Don’t make people the safety net
The writer Cory Doctorow has a useful way of describing how people and machines can work together. In automation theory, he writes, a centaur is “a person who is assisted by a machine”. A reverse centaur is the opposite: “a machine head on a human body”, a person whose job is to serve the machine rather than be served by it.1
His sharpest example is a radiologist asked to sign off on an AI’s diagnoses at a pace no one could keep up with. When the AI misses a tumour, it is the radiologist’s signature on the report. Their real job, Doctorow argues, “isn’t really to oversee the AI’s work, it’s to take the blame for the AI’s mistakes”. Borrowing Dan Davies’ term, he calls this an accountability sink.
The same trap in research publishing
Anyone who has worked on equity research will recognise the shape of it. A report goes out under deadline, in two languages, full of figures that have to agree with an earnings model, the company’s own filings, the tables and each other. Somewhere in it there may be a transposed digit, a stale target price, or a 当社 that should have been 弊社.
The traditional answer is a person: a translator reading every line, a supervisory analyst or 審査担当者 checking every number. Adding AI to the drafting does not change that answer if all it does is produce more text for the same person to check, faster. As Doctorow points out, AI errors are also the hardest kind to catch, because they are designed to look plausible.2
Asking people to stay vigilant over every detail of a complex document is not a division of labour. It moves the blame.
Redressing the balance
We want to play to the strengths of both software and people, rather than transferring responsibility from one to the other. Software is good at tireless, exhaustive checking. People are good at judgement.
That is the division of labour our products are built around:
- Our AI translation editor highlights numbers and terms that don’t match between Japanese and English, and flags company names that aren’t in the source, so translators are not hunting for them line by line.
- Second, our SA precheck app, reconciles every figure in a draft against the earnings model and the company’s TDnet and EDINET filings, and applies house style and review rules before a report reaches review.
In both, every finding shows its reason, and a person decides what to do with it. The software checks; the expert decides.
When you evaluate any AI tool for research, it is worth asking a simple question: who is doing the checking, and who will carry the blame?
Gareth Edwards
CEO, Sentience Inc.
Footnotes
-
Cory Doctorow, “The Reverse-Centaur’s Guide to Criticizing AI”, Pluralistic, 5 December 2025. ↩
-
Cory Doctorow, “‘Humans in the loop’ must detect the hardest-to-spot errors, at superhuman speed”, Pluralistic, 23 April 2024. ↩