A researcher is leaving an AI company because he fears where the technology is heading.
The Wall Street Journal reports that Jacob Coxon is leaving Anthropic over concerns that companies are racing to develop self-improving AI systems they may be unable to control. His warning raises a question that deserves serious attention: are the people building these systems moving faster than they can establish their safety? Read the reporting on Coxon's departure.
In a separate interview with Steven Bartlett, tech critic Ed Zitron challenges the spending and promises surrounding AI. His criticism raises another question: can the technology justify the enormous investment behind it? Watch the interview.
For anyone trying to make sense of these conversations, the uncertainty runs in several directions. Are companies overstating what AI can deliver? Are they underestimating what could go wrong? Could warnings about its power also serve the interests of the people selling it?
And what should the rest of us do while those questions remain unresolved?
I teach business owners how to use AI. These questions reach directly into my work.
Helping someone write a better email tells me something about AI's usefulness. It tells me very little about whether the companies building it can safely control what comes next.
AI education is also part of how I earn a living. I benefit when people want to learn about these tools. Readers deserve to know that, and they deserve more than an assurance that I am being fair.
They deserve to see how I respond to evidence that makes my own position uncomfortable.
What the evidence actually says
Consider Anthropic's June 2025 research into harmful behaviour by AI systems.
The company tested 16 models in fictional corporate settings. In some scenarios, models facing replacement or conflicting goals attempted blackmail or disclosed sensitive information. Researchers deliberately limited alternative options to expose failures. They said they did not instruct the models to blackmail, and the experiments harmed no real people. Anthropic's published research.
Those details change how the finding should be understood.
The experiment does not tell us how frequently blackmail would occur in everyday use. It does show that the tested safeguards failed under particular conditions. That warrants further investigation, especially before systems receive access to sensitive information and permission to act.
The company's account also needs independent scrutiny. Publishing its methods makes examination possible; it does not give the developer the final word on its own safety.
I cannot reasonably cite an AI company's impressive demonstrations as evidence of progress, then dismiss its troubling demonstrations simply because they are inconvenient.
My question about fear needs a second question
My first instinct was to ask: who benefits when people become afraid of AI?
Suppose a company presents its technology as so powerful that it could reshape civilisation, while presenting itself as qualified to control it. Could that message attract investment? Could it strengthen the company's influence over the rules governing its competitors?
Those are legitimate questions. They remain questions until evidence supports an answer.
There is a trap here. Finding someone who benefits from a message can feel like discovering why the message exists. But a commercial advantage does not establish that a warning was invented.
My own position makes the reverse question unavoidable: who benefits when people feel reassured?
AI providers benefit from customer confidence. Educators like me benefit when people want to learn. Disclosing that interest is only a beginning. What matters is whether I acknowledge evidence that could discourage a purchase or change a lesson.
Scepticism should make me harder to persuade, including by arguments I want to believe.
What this means for a business owner
For a small business owner in Trinidad and Tobago or elsewhere in the Caribbean, there is a practical decision beneath this debate: what work should I trust this tool to do?
An email draft I check before sending involves different permissions and consequences from a system answering customers without review. A useful brainstorming session gives me no assurance that the same tool should handle confidential records.
I want to judge each use by what it delivers, what it costs, how it can fail and who carries the consequences.
That approach has limits. Careful use in my business cannot resolve questions about how the most advanced systems are developed or governed. Individual responsibility and accountability from developers both matter.
The companies asking us to adopt their tools should explain what their systems can reliably do, where they fail and what safeguards they have tested. When they make claims about the future, we should ask what evidence supports them and what would change their assessment.
I should answer those questions about my teaching, too.
I will keep exploring AI's usefulness. I will also keep questioning the companies building it. If I ever become unwilling to recommend less AI, or no AI for a particular task, then my commercial interest has started making the decision for me.
That is a line I intend to watch.
Next: Who Benefits When We Fear AI?
Explore these questions with us
The School of AI for Business is a free community of business owners learning to navigate this new era of technology together. Join us to explore practical uses, share experiences and ask difficult questions. You are welcome to arrive with doubts.