What happened

Education Accord NL was meant to be a big deal. It was the final report of an 18-month review of education in Newfoundland and Labrador - a 418-page document laying out roughly 110 calls to action and a 10-year plan for the province's public schools and post-secondary system. It was co-chaired by two Memorial University Faculty of Education professors, Anne Burke and Karen Goodnough, and released alongside the education minister in late August 2025. The province's reporting put the cost of the work at around $755,000.

Within weeks, the report's credibility collapsed over its bibliography. Memorial University academics reviewing the document found that at least 15 of its citations pointed to articles, documents, and authors that do not exist. The references had the unmistakable signature of generative-AI hallucination: titles in the right subject area, plausible-sounding authors, formatted correctly, and entirely fabricated.

The film that does not exist

The single most quoted example is almost too on the nose. The report cited a National Film Board of Canada film called "Schoolyard Games." The NFB has no such film. The citation did not even invent the movie from scratch; it reproduced, almost word for word, a fake reference that appears in a University of Victoria citation style guide.

That guide, like many, uses invented examples to show students how to format different source types. The "Schoolyard Games" entry is a deliberately fictional placeholder - a teaching dummy, not a real film. Its verbatim appearance in a government report is a fingerprint. A language model trained on the open web had ingested the style guide, treated its instructional fake citation as a real source, and dropped it into a 10-year education plan as supporting evidence. No human checking the reference against reality would have included it, because checking it would have taken one search.

Aaron Tucker, a Memorial University assistant professor whose research includes the history of AI in Canada, was among those who flagged the fabricated citations. He was careful about the inference: he could not definitively prove a language model wrote the references, but, as he put it, fabricating sources is a telltale sign of AI, and at minimum the fakes reflected a lack of attention and a lack of transparency. The fingerprints point hard at generative AI; what is certain is that no adequate human verification happened.

The irony

This is the part that makes the episode more than a routine citation error. The report does not merely happen to involve AI. Among its 110 calls to action is a recommendation that the province equip learners and educators with essential AI knowledge - including ethics, data privacy, and responsible use of technology.

A document advocating that students be taught to use AI ethically and responsibly was itself, by the strong appearance of the evidence, padded with sources hallucinated by AI and never verified by a human. The report telling the next generation to check their AI output did not check its own. If you wanted a single artifact to capture the gap between how organizations talk about AI and how they actually use it, this is a strong candidate.

The response

The co-chairs, Burke and Goodnough, said they did not know where the errors came from. Their position, as relayed to CBC, was that they could only assume the fabricated citations were introduced after they submitted the document - in effect, pointing responsibility at the government's handling of the report rather than their own drafting. The former education minister, Bernard Davis, denied that AI had been used at all, reportedly calling the suggestion preposterous.

Those two positions sit awkwardly together. If no AI was used, the verbatim fake film citation from a style guide needs some other explanation, and none has been offered. If AI was used somewhere in the pipeline - by the authors, by contractors, by someone touching the document before release - then the denials and the "we don't know how this happened" both describe a process where nobody owned verification.

The province acknowledged a small number of potentially incorrect citations and said the online report would be corrected. The document was, at least for a period, pulled from the Education Accord NL website pending those corrections. Coverage after a change in government indicated the incoming administration was not prioritizing a promised review of AI use in the report, despite the scandal - which is its own small lesson about how quickly these episodes drop off the agenda once the news cycle moves on.

The exact count of fabricated sources is reported as "at least 15," out of a reference list numbering in the hundreds, so the fakes are a minority of the citations. That is worth stating fairly: the report is not entirely fictional, and many of its references are genuine. But "only 15 of our sources were invented" is not the reassurance it might sound like. A reader cannot tell which 15 without checking all of them, which is exactly the verification work the authors were supposed to have done.

Why it matters

A citation is a promise. It says: this claim rests on real work you can go and read. When a government commissions a flagship report - the kind that is supposed to steer a province's education system for a decade - the citations are how the public verifies that the recommendations are grounded in evidence rather than assertion. Fabricated references do more than add a few bad entries to a list. They poison the reader's ability to trust the rest of the document, because once you know some of the sources are invented, you cannot take any of them on faith.

The deeper failure is a process one, and it is increasingly common. Generative tools make it trivial to produce text that looks like the output of careful research - complete with a confident, well-formatted bibliography of sources that were never consulted because they never existed. The only defense against that is human verification: someone clicking through to each citation and confirming it is real. That step is boring, it is slow, and it is exactly the step that gets skipped when a tool will happily generate a finished-looking reference list in seconds.

This incident sits alongside a growing list of AI-hallucinated citations in serious documents - consulting reports, government policy papers, court filings - that share one trait. In every case, the technology did the easy part (producing plausible text) and a human failed to do the hard part (confirming it was true). The "Schoolyard Games" film is funny in isolation. It is less funny as the load-bearing example of how a $755,000 plan for educating a province's children got its homework wrong.