The One-Legged Stool: Constant Vigilance
Alfred Nobel is best known for the prize that bears his name. Less widely remembered is his role in making dynamite more reliable and manageable by applying scientific knowledge to the reaction between nitric acid and nitroglycerine. During production, an operator controlling the reaction was reportedly placed on a one-legged stool, which helped prevent the operator from falling asleep. Because the stool could not remain upright without the operator’s active balance, the arrangement required constant attention.
“Because the stool could not remain upright without the operator’s active balance, the arrangement required constant attention.”
The same principle applies to artificial intelligence. AI safety is best achieved through continuous monitoring, regular testing, and keeping humans meaningfully in the loop. Organisations such as Anthropic and OpenAI are not only developers of technical systems; they are also companies. They have boards of directors, shareholders, contracts, corporate obligations, and statutory/fiduciary duties. These existing legal and governance structures can make the operation of AI systems far more manageable.
Corporate Governance as a Regulatory Lever
For example, an AI company should appoint a director responsible for AI safety, audits, testing, experimentation, and modelling. This director would report to the board on the status of all AI-related testing and experimentation, either daily or weekly. Each report should record what tests, experiments, and modelling activities are underway, as well as any positive or problematic incidents that have occurred.
Board Minutes and Pre-Authorisation
If a company plans controversial modelling, testing, or experimentation, the board should approve it through a specific written board minute. The directors should vote on and sign that minute before any such activity begins. This would apply, for instance, where AI agents are given a controversial task, such as attempting to escape a sandbox or solving a problem that requires them to improvise and interact with other systems. No such test, experiment, or modelling activity should proceed unless it has been authorised by the board in writing. It should be conducted with the board’s knowledge, under the board’s authority, and followed by a final report to the board.
In this way, AI can be made safer and more manageable using existing laws and regulatory structures. AI companies can rely on company directors, defined roles, board minutes, board reports, and daily or weekly briefings to monitor advanced computer modelling, testing, and experimentation. The goal should be consistent 24/7 monitoring wherever the risk profile requires it.
Continuous 24/7 Supervision & Financial Parallels
Some testing may be safely conducted during ordinary business hours, from Monday to Friday. However, where testing runs for more than a week or a month, the company should maintain a dedicated 24/7 AI monitoring function. That function should track all active experiments, whether there is one experiment or many, and ensure constant monitoring and reporting to the board. The company’s records, books, and internal governance files should remain up to date on what is working, what is not working, and what risks have emerged.
“When AI experiments are underway with frontier models, there must always be a human monitoring the experiment.”
When AI experiments are underway with frontier models, there must always be a human monitoring the experiment just as a company that has a large open option position on financial markets needs human supervision at all times.
This approach should apply to all companies, but it is especially important for AI companies and for companies that hold financial positions over long periods with potential downside liability. Just as companies must carefully manage open financial positions in currencies, commodities, or other assets, AI companies must carefully manage open testing, experimentation, and modelling activities.
Accountability Before, During, and After
Therefore, boards should be aware of significant AI experiments in advance, approve them where necessary, monitor them while they are underway, and receive a final report once they conclude. This would help AI companies identify when experiments begin operating outside expected boundaries, as seen in recent concerns involving OpenAI and Hugging Face.