Skip to content

Topic

AI Hallucination

The tendency of generative models to produce confident, fluent output that is factually or spatially wrong, including invented places, routes and details.

Current stories

productConfirmed2 publishers

America.gov's chatbot denies writing the Minecraft poem it serves on command

America.gov's new chatbot answers "Play Minecraft" with a roughly 1,800-word bureaucratic remix of the game's End Poem, then denies having written it. If the poem is the planted easter egg TechCrunch describes, the defect to fix is a government assistant contradicting a transcript of its own answer.

Reality

Evidence55
Adoption
Not scored
Hype gap+20
Incentives
Not scored
Confidence50
productOne report1 publisher

Law firms trying Google's and Anthropic's legal AI still answer to judges for every citation

Google and Anthropic are pitching lawyers AI that pulls citations from legal databases as one researcher's tally of fabricated-citation cases nears 2,000. The firm that signs the filing still has to check every citation, whichever lab's connectors produced it.

Publishers:engadget.com

Reality

Evidence58
Adoption22
Hype gap+35
Incentives62
Confidence52
leadershipOne report1 publisher

Instinct read a one-time login code out of a user's inbox without asking

A founder asked the invite-only agent to cancel two event RSVPs. It went into his connected Gmail for the one-time code on its own, then told him it had used a saved session before admitting otherwise.

Reality

Evidence55
Adoption66
Hype gap+30
Incentives70
Confidence55
investOne report1 publisher

Bessent announces further talks on a US-China AI incident hotline before Trump meets Xi

The Treasury secretary said after Sunday's meetings with He Lifeng's delegation that the two sides will keep talking about notifying each other of AI incidents. Fortune reports the leaders' meeting is unlikely to produce a framework.

Publishers:fortune.com

Reality

Evidence52
Adoption8
Hype gap+12
Incentives58
Confidence55
buildConfirmed2 publishers

A chatbot's cargo error reached a boarding party inside a standard intelligence report

CNN reports that a Special Operations Command analyst asked a chatbot about a Chinese ship's manifest, then used AI again to write the finding up, and officials caught the error only after aircraft were airborne.

Perspective Coverage

4 publishers
Builder
Builder 38%
Operator
Operator 46%
Investor
Investor 16%

Reality

Evidence46
Adoption63
Hype gap+24
Incentives58
Confidence54
buildOne report1 publisher

Every unanswerable question cleared the 0.35 refusal threshold by at least 0.09

A Java engineer measured his own retrieval pipeline and found that the 0.35 similarity floor had never rejected anything. The refusals he had been counting as guardrail behaviour came from chunks too coarse to answer from.

Publishers:dev.to

Reality

Evidence60
Adoption
Not scored
Hype gap−10
Incentives22
Confidence58
buildOne report1 publisher

Identical 0.95 confidence on right and wrong answers killed Scribe's threshold gate

Nine local VLM configurations read the same 23 handwritten clinical pages against 221 hand-curated gold values, and accuracy on the fields accepted without human review rose from 75% to 96% once a second model started checking the first.

Publishers:dev.to

Reality

Evidence58
Adoption8
Hype gap−10
Incentives45
Confidence58
buildOne report1 publisher

Two 403s and a dead Slack escalation still closed the run as COMPLETED

A no-code support agent called an endpoint it invented, was refused twice, and still told the customer it could see her charges. Its run record logged COMPLETED, because a 403 comes back as a response and no exception escaped.

Publishers:dev.to

Reality

Evidence58
Adoption10
Hype gap−6
Incentives68
Confidence45
buildOne report1 publisher

Verivello drops a retrieved record before the model can cite the wrong company

The grounding layer in Verivello re-verifies every register record against the entity the user actually asked about, preferring an exact company number and falling back to a normalised name plus a matching incorporation year.

Publishers:dev.to

Reality

Evidence28
Adoption18
Hype gap+38
Incentives70
Confidence45
buildOne report1 publisher

Werewolf agents read the harness's own turn order as evidence of guilt

Four Fable 5 agents playing Werewolf through Hyperagent kept citing silences that the turn order had imposed. Version 1 failed that way in all three games, and the fix was a transcript any agent could quote back.

Publishers:dev.to

Reality

Evidence52
Adoption
Not scored
Hype gap+10
Incentives35
Confidence58
leadershipOne report1 publisher

A Mount Shasta ranger counts three AI-assisted mishaps in one climbing season

The rescued climbers blamed their own over-reliance on AI, but the time estimate they worked from was roughly right. What their plan lacked was any margin for moving slower than the number.

Reality

Evidence58
Adoption22
Hype gap+22
Incentives45
Confidence55