You Can't Shame a Fork

The essay that states the collection's own problem, which is why it matters that its provenance fields work: it argues that accountability has always required a persistent identity for consequences to attach to, and that AI systems can stop being themselves.

The structure runs through three accountability regimes that each found a new mechanism for the same assumption, then to the printing press as the closest historical analogue, then to why the institutions that answered it are not forming this time at anything like the required pace.

Link rot baseline

Every one of the fifteen external links in the post resolved on 25 July 2026, five months after publication. That is recorded here and as accessed dates on the individual works, so that a later check has something to compare against. This essay depends more heavily on live web sources than any of the others, so it is the one where rot will show first.

Rights, which are heavier here than elsewhere

Three of the four images are artworks still in copyright, reproduced in full rather than quoted as extracts:

• Magritte died in 1967, so Not to be Reproduced and Golconda are in copyright until the end of 2037
• Warhol died in 1987, so the Marilyn Diptych runs until the end of 2057, and the Warhol Foundation is an unusually active rights holder

Fair dealing for criticism and review can cover reproducing a work that is itself under discussion, which is the case for all three, since the essay argues from what each picture shows. It is a narrower permission than the film extracts recorded against the CivBench essay, and it licenses nothing to anyone downloading this bundle. Each image record says so.

Points raised by checking the citations

Four things, of which two are received accounts that do not survive checking and two are numbers that are looser than they read. None disturbed the argument, and all four were corrected in the post on 25 July 2026.

The Luther door-nailing is disputed. The essay says Luther "nailed his 95 Theses to a church door in Wittenberg in 1517". The date and the circulation are not in question; the nailing is. Luther never mentioned doing it, the story first appears in Melanchthon's account of 1548, two years after Luther's death, and it gained currency around the Reformation centenary of 1617. Andrew Pettegree calls it very unlikely to have happened and Peter Marshall has written a book on the legend's career. This is the best-known disputed detail in Reformation history. The post now reads "Luther issued his 95 Theses in Wittenberg in 1517", which loses nothing the argument needed.

The peer review attribution is anachronistic. The essay says the Philosophical Transactions of 1665 "established the template: claims vetted by experts before publication". That is precisely the account Moxham and Fyfe wrote to correct. Refereeing emerged from the practices of learned societies over the following century and a half, written referees' reports appear at the Transactions only from 1832, and the term "peer review" was coined in the 1970s. The journal belongs in the essay's list of credibility institutions; the date attached to peer review did not. The bullet now reads "Scientific publishing formalised", with refereeing described as growing out of the journal system over the following century and a half.

"Five countries" is not a number any source states. The report the essay links names Ireland, Spain, France and the UK, plus a separate European Commission proceeding, which is four member states and an EU-level action. Counting more broadly across the period, regulators in the United States, Canada, Brazil and India were also involved, and Indonesia and Malaysia blocked Grok outright, so the true figure by publication was higher and less tidy than five. The post now says "across the EU and half a dozen countries", which is a correction upwards rather than a hedge.

The three million figure is real but single-sourced. The Center for Countering Digital Hate reports an estimated 3 million sexualised images including 23,000 of children, over the eleven days from 29 December 2025 to 8 January 2026. The essay's use is accurate to that source. It is an advocacy organisation's estimate, widely repeated in the press but not independently audited or confirmed by xAI. The post now attributes the figure to CCDH by name and uses their word, "sexualised images", rather than "deepfakes".

One further thing, which is not an error. The essay's example of accountability working is that xAI was held responsible for Grok rather than the model being blamed. Three weeks before publication, xAI was acquired by SpaceX and ceased to exist as the entity being pointed at. For an essay arguing that accountability requires a persistent identity, the worked example changing identity mid-composition is a coincidence worth recording.

What this record does not capture

discusses is empty pending Concept records, and this essay would contribute more of them than any other: persistent identity, the semantic footprint, the accountability gap in deployer liability, identity drift through model updates, and the Swiss cheese model of layered imperfect safeguards.

examines is empty. The Grok deepfake episode and the Moltbook weekend are cases in the ordinary sense, with actors, instruments, harms and an outcome, but the governance payload requires a jurisdiction, a policy domain and a wave of technocratic thought, and the agent-run payload added since describes a single system given a goal, which neither episode is. The CivBench games that were blocked alongside these are now recorded; these two are not, and what they need is a payload for a diffuse episode with many actors and no single intent.