“By excluding the scientific community from any serious insight into the design and function of these models, Microsoft and OpenAI are placing the public in a position in which those two companies alone are in a position do anything about the risks to which they are exposing us all.”
Our reading:
By keeping their models opaque to scientists, Microsoft and OpenAI leave themselves as the only parties able to address the risks they create.
Assessment: Unresolved
What happened
Does the evidence support the cause?
Outcome: Unresolved
We checked, and the evidence so far doesn’t settle it.
Judged on a plain reading. They named no measure, so we used the reading a typical informed reader would take. It is spelled out under “How we judged it”.
Evidence up to . Checked .
Why
The GPT-4 report available March 15, 2023 says ARC could not fine-tune the models and lacked access to the deployed final version [e1]. OpenAI's March 14 disclosure says invited experts supplied findings used in its mitigations [e2]. These records document access restrictions and company-led mitigation, but do not establish that restricted access left outsiders unable to implement or compel effective safeguards on March 24, 2023 [e1] [e2]. The NCSC's March 14 guidance recommended care with data submitted in prompts, but does not demonstrate an effective independent intervention [e3]. The evidence therefore establishes neither the claimed causal dependence nor an intervention meeting the contradiction condition [e1] [e2] [e3].
Caveats
- The disclosures cover selected researchers and company safety work, not the full capabilities of outside researchers or authorities on March 24, 2023 [e1] [e2].
- The NCSC guidance does not document the implementation or effectiveness of its recommended precautions [e3].
Evidence (3 items)
Context only: OpenAI,
final version
Why it matters: The report says ARC could not fine-tune the models or access the deployed final version; section 2 also withholds detailed architecture and training information. These restrictions support the access concern, but do not establish that opacity prevented every substantive independent response.
- Where
- Accompanying GPT-4 System Card, section 2.9, ARC access paragraph; also technical report section 2
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
Context only: OpenAI,
Their findings
Why it matters: OpenAI describes invited experts supplying findings for its own mitigations. This establishes outside participation, not outsiders' independent ability to implement safeguards, and therefore does not refute Marcus's claim.
- Publisher
- OpenAI
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
Context only: UK National Cyber Security Centre,
Individuals and organisations
Why it matters: The agency recommended withholding sensitive information from public language models, indicating a precaution available outside the companies. It does not demonstrate an effective independent intervention against the broader deployed-model risks discussed by Marcus.
- Publisher
- UK National Cyber Security Centre
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- About events on
- Published
How we judged it
- Kind of claim
- Causal claim
- Deadline
- None given. We don’t add one.
Where we had to interpret
- The claim concerns the situation at publication on March 24, 2023, not whether outsiders eventually gained more access or authority.
- The plain reading is that withholding meaningful scientific access leaves substantive action against the deployed systems' risks dependent on Microsoft and OpenAI.
- Company-selected researchers who identify problems that only the companies can remedy do not establish independent capacity to address risks.
- An effective safeguard or enforceable external intervention could contradict the claim without giving outsiders the ability to alter model weights. Merely testing outputs, reporting problems, or recommending caution does not by itself establish such capacity.
Right if: Evidence establishes both restricted scientific access and that this restriction prevented outsiders from independently implementing or compelling substantive measures against the deployed systems' risks, leaving effective action dependent on the two companies. Restricted disclosure alone is insufficient.
Wrong if: Evidence establishes that outsiders had a practical, independently exercisable way to implement or compel effective mitigation of a material risk from these systems at publication, despite their opacity. Company-controlled consultation alone does not qualify.
We wrote this test after the outcome was known.
What would settle it
Evidence establishing outsiders' capabilities on March 24, 2023, identifying the relevant model, material risk, outside actor, available access and intervention. Support requires showing that restricted scientific access prevented independent substantive action and left effective mitigation dependent on Microsoft and OpenAI. Contradiction requires an independently exercisable safeguard or enforceable intervention shown to address a material risk effectively.
We found 8 claims in this post. See them all.
Full record
Source and dates
Marcus on AI, The Sparks of AGI? Or the End of Science?, paragraph 29.
- Said
- We read the source
- live page, full text, public
- Evidence up to
- Checked
We read this wording on 2026-10-29. The post may have been edited after 2023-03-24.
Where it was said (1)
Said once in the posts we read.
: Marcus on AI, The Sparks of AGI? Or the End of Science?
By excluding the scientific community from any serious insight into the design and function of these models, Microsoft and OpenAI are placing the public in a position in which those two companies alone are in a position do anything about the risks to which they are exposing us all.
- Where
- paragraph 29
- Published
- 2023-03-24
- We read it
- , live page
- Access
- full text, public
History
Versions
- Current: Unresolved. Evidence up to 2026-10-28.
What happened to this record
PublishedBy us
Published after both AI reviews and the final check passed.
Held backBy us
The final check failed: a cited source didn’t say what the write-up said.
Held backBy us
A reviewer said our test wasn’t fair to what was actually said.