Change the prompt, framing, or evaluation method, and the model behaves differently. I will never have the complete picture; like a particle physicist, I can only approximate. (And yes, my Heisenberg parallel isn’t original – wiser minds have got there before me.)
My economics training then brought me back to Goodhart’s Law. As the former Bank of England economist put it (with classic policymaker clarity), “Any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes.” When regulators fixate on a single metric, businesses adapt to hit the target rather than the spirit behind it. The parallels with AI are obvious – if regulation defines fairness, accuracy, or explainability too narrowly, the market will optimise for those benchmarks at the expense of real outcomes.
But the most helpful insight came from my art history studies. Time and again, we see how radical innovation hardens into orthodoxy once institutions codify it. Renaissance perspective, Impressionist brushwork, Cubist form – all began as disruptive ways of seeing the world. Yet once absorbed by academies and critics, they calcified into rules to be followed, cliches to be reproduced. Creativity became convention.
The risk is clear – if we reduce AI regulation to fixed formulas or metrics, we risk turning an evolving discipline into a compliance exercise. The challenge for policymakers and practitioners like me alike is to design governance frameworks that provide trust and accountability without impacting outcomes.
This is where the idea of curation feels most useful. In art, a curator doesn’t dictate what the artist creates, nor do they reduce works to a checklist. Instead, they frame, contextualise, and guide interpretation – balancing preservation with the freedom for new movements to emerge. Curation is active but not domineering; it respects uncertainty while providing enough structure for meaning to surface.
AI may require the same stance.
Regulators, practitioners and auditors are but curators of AI –
1. Framing : We set boundaries and standards of safety, without dictating the exact form innovation must take.
2. Contextualising : We ensure the organisation understands the risks and trade-offs of each AI module.
3. Preserving : We safeguard against harm, threats and abuse, much like protecting fragile works of art.
4. Inviting interpretation : We leave space for multiple approaches, perspectives, and future breakthroughs.
If regulation becomes curation rather than codification, we may find a balance between accountability and adaptability. The role of governance, then, is not to eliminate uncertainty, or to audit by numbers, but to shepherd AI’s evolution in a way that both protects and provokes – much like the best curators do with art.