In the first week of January 2017, the most influential people working in artificial intelligence gathered at a small conference centre on the California coast. Asilomar is the kind of place that hosts academic retreats and yoga weekends. For three days that month, it hosted what may have been the most consequential meeting about AI ever held.
A short version of the guest list reads like a who’s-who of the field. Stuart Russell, who wrote the textbook most computer science students still learn AI from. Demis Hassabis, founder of DeepMind. Yoshua Bengio and Yann LeCun, two of the three so-called godfathers of deep learning. Ilya Sutskever, then chief scientist at OpenAI. Elon Musk. More than one hundred researchers, ethicists, lawyers and economists in total.
What came out of the meeting was a document called the Asilomar AI Principles. Twenty-three short statements, organised into three groups. Five principles on research issues. Thirteen on ethics and values. Five on longer-term issues. The open letter has since been signed by nearly two thousand AI researchers and several thousand others, including Stephen Hawking, who added his name from Cambridge a year before his death.
The principles are nine years old. They were written before ChatGPT. Before the transformer architecture had been published for a year. Before most of the public had heard of generative AI. And reading them today, they have aged remarkably well. Better, in some respects, than the regulation that has been written since.
It is worth pulling a few out in detail.
The fourteenth principle is called Shared Benefit. It states that AI technologies should benefit and empower as many people as possible. Read that in 2026, with the productivity gains from AI accruing largely to a handful of companies and a much smaller handful of individuals, and you start to feel the gap between intention and reality. The principle does not say AI should be free or that it should be open. It says the benefit should be widely shared. Whether the current distribution is moving in that direction is a question every business leader deploying AI ought to have an answer to.
The sixteenth is Human Control. Humans should choose how and whether to delegate decisions to AI systems, to accomplish human-chosen objectives. For mortgage advice, lending decisions, fraud screening, recruitment, this is not abstract. It is the operational question that determines whether Consumer Duty is actually being met. If a customer is being declined by an algorithm, who chose to delegate that decision? Who can override it? Who is accountable when it gets it wrong? Asilomar was clear in 2017. The answer is humans, on every count.
The eighteenth is the AI Arms Race principle. An arms race in lethal autonomous weapons should be avoided. Nine years on, this one reads less as guidance and more as elegy. The signatories knew, even then, the direction things were heading.
The twenty-second is the one that should keep AI researchers awake at night, and indeed it does for many of them. AI systems designed to recursively self-improve or self-replicate in a manner that could lead to rapidly increasing quality or quantity must be subject to strict safety and control measures. Read in the context of Sakana AI’s automated research system, or DeepMind’s mathematics agents that train themselves in the middle of a competition, this principle is no longer theoretical. The systems exist. Whether the strict safety and control measures exist around them is a different question.
Why does any of this matter for a UK business audience?
Three reasons, none of them academic.
The first is procurement. When you buy AI tools for your firm, the principles give you a vocabulary for asking the right questions. Does this system have failure transparency? Will it explain its decisions in a way a regulator would accept, and is human control built into the design rather than bolted on after?
The second is regulation. The UK government’s own five AI principles, set out in its 2023 White Paper, and the strategic approach the FCA and Bank of England published in April 2024, sit in the same intellectual tradition as Asilomar, even when they do not cite it. Understanding the source helps you read the regulation. The Asilomar Principles are upstream of much of what your compliance team is going to be dealing with for the next decade.
The third is judgement. Most of the genuinely hard decisions you will face about AI in the coming years will not be answered by ticking compliance boxes. They will be questions about whether something is a good idea, whether it is fair, whether it respects the customers and colleagues who depend on you. Frameworks like Asilomar are not laws. They are a starting point for thinking like the people who built these systems and worried about what they had built.
Most of the principles were written before transformers existed. They have held up. The question is whether they will hold up for another nine years, as the systems they were trying to govern start, in small but real ways, to design themselves.
That is not a comforting note to end on. But it is probably the honest one.
The full list of principles, and the signatory list, is hosted by the Future of Life Institute at futureoflife.org/open-letter/ai-principles.