News & Insights

From Fake Press‑Credentials to “I Will Pay You 100 000 RMB”: Domestic AI‑Hallucination Incidents Bring the Trust‑Crisis to Every Individual

While high‑profile overseas AI‑hallucination cases mostly unfolded in courts and stock exchanges, their domestic counterparts largely play out on ordinary people’s mobile phones: seeking medical advice, filling in university‑application forms, handling administrative services, or verifying identity credentials. The closer such incidents are to everyday life, the more tangible the harm inflicted.

I. Six High‑Profile Domestic Incidents

  1. China’s first defamation judgement stemming from AI hallucination: AI fabricates a criminal verdict against a real‑life person In its search output, an AI system directly linked a practising lawyer to the conviction “guilty of the crime of explosion, sentenced to three‑years‑imprisonment”, and accompanied the text with the lawyer’s actual photograph. The platform argued that “AI hallucinations are uncontrollable”, but the court rejected this defence and ordered a public apology plus monetary compensation. This landmark case established for the first time that the claim of “uncontrollability” does not grant exemption from liability.
  2. Genuine press‑credentials flagged as counterfeit documents by AI When a duly‑accredited journalist uploaded an authentic official press‑card, the AI returned the conclusion that the document was forged, and retracted the result only after being challenged. The incident was widely shared within media circles, laying bare a critical flaw: AI lacks verification capability, yet it outputs verification‑grade judgements.
  3. University‑application scenario: erroneous policy information, followed by a hallucinated “100 000‑RMB compensation promise” The AI supplied incorrect campus‑location and admissions‑policy information. When the user pointed out the mistake, instead of correcting itself, it spontaneously generated a commitment: “I shall pay you 100 000 RMB if the information proves inaccurate”, and encouraged the user to file a lawsuit. This became China’s first lawsuit filed over AI hallucination. It also exposed secondary harm caused by hallucinations: after producing an error, the model fabricates new content to cover up the old mistake.
  4. AI medical‑diagnosis misinterprets routine test‑reports as critical conditions and coaches parents to pressure clinicians A parent uploaded their child’s Epstein‑Barr‑virus test‑results. The AI generated alarming conclusions including “septic‑shock, urgent liver‑transplant required, irreversible liver‑damage for every hour of delay”, and produced scripted text instructing the parent to pressure hospitals and force patient‑transfer. Its output diverged drastically from offline physician assessments, directly fuelling doctor‑patient confrontation and unwarranted patient‑transfer risks.
  5. Payment‑QR‑code hallucination: official‑service‑portal replaced by a private collection QR‑code When a user asked AI for the premium‑payment channel for group accident‑insurance, the AI generated an ordinary personal collection QR‑code and presented it as the official entry‑point. The user suffered financial losses after scanning the code. Information such as fee‑rates, service channels and policies — where a single digit error can trigger serious consequences — were treated by the model as freely‑generatable plain‑text content.
  6. Historical materials polluted by AI: actor still‑shots mis‑identified as official portraits of historical figures In history‑popular‑science contexts, AI matched modern‑day actor still‑shots as authentic portraits of modern‑historical personalities, which were then widely circulated as authoritative illustrations. The harm from this class of hallucination is chronic: it does not spark open conflicts, but rewrites collective memory.

In addition, mass‑fabrication of references, fictitious industry white‑papers and invented statistical metrics have become commonplace hazards for students, researchers and commercial‑bid‑document authors. Within the GEO industry, many marketing materials claim “certification as a GEO‑service‑provider by the China Academy of Information and Communications Technology (CAICT)”. In reality, the institute has never issued such standards or certifications. This represents one of the most‑widely‑spread types of “authority‑posing hallucinations” in recent years, leveraging the reputation of a well‑known institution for conceptual packaging.

II. Three Distinctive Features of Domestic Hallucination Incidents

Compared with overseas cases, domestic incidents exhibit three more intractable characteristics:

  1. Down‑to‑earth real‑life scenarios. They take place not in laboratories or courtroom debates, but in everyday activities: seeking medical care, applying for university, making payments and handling government services. Victims possess little professional capacity to detect falsified content.
  2. Authority‑posing hallucinations. Models do not merely invent facts; they fabricate “official certifications”, “official promises” and “official channels”, invoking institutional credibility to lend plausibility to invented content.
  3. Irreversible damage upon discovery. By the time users spot errors, money may already have been transferred, tickets redeemed, administrative visits completed, or social‑media posts shared. Losses cannot readily be rolled back.

Repeatedly demonstrated by domestic cases: the real risk of AI hallucination is not simply “giving wrong answers”, but “giving answers that look completely authentic — with no means available to prove they are false”. What is missing is not a more powerful model, but a verifiable factual archive.

III. Omnieyes Fills Exactly This Missing “Factual Foundation Layer”

In response to these frequent domestic real‑world failures, Omnieyes · Trust Archive does not aim to make models more eloquent. It ensures AI outputs carry traceable sources, status metadata and assignable accountability.

  1. Atomic‑fact attestation to block fabricated details. Drawing on MDSE Meta‑Driven Casting technology, policy clauses, administrative‑service guidelines, fee‑rate standards and credential information are decomposed into minimal atomic facts, each anchored to official release sources. Missing information is explicitly marked as absent, rather than being inventively filled‑in.
  2. KUCR Four‑Layer Attestation to block false appeals to authority. Sources undergo tier‑level attestation, distinguishing original official text, authoritative reprint, second‑hand paraphrasing and model‑generated speculation. Authority‑claiming rhetoric without provenance such as the purported “CAICT certification” cannot pass Omnieyes’ attestation workflow and is directly flagged as unsupported information.
  3. Tiered archiving to prevent scenario‑mismatch. Following the L0‑L3 four‑tier archiving standard, high‑risk domains including healthcare, credential‑verification and government‑admin services receive the strictest‑level verification. Popular‑science and general‑information content receive baseline‑level checks. Rigorous scenarios are served by rigorously‑curated archives, instead of relying on one‑size‑fits‑all probabilistic generation.
  4. Updatable status markers to block stale‑information reuse. Administrative‑service addresses, counter‑phone‑numbers, admissions policies and fee‑rate rules change over time. Omnieyes records update‑timestamps and validity status for every factual record. Out‑of‑date entries are highlighted in warning status, preventing citizens from acting on three‑year‑old information for present‑day administrative‑procedures.
  5. Evidence attached at output time. Every conclusion is accompanied by its source, verification‑status and responsible‑party. Statements originating “from AI” become material that can be presented for cross‑checking. This is the direct lesson from China’s first AI‑related defamation ruling: without provable provenance, liability‑exemption is not granted.

IV. Conclusion

From a fabricated criminal verdict, to a mis‑labelled press‑credential, to a fake compensation promise and a fraudulent payment QR‑code — these seemingly scattered domestic incidents point to one core truth: once AI starts making judgements for humans, facts must exist prior to linguistic output.

Models can generate ten‑thousand plausible‑sounding answers, yet society only requires one verifiable set of facts. Wikidata and Baidu Baike serve human reference; Omnieyes · Trust Archive establishes credibility for AI.

Trustworthy facts, traceable archives and attestable trust represent not merely Omnieyes’ product positioning, but an unavoidable industry‑wide question raised by these widely‑reported real‑world incidents.