“Moreover, benchmarks are often easily gamed.”
Our reading, the sentence that was checked:
AI benchmarks are often easily gamed.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Rumors of AGI’s arrival have been greatly exaggerated, paragraph 20.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-rumors-of-agis-arrival-have-been-0b1c6b-a1
We recorded 18 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 2026-02-17. 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
- benchmarks
- often
- easily gamed
Our interpretation
- Plain reading: many widely used AI benchmarks and leaderboards have documented weaknesses that let scores be inflated without the matching capability. Examples are selective disclosure, test-set exposure, or graders that accept trivial or empty answers. These weaknesses are not rare, and exploiting them takes little effort.
- He describes an ongoing state with no period. Evaluator uses documentation published from 2024 up to the post date of 2026-02-17. Later audits are context only.
Counts as supported if: Pre-post audits show several prominent benchmarks or leaderboards whose scores can be inflated by simple means or were distorted by gaming practices.
Counts as contradicted if: Audits of prominent benchmarks before the post find them generally robust, with gaming rare or requiring great effort.
Retrospective: these criteria were written after the outcome was public.
Current assessment
Was it true for the period described?
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
- No horizon
- Assessment recorded
- Policy version
- 2026-10-28.7
Rationale
Audits before the post show prominent AI benchmarks could be gamed with little effort. A July 2025 audit by researchers from Stanford, Berkeley, UK AISI and others found 'many agentic benchmarks' with flawed setup or scoring. SWE-bench Verified used too few tests, and TAU-bench counted empty responses as successes. 'The Leaderboard Illusion' (April 2025) found that providers testing many private variants, such as 27 by Meta before Llama 4, could pick their best score and bias Chatbot Arena rankings. An April 2026 Berkeley audit, used as context because it came after the post, exploited eight major agent benchmarks to near-perfect scores without solving any tasks.
Limitations
- 'Often' has no threshold. The evidence covers prominent benchmarks, not all benchmarks.
- The strongest demonstration (Berkeley, April 2026) postdates the post and is context only.
Evidence (3 items)
Supports the claim: arXiv (Zhu et al., incl. Stanford, Berkeley, UK AISI researchers),
many agentic benchmarks have issues in task setup or reward design. For example, SWE-bench Verified uses insufficient test cases, while TAU-bench counts empty responses as successful.
Why it matters here: Many prominent benchmarks had flaws that allow inflated scores. One even credits empty responses, which shows gaming can be easy.
- Where
- Abstract
- Publisher
- arXiv (Zhu et al., incl. Stanford, Berkeley, UK AISI researchers)
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: arXiv (Singh et al.),
the ability of these providers to choose the best score leads to biased Arena scores due to selective disclosure of performance results. At an extreme, we identify 27 private LLM variants tested by Meta in the lead-up to the Llama-4 release.
Why it matters here: The leading public leaderboard was distorted by best-of-many private testing, a simple way to game rankings.
- Where
- Abstract
- Publisher
- arXiv (Singh et al.)
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: UC Berkeley RDI,
discovered that every single one can be exploited to achieve near-perfect scores without solving a single task. No reasoning. No capability. Just exploitation of how the score is computed.
Why it matters here: Later confirmation that major benchmarks in use before the post were easy to exploit. It is dated after the post, so it is context only.
- Publisher
- UC Berkeley RDI
- Kind of source
- Independent measurement
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Rumors of AGI’s arrival have been greatly exaggerated
Moreover, benchmarks are often easily gamed.
- Where
- paragraph 20
- Published
- 2026-02-17
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-rumors-of-agis-arrival-have-been-0b1c6b-a1: 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-rumors-of-agis-arrival-have-been-0b1c6b-a1.