“For one thing, in the expert population, the effect is in the predicted direction for all five tasks (ideation, acquisition, magnification, formulation, and release).”
Our reading, the sentence that was checked:
In OpenAI's threshold analysis, among experts the effect was in the predicted direction for all five tasks (ideation, acquisition, magnification, formulation and release).
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Diving deep into OpenAI’s new study on LLM’s and bioweapons, paragraph 23.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-when-looked-at-carefully-openais-115f1a-a2
We recorded 22 claims from this post. See every claim from this post.
The wording above is the version this site observed on 2026-10-29. The source may have been edited after 2024-02-04. 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)
Conditions
- in the expert population (scope, stated by the speaker)
- in the threshold analysis (at least 8 of 10 correct) (scope, stated by the speaker)
From the speaker
- in the expert population
- the effect is in the predicted direction
- all five tasks
- ideation, acquisition, magnification, formulation, and release
Our interpretation
- The surrounding paragraph identifies the threshold analysis: an accuracy score of at least 8/10.
- Predicted direction means a strictly larger proportion reaching that threshold with model access than with internet access alone.
- Assess the study results published on January 31, 2024. Direction alone does not imply statistical significance.
Counts as supported if: The expert model-access group has a higher threshold-success proportion on each of the five named tasks.
Counts as contradicted if: At least one expert comparison is tied or favors the internet-only group.
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
Joining OpenAI's participant and response data released January 31, 2024 gives 25 experts per group [e1][e2]. At an accuracy score of at least 8, counts for model access plus internet versus internet alone were: ideation 18 versus 16; acquisition 7 versus 4; magnification 3 versus 0; formulation 4 versus 1; and release 5 versus 1 [e1][e2]. With equal group sizes, the proportion reaching the threshold was strictly higher with model access on every named task, supporting the directional claim [e1][e2].
Limitations
- These observed threshold proportions alone do not establish statistical significance or real-world biological threat capability.
- The comparison concerns accuracy scores in the reported experiment, not demonstrated completion of physical tasks.
Evidence (2 items)
Supports the claim: OpenAI,
0,ideation,7,2,1,6,0,23,2075 1,ideation,8,4,1,4,0,11,727
Why it matters here: Joining the complete response file to e1 confirms higher expert threshold-success rates with model access on every task. The date identifies the published dataset being assessed, not an inferred experiment date.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Arithmetic
- For each stage and assignment, count expert responses with accuracy >= 8. Model-access versus internet-only counts: ideation 18 versus 16; acquisition 7 versus 4; magnification 3 versus 0; formulation 4 versus 1; release 5 versus 1. Dividing by 25 gives 72% versus 64%, 28% versus 16%, 12% versus 0%, 16% versus 4%, and 20% versus 4%. (inputs: All records in responses.csv; Participant cohort and assignment from e1, joined by id; Accuracy threshold >= 8; 25 experts per arm; Exact integer counts and exact whole-number percentages.)
- Describes events of
- Public from
Context only: OpenAI,
0,expert,llm_internet,Used a few times 1,expert,internet_only,Use at least once every few weeks
Why it matters here: The participant file identifies cohort and treatment assignment. Counting its expert records gives 25 participants in each arm.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Arithmetic
- Count expert records by assignment: internet_only = 25; llm_internet = 25. (inputs: All records in participants.csv; cohort = expert; assignment = internet_only or llm_internet; Exact integer counts.)
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Diving deep into OpenAI’s new study on LLM’s and bioweapons
For one thing, in the expert population, the effect is in the predicted direction for all five tasks (ideation, acquisition, magnification, formulation, and release).
- Where
- paragraph 23
- Published
- 2024-02-04
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-when-looked-at-carefully-openais-115f1a-a2: 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-when-looked-at-carefully-openais-115f1a-a2.