AiatSea: Who Pays the Inspector? The AI Industry Just Met Shipping's Oldest Question

Last week Dario Amodei, chief executive of Anthropic, asked the AI industry to slow down and promised one concrete step: outside evaluators working inside his company with access close to an employee's. On Friday he delivered it. Anthropic named Faculty, the AI business Accenture bought in January, as its first embedded evaluator. Faculty's people will test and try to break the models, check how they behave, test the safeguards, and report incidents. Both companies say they expect to invest at least $1 billion each over five years in this kind of work.

We said last week that an evaluator who cannot compel anything is not an auditor. So here is the fair scorecard. On the plus side, it happened in six days, the access described is real access to how models are built rather than a look at the finished product, and it is not exclusive: Anthropic says it is also talking to METR, the non-profit research group, and others. On the other side, Anthropic states plainly that there are no standards yet for what an embedded evaluator may see or how it should report, and no settled way to pay for it. So, in its own words, Anthropic will fund Accenture's work directly. The company being inspected pays the inspector. Anthropic says long-term funding should come from pooled or government sources. It also says the evaluators do not reduce its own responsibility for safety. Accenture's shares rose about 8% after hours.

Shipping recognises this problem immediately, because it is ours. The owner pays the class society. The yard pays for the plan approval. The equipment maker pays for the type approval, as TS Cyanergy did this week. Nobody in shipping pretends that arrangement is pure. What we did instead was build layers around it over 150 years, so that one paid relationship is never the only check: a flag State that answers to a government, port State control that inspects without warning and cannot be paid, IACS auditing the societies themselves, and insurers who pay when everyone else was wrong and so have their own reasons to look.

The AI industry has one layer so far, and it is the paid one. That is not a scandal. It is day one. But it is a useful reminder when a maritime AI supplier tells you their product has been "independently validated." Ask who paid. Then look for a second check that nobody paid for.

This week had one. Rijkswaterstaat's check of Shipping Technology's 5% fuel saving is exactly that second layer: a public body, measuring on real vessels, with nothing to gain. It is also the rarest thing in this newsletter. In seven stories this week, it is the only number checked by someone with no stake in the answer.

Three questions for your next AI contract. Who has tested this product other than the supplier, and who paid for that test? Is there any check on it that nobody paid for, from a regulator, an insurer or a customer's own trial? And if a result is disputed, whose measurement wins: the vendor's dashboard or yours? Put the answer to the last one in writing before the trial starts, not after.

Disclosure: this newsletter is produced with the help of Claude, an AI model built by Anthropic, the company discussed in this piece. We have tried to hold it to the same standard as every other company in this issue. [Anthropic · CNBC · TechCrunch · AiatSea, 13 September]

This story is part of the Maritime AI Digest of 20 September 2026.