“Hallucinations (which should really be called confabulations) will continue to haunt generative AI.”
Our reading, the sentence that was checked:
Through 2025 hallucinations will continue to plague generative AI.
Assessment: Supported Deadline passed on 2026-10-29 Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, 25 AI Predictions for 2025, from Marcus on AI, paragraph 30.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-25-ai-predictions-for-2025-from-marcus-64d30d-a3
We recorded 37 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 2025-01-01. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Forecast
- Deadline, as the speaker wrote it
- “end of this year”
- Deadline state when assessed
- Deadline passed
- Deadline state on 2026-10-29
- Deadline passed (the release’s reference date)
From the speaker
- Hallucinations
- will continue to haunt generative AI
- where we will be at the end of this year
Our interpretation
- Plain reading: through the end of 2025, hallucinations (confident false outputs) stay a well-documented, consequential problem for leading generative AI models and their deployments. They need not be unchanged or rising. They need to remain a recognized, material problem.
- Evidence can be lab evaluations of leading models, statements by the labs themselves, or documented real-world incidents during 2025.
Counts as supported if: Credible 2025 evidence shows that leading models still hallucinate at material rates, or that the problem keeps causing documented harm, and that the labs themselves still treat it as unsolved.
Counts as contradicted if: Credible evidence shows that by late 2025 hallucinations in leading models had become rare enough that they were no longer a significant practical problem, for example near-zero rates on standard factuality evaluations and no notable incidents.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Did the forecast come true?
Outcome: Supported
The dated evidence matches what the claim committed to, under the criteria frozen before grading.
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
- Deadline passed
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
The forecast period ended on 31 December 2025. OpenAI's system card of 16 April 2025 gave PersonQA hallucination rates of 0.33 for o3 and 0.48 for o4-mini, against 0.16 for o1 (e1). For o3 compared with o1, OpenAI wrote that more research was needed to understand the cause (e1). On 4 September 2025, OpenAI researchers wrote that hallucinations persist even in state-of-the-art systems and undermine trust (e2). On 6 October 2025, The Register reported that Deloitte had agreed to refund part of an Australian government contract. Deloitte had admitted using generative AI to produce the report, which contained fake citations and a made-up court quote (e3). In August 2025, OpenAI said GPT-5's responses were about 45% less likely than GPT-4o's to contain a factual error. With thinking on, it put the figure at about 80% less than o3 (e4). The same article called hallucination a major problem (e4). The reported reductions do not show that the problem stopped being material, so the forecast is supported.
Limitations
- 'Haunt' has no threshold. The verdict rests on the plain reading that the problem stayed material and recognized.
- The GPT-5 figures are OpenAI's own and are relative, not absolute rates.
- No independent cross-lab measure of hallucination rates at the end of 2025 was cited.
- The Deloitte case is one reported incident. It does not show how often such incidents happen.
Evidence (4 items)
Supports the claim: OpenAI,
More research is needed to understand the cause of these results.
Why it matters here: OpenAI's April 2025 frontier reasoning models hallucinated on 33% (o3) and 48% (o4-mini) of PersonQA questions, more than o1. OpenAI said it did not know why.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Supports the claim: OpenAI,
Such “hallucinations” persist even in state-of-the-art systems and undermine trust.
Why it matters here: In September 2025, OpenAI's own researchers wrote that hallucinations persist even in state-of-the-art systems and undermine trust.
- Where
- Abstract
- Publisher
- OpenAI
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: The Register,
Deloitte has agreed to refund part of an Australian government contract after admitting it used generative AI to produce a report riddled with fake citations, phantom footnotes, and even a made-up quote from a Federal Court judgment.
Why it matters here: A late-2025 real-world incident in which fabricated citations from a GPT-4o-based tool led to a partial refund on a government contract.
- Where
- paragraph 1
- Publisher
- The Register
- Kind of source
- News report
- Describes events of
- Public from
Goes against the claim: The Register,
GPT-5's responses are around 45 percent less likely to contain a factual error than GPT-4o and when thinking GPT-5's responses are around 80 percent less likely to contain a factual error than OpenAI o3
Why it matters here: OpenAI reported large relative cuts in factual errors with GPT-5. This shows real progress. It does not show that the problem went away, and the same article calls hallucination 'a major problem'.
- Publisher
- The Register
- Kind of source
- News report
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, 25 AI Predictions for 2025, from Marcus on AI
Hallucinations (which should really be called confabulations) will continue to haunt generative AI.
- Where
- paragraph 30
- Published
- 2025-01-01
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-25-ai-predictions-for-2025-from-marcus-64d30d-a3: Current. Outcome: Supported. 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.
About assessment marcus-25-ai-predictions-for-2025-from-marcus-64d30d-a3.