News & Insights

From “Six Fictitious Case Citations” to “Glue on Pizza”: Global AI‑Hallucination Headlines Are Repricing Trust

Over the past three years, AI hallucination has not remained a technical term confined to laboratories; it has repeatedly surged into global headlines as a public event. All such incidents share one common feature: the model spoke with extreme confidence about facts that did not exist at all — and real people, real companies and real markets bore the cost.

I. Four Overseas Incidents That Changed Industry Rules

1. Six non‑existent precedents put “AI hallucination” into a legal ruling (2023, United States)

A lawyer used ChatGPT to retrieve case law on aviation accidents and submitted six entirely fictitious cases to court, complete with docket numbers, judges and verbatim judgment texts. In the ruling, the judge wrote the words “six fictitious case citations” directly and imposed a fine on the lawyer. This was the first time worldwide that a judicial body formally recognised AI hallucination; since then, virtually every article discussing AI trustworthiness begins with this ruling.

2. A single sentence — “put glue on pizza” — discredited AI search across the board (2024, Google AI Overviews)

Google’s newly launched AI search summaries advised users to add non‑toxic glue to pizza sauce to keep the cheese in place, and also recommended eating one rock per day to supplement minerals. Screenshots spread virally across global social platforms, and Google was forced to urgently take down large numbers of summaries. The destructive power of this incident lay not in its absurdity, but in the fact that it happened at a company holding the world’s largest index — enormous data volumes, and yet facts were still fabricated.

3. One astronomical howler wiped out a hundred billion dollars in market value (2023, Google Bard)

Bard claimed that the James Webb Space Telescope had captured the first image of a planet outside the solar system, and was immediately corrected by the astronomical community. On the day the statement was published, Google’s share price fell roughly 7% in a single session, wiping out about USD 100 billion in market value. For the first time, hallucination was quantified in the form of “capital‑market losses”.

4. An official customer‑service bot invented a refund policy, and the enterprise was ordered to honour it (2024, Air Canada)

Air Canada’s website customer‑service bot told a user that bereavement fares could be refunded retroactively. The user acted on that advice, was then refused compensation, and took the matter to court. The ruling was explicit: an enterprise must be responsible for the content stated by its AI customer service. This was the first precedent in which “AI hallucination entails legal consequences for the enterprise”, directly converting “hallucination” from a technical defect into a compliance liability.

In addition, Sports Illustrated was exposed for using AI to generate bylined articles and headshots attributed to non‑existent reporters, and its brand licensing was subsequently terminated. Nature has published multiple dedicated reports on AI mass‑fabricating references and invalid DOIs, forcing the academic publishing community to collectively revise its submission rules.

II. The Same Structural Problem Behind All Four Incidents

Placed side by side, these incidents are not four different kinds of error, but four outlets of one and the same mechanism:

Absence becomes fabrication. When no authoritative source can be retrieved, a general‑purpose large model does not return “no data available”; it completes the most plausible‑looking answer.

Formal credibility overwhelms substantive credibility. Docket‑number formats, academic citation formats, customer‑service scripts and policy wording are all reproduced perfectly, making counterfeits hard to detect even for professionals by visual inspection alone.

The accountable subject is blurred. The model said it wrong; the user assumed it was official; the official side assumed it was the model — and in the end nobody owned up, until the court stepped in.

Correction costs spill over. The cost of generating a sentence by a model is close to zero; the cost of verifying that sentence, correcting that sentence and paying compensation for that sentence is entirely transferred to users, enterprises and the judicial system.

表格

The conclusion drawn from international experience is clear: AI hallucination is not a problem that can be solved by “nudging accuracy up a few points”. What it requires is a set of factual infrastructure that is verifiable, traceable and accountability‑enforceable — not stronger generative capability.

III. Why This Matters Directly to Omnieyes

These overseas precedents correspond precisely to the core proposition that Omnieyes · Trust Archive exists to solve — ensuring that, before AI utters any sentence, there is already a factual source that can be checked, verified and held to account.

First, replace fabrication with archives. Omnieyes does not perform “probabilistic completion”. Instead, drawing on MDSE Meta‑Driven Casting technology, it builds an atomic‑fact archive repository, in which statutes, precedents, policies, rate schedules, clinical guidelines and similar information are anchored item‑by‑item to authoritative primary sources. If something cannot be found, it explicitly outputs “no authoritative factual record” — and never generates a “looks correct” answer.

Second, replace self‑certification with attestation. Powered by the KUCR four‑layer attestation algorithm, each fact undergoes atomic decomposition, source anchoring, status labelling and entitlement attestation. Outputs carry their own source, verification status and update time. The essence of the Air Canada case was that the enterprise could not prove “where on earth this sentence came from”; whereas every single output from Omnieyes inherently carries its provenance.

Third, replace one‑size‑fits‑all treatment with tiered archiving. Under the L0–L3 four‑tier archiving standard, the highest level of verification and dual‑track review is applied to high‑rigour scenarios (judicial, medical, financial), thoroughly separating “what can be generated” from “what can be relied upon” — precisely the direction that overseas regulators and courts are now demanding.

Fourth, replace blurred accountability with traceable accountability. The significance of the Trust Archive is not merely making content more accurate, but making every AI output traceable back to the accountable subject. Compliance is not a post‑hoc defence; it is a structure written into the archive.

IV. Conclusion

From a sheet of fictitious case citations, to a share‑price collapse, to a customer‑service promise that a court ruled must be honoured — the global market has already, with real money and judicial rulings, put a price tag on AI hallucination.

The race for generative capability is over. The race for trustworthy capability has only just begun. Wikidata and Baidu Baike serve human reference; Omnieyes · Trust Archive establishes credibility for AI. Trustworthy facts, traceable archives, attestable trust — this is not a slogan, but a lesson that these overseas precedents are now compelling every AI user to complete.