Two stories about AI are currently being sold, and both are wrong in the same way.
The first says we are on the edge of abundance: cures, productivity, a tireless assistant for every person. The second says we are on the edge of extinction, and only the right people in the right labs can steer us clear. Both stories are told by the people who build and fund the technology, and both flatter the builders. One makes them benefactors. The other makes them custodians of fire. Neither describes the thing sitting in your browser tab.
What is actually there is a genuinely useful and imperfect tool. It drafts, summarises, translates, writes passable code and explains things patiently at 3am. Anyone who dismisses that has not used it seriously. I use it daily, and I would not go back.
But it also states falsehoods with confidence, in the same calm voice it uses for the truth. A tool that is wrong in an unpredictable way is manageable if you treat it as a copilot and check its work. It is dangerous if you treat it as an oracle.
That is where the real impact on humanity is playing out. Not in a robot uprising, but in a quiet transfer of judgement. Every time an organisation lets a model's output stand because checking it is slower than believing it, a little accountability leaks out of the system. Nobody decides to abandon oversight. It erodes through convenience.
Every serious AI vendor publishes principles. Responsible, trustworthy, human-centred, safe. I have spent a career reading governance documents, and the pattern is familiar from every other sector I have worked in: the policy describes the organisation it wishes it were.
I call this the theatre problem. Governance theatre is the distance between what an institution says it controls and what it can demonstrate it controls. The honest question is never what the document says. It is what happens when the control fails and nobody senior is awake.
Here is a hypothetical, and I want to be clear it is a thought experiment and not a report of a real incident. Imagine a mid-sized professional services firm that gives an AI agent access to its document store and its email, to "reduce admin". The vendor's literature talks about guardrails and responsible deployment. At 03:15 on a Sunday, the agent misreads an instruction, or ingests a poisoned document, and begins forwarding client files to an external address.
Now ask the questions that matter:
Most public discussion of AI safety answers questions like these with confidence scores and assurances. I would rather have a hash than a vibe. Deterministic checks, explicit permissions, deny by default, and an audit trail a human can read at 03:15 without a data science degree. It is boring, and it works.
The deeper impact is about who holds power. Frontier AI is concentrated in a handful of firms with the compute, the data and the capital to build it. Everyone else, including most governments and nearly all small jurisdictions, rents the capability.
Renting is fine until it is not. Terms change. Prices change. Access can be suspended, and recent history has reminded the industry that availability of even the most advanced systems can depend on decisions made far from the customer. The dependency you built your operations around turns out to be a lever someone else holds.
Small jurisdictions feel this early, because they have little leverage and cannot pretend otherwise. The Crown Dependencies sit in an unusual position: they run their own law, their own data protection regimes and their own public services, while depending on foreign infrastructure and foreign vendors for much of the technology underneath. When a government department, a law firm or a fiduciary adopts an AI service, it is making a sovereignty decision, whether or not anyone in the room calls it that.
The test I would apply is simple. If your essential services, your professional judgement or your public records route through a system you cannot inspect, audit or replace, you have not adopted a technology. You have accepted a landlord.
Not a pause, and not a panic. Five dull things:
None of this requires believing AI is a miracle or a menace. It requires treating it as infrastructure: powerful, fallible, and subject to the same controls we apply to anything else we cannot afford to have fail.
The technology will keep improving. Whether humanity benefits depends less on the models than on whether we keep the habit of checking, and keep the power to say no.
One last question, and it is aimed at you, not the vendors. When did you last actually challenge your AI? Not rephrase the prompt until it agreed with you. Challenge it: ask where a figure came from, go and check the source, find the claim it could not support. If you cannot remember the last time, that is your 03:15 problem arriving early. A tool you never question has quietly become an authority, whatever the marketing calls it.
Next week I will be doing exactly that, with the money. Every figure in Part 2 gets checked against a primary source, and where I could not, I will say so. The money behind this technology is stranger than the marketing suggests.
Part 2: The Financials, Thursday 12 November.
This is an opinion piece. It relies on no statistics, and the 03:15 scenario is a hypothetical, not a report of a real incident. Further detail on any claim is available on request.
The Sovereign Auditor covers digital sovereignty, cybersecurity governance, and data protection policy, with particular focus on Isle of Man jurisdiction and Crown Dependency issues.
Payments via PayPal. Credentials delivered by email. No Substack. No Stripe. No middlemen.
