“RAG may reduce hallucinations somewhat, but it certainly won’t come close to eliminating them.”
Our reading, the sentence that was checked:
RAG may somewhat reduce hallucinations but will not come close to eliminating them.
Assessment: Unresolved No deadline stated on 2026-10-29 Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, No, RAG is probably not going to rescue the current situation, paragraph 24.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-no-rag-is-probably-not-going-to-rescue-afdf2d-a2
We recorded 11 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 2024-02-23. This page does not claim the wording was captured then.
What the claim commits to
- Kind
- Forecast
- Deadline, as the speaker wrote it
- None stated. This site does not add one.
- Deadline state when assessed
- No deadline stated
- Deadline state on 2026-10-29
- No deadline stated (the release’s reference date)
From the speaker
- RAG may reduce hallucinations somewhat
- certainly won’t come close to eliminating them
Our interpretation
- The firm commitment is that RAG will not achieve near-zero hallucination rates across representative factual answering tasks.
- The surrounding discussion concerns broad reliability. A single narrow task does not establish that broad result.
- No numerical definition of close or deadline is supplied. The separate statement about perfectly reliable chatbots next year is not imported into this forecast.
- The tentative statement that RAG may help is not a separate firm prediction.
Counts as supported if: An open-ended negative prediction cannot be established merely by observing continued failures for a finite period. Decisive support would require a demonstrated general barrier covering the relevant RAG methods.
Counts as contradicted if: Reproducible evaluations demonstrate that RAG brings hallucinations near zero across representative factual tasks, with substantial answer coverage and evidence connecting the improvement to retrieval.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Did the forecast come true?
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 deadline stated
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
CRAG's June 7, 2024 report found weighted hallucination rates of 17%–25% among the four industry RAG systems tested [e1]. Vectara's February 28, 2025 leaderboard reported 0.7% hallucination with 100% answer coverage for Gemini-2.0-Flash-001 on document summarization [e2]. Its November 19, 2025 report on a harder summarization benchmark reported a leading rate of 3.3% [e3]. These evaluations establish neither a general barrier to future RAG reliability nor broadly representative near-zero factual error attributable to retrieval, so they do not settle the open-ended forecast [e1, e2, e3].
Limitations
- The claim supplies no deadline or numerical definition of near-elimination.
- The summarization benchmarks measure source faithfulness, not factual truth across representative queries, and do not isolate retrieval's contribution.
- The benchmarks use different datasets; their rates do not establish a trend for a fixed system.
- The packet cannot establish that no qualifying demonstration exists elsewhere.
Evidence (3 items)
Context only: arXiv; Yang and colleagues, Meta and HKUST,
17% to 25%
Why it matters here: CRAG found substantial hallucinations across its tested industry systems. This documents failures at that time, but cannot establish that future RAG systems will never approach elimination.
- Publisher
- arXiv; Yang and colleagues, Meta and HKUST
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Vectara,
summarizing a document.
Why it matters here: The table reports a 0.7% hallucination rate with 100% answer rate for Gemini-2.0-Flash-001. This materially challenges a blanket dismissal of low error rates, but the test measures judged summary faithfulness, not broad retrieval-based factual answering or retrieval's causal contribution.
- Publisher
- Vectara
- Kind of source
- Benchmark result
- Describes events of
- Public from
Context only: Vectara,
3.3% hallucination rate
Why it matters here: The revised summarization benchmark uses longer, harder, more varied documents and reports higher rates. It shows why low scores on the earlier dataset cannot establish general elimination.
- Where
- Leaderboard Results
- Publisher
- Vectara
- Kind of source
- Benchmark result
- Describes events of
- Public from
What would settle it
A dated, reproducible evaluation showing near-zero hallucinations across representative factual tasks, substantial answer coverage, reliable grading, and improvement attributable to retrieval would contradict the forecast. A demonstrated general barrier covering the relevant RAG methods would support it.
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, No, RAG is probably not going to rescue the current situation
RAG may reduce hallucinations somewhat, but it certainly won’t come close to eliminating them.
- Where
- paragraph 24
- Published
- 2024-02-23
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-no-rag-is-probably-not-going-to-rescue-afdf2d-a2: 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.
About assessment marcus-no-rag-is-probably-not-going-to-rescue-afdf2d-a2.