“While there's tons of evidence that they're impossible to fix, surely a safety-minded culture would be one that sought to fix or mitigate the most obvious and dangerous problem of Large Language Models.”
Our reading, the sentence that was checked:
There is extensive evidence that LLM hallucinations are impossible to fix, and they are LLMs' most obvious and dangerous problem.
Assessment: Unresolved Evidence as of
Where and when Ed Zitron wrote it
Where's Your Ed At, Put Up Or Shut Up, paragraph 34.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment wyea-put-up-or-shut-up-f54837-a1
We recorded 19 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-28. The source may have been edited after 2024-07-16. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Observation
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No horizon
- Deadline state on 2026-10-29
- No horizon (the release’s reference date)
From the speaker
- tons of evidence
- they're impossible to fix
- fix or mitigate
- the most obvious and dangerous problem of Large Language Models
Our interpretation
- They refers to hallucinations in the immediately preceding sentence.
- The explicit distinction between fixing and mitigating makes complete elimination the ordinary reading of fix.
- The observation concerns the research record available by publication, so the target period includes earlier research through 2024-07-16.
- Extensive evidence is a qualitative characterization of the literature, not a specified paper count.
- Most dangerous requires a comparison with other LLM risks. Most obvious has no objective ranking supplied by the speaker.
Counts as supported if: A substantial research record establishes general limits preventing complete elimination of LLM hallucinations, with comparative evidence supporting their claimed primacy in danger.
Counts as contradicted if: Research establishes a general method that eliminates hallucinations under the relevant scope, or comparable harm evidence clearly places another problem above hallucinations. Partial mitigation alone does not contradict the claim.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
Outcome: Unresolved
The record was checked, and the evidence available by the cutoff does not settle it.
Scored on a plain reading. The speaker named no measure for this claim, so it was judged on the reading an informed reader would take, stated below under our interpretation. Both reviews had to find that reading fair.
- Evidence as of
- Deadline state when assessed
- No horizon
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
Relevant impossibility research existed before the post, so the assertion has a real evidentiary basis. However, the cited arguments have defined scopes and assumptions. Detection and mitigation research does not refute a claim about complete elimination. The records do not settle the stronger characterization of the literature or the claim that hallucinations are the most dangerous problem. Partial support and remaining uncertainty do not justify a mixed verdict without actual contradictory outcome evidence.
Limitations
- Theoretical impossibility over unrestricted tasks is different from dependable performance on bounded tasks.
- No threshold was supplied for extensive evidence.
- The danger ranking lacks a shared harm measure.
Evidence (3 items)
Supports the claim: arXiv,
In this paper, we formalize the problem and show that it is impossible to eliminate hallucination in LLMs.
Why it matters here: This directly establishes that a pre-post paper argued for impossibility. Its formal setting concerns learning computable ground-truth functions; it does not by itself establish the strength of the entire literature or a comparative ranking of harms.
- Where
- Abstract
- Publisher
- arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: arXiv,
Therefore, different architectures and learning algorithms may mitigate these latter types of hallucinations.
Why it matters here: The paper derives a hallucination requirement for calibrated models generating certain arbitrary facts, while allowing mitigation and distinguishing other fact types. Those qualifications prevent treating its result as an unrestricted proof, but mitigation does not contradict Zitron's elimination claim.
- Publisher
- arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Nature,
Encouraging truthfulness through supervision or reinforcement has been only partially successful8.
Why it matters here: The study presents a method for detecting a subset of hallucinations and describes serious risks from unreliable answers. It establishes neither complete elimination nor that hallucinations outrank every other LLM danger.
- Where
- Abstract
- Publisher
- Nature
- Kind of source
- Peer-reviewed paper
- Describes events of
- Public from
What would settle it
A pre-post synthesis establishing the breadth and applicability of the impossibility results, plus comparable evidence ranking LLM harms. A general elimination result within the intended scope would contradict the technical component.
Where it was said (1)
Said once in the sources this register covers.
: Where's Your Ed At, Put Up Or Shut Up
While there's tons of evidence that they're impossible to fix, surely a safety-minded culture would be one that sought to fix or mitigate the most obvious and dangerous problem of Large Language Models.
- Where
- paragraph 34
- Published
- 2024-07-16
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment wyea-put-up-or-shut-up-f54837-a1: Current. Outcome: Unresolved. Evidence as of 2026-10-28. Public date not recorded.
Recorded events
Assessment publishedBy this site
First assessment published after both reviews and the final-text check passed.