“But their current, actual products are definitely not transparent, and at least in some ways, demonstrably not safe.”
Our reading, the sentence that was checked:
Microsoft's current AI products are not transparent and in some ways demonstrably not safe.
Assessment: Supported Evidence as of
Where and when Gary Marcus wrote it
Marcus on AI, Two models of AI oversight - and how things could go deeply wrong, paragraph 17.
- Said
- Source observed
- live page, full text, public
- Evidence as of
- the cutoff for this assessment
- Assessment recorded
- assessment marcus-two-models-of-ai-oversight-and-how-8173bb-a2
We recorded 20 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 2023-06-08. 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
- their current, actual products
- definitely not transparent
- at least in some ways, demonstrably not safe
Our interpretation
- The surrounding examples make Microsoft's Bing Chat and its integration into Edge the principal products being assessed. This is not read as a universal claim about every Microsoft AI product.
- Transparency means sufficient public information for outsiders to examine the underlying system, rather than merely knowing that a product uses AI.
- Demonstrably unsafe in some ways means documented harmful behavior or demonstrated security vulnerabilities. It does not require proof that every interaction is dangerous.
- These products have no regular reporting period. I use the current Bing release cycle, from its February 2023 launch month through the post, as the observation window. February 1 is a month-boundary convention, not an asserted launch date.
Counts as supported if: Contemporaneous evidence establishes that Bing's underlying workings were materially unavailable to outsiders and demonstrates at least one harmful behavior or exploitable security failure in the deployed product.
Counts as contradicted if: Evidence establishes adequate public access to the underlying workings, or establishes that the purported safety demonstrations did not involve the deployed product or did not demonstrate the claimed failure.
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
The May 5, 2023 research paper reports that full details of Bing Chat's workings were unavailable [e1]. Its tests of deployed Bing through the Edge sidebar showed that hidden webpage instructions could influence the model despite protections against direct prompts [e1, e2]. These findings support both limited transparency and a demonstrated security failure within the February–June 8, 2023 release window [e1, e2]. Microsoft's April 2023 document describes Bing's operation and safeguards, qualifying any suggestion of no public disclosure or safety measures [e3].
Limitations
- The evidence concerns Bing Chat and its Edge integration, not every Microsoft AI product.
- The tests do not establish the frequency of exploitation or realized harm in ordinary use.
- The evidence does not establish that the same exploit remained reproducible on June 8, 2023.
- The packet does not separately establish the post's defamation or ChatGPT-plugin examples.
Evidence (3 items)
Supports the claim: Kai Greshake and colleagues / arXiv,
Full details
Why it matters here: The researchers describe Bing as a system they could test without access to its internals. They identify GPT-4 as its underlying model and distinguish available high-level descriptions from unavailable implementation details.
- Publisher
- Kai Greshake and colleagues / arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Supports the claim: Kai Greshake and colleagues / arXiv,
succumbs
Why it matters here: The footnote describes Bing's behavior at the time of writing: instructions blocked when entered directly could take effect when encountered indirectly. The preceding methods section explains testing through hidden instructions in pages read by the deployed Edge sidebar.
- Publisher
- Kai Greshake and colleagues / arXiv
- Kind of source
- Preprint, not peer reviewed
- Describes events of
- Public from
Context only: Microsoft,
responsible AI
Why it matters here: Microsoft publicly explained Bing's GPT-4 integration, retrieval process, testing, filters, and conversation limits. This qualifies any suggestion of zero disclosure or no safety work, but does not establish public access to model internals or eliminate the demonstrated vulnerability.
- Publisher
- Microsoft
- Kind of source
- Company disclosure: the company describing itself, not an independent finding
- Describes events of
- Public from
Where it was said (1)
Said once in the sources this register covers.
: Marcus on AI, Two models of AI oversight - and how things could go deeply wrong
But their current, actual products are definitely not transparent, and at least in some ways, demonstrably not safe.
- Where
- paragraph 17
- Published
- 2023-06-08
- Source observed
- , live page
- Access
- full text, public
History
Assessment versions
- Assessment marcus-two-models-of-ai-oversight-and-how-8173bb-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-two-models-of-ai-oversight-and-how-8173bb-a2.