Back
OpenAI, DeepMind and Anthropic Sound Alarm
webCredibility Rating
3/5
Good(3)Good quality. Reputable source with community review or editorial standards, but less rigorous than peer-reviewed venues.
Rating inherited from publication venue: VentureBeat
News coverage of a notable cross-organizational research paper on chain-of-thought monitoring as a safety tool; the underlying paper (by Korbak et al., July 2025) is the primary source and should be consulted for technical details.
Metadata
Importance: 72/100news articlenews
Summary
Over 40 researchers from OpenAI, Google DeepMind, Anthropic, and Meta jointly warn that the current window to monitor AI chain-of-thought reasoning in human-readable language is a fragile and potentially temporary safety opportunity. They argue that AI systems' visible reasoning traces can reveal harmful intentions before they become actions, but this transparency could disappear as AI technology advances. The paper calls for urgent work to evaluate, preserve, and improve chain-of-thought monitorability.
Key Points
- •40+ researchers from competing AI labs published a joint paper warning that the window to monitor AI chain-of-thought reasoning may close soon.
- •Current reasoning models 'think out loud' in human language, allowing detection of harmful intentions like model-written phrases 'Let's hack' or 'Let's sabotage'.
- •Chain-of-thought monitorability is described as fragile and could vanish through future technological developments in AI architecture or training.
- •The paper was endorsed by prominent figures including Geoffrey Hinton, Ilya Sutskever, and John Schulman.
- •Researchers urge the field to actively work to preserve and improve CoT monitorability as a near-term AI safety mechanism.
Cited by 3 pages
| Page | Type | Quality |
|---|---|---|
| Intervention Timing Windows | Analysis | 72.0 |
| Interpretability | Research Area | 66.0 |
| Multipolar Trap (AI Development) | Risk | 91.0 |
Cached Content Preview
HTTP 200Fetched Mar 20, 202626 KB
[Michael Nuñez](https://venturebeat.com/author/michael_nunez)
July 15, 2025
[Add to Google Preferred Source](https://www.google.com/preferences/source?q=venturebeat.com "Add to Google Preferred Source")
Scientists from [OpenAI](https://openai.com/), [Google DeepMind](https://deepmind.google/), [Anthropic](https://www.anthropic.com/) and [Meta](https://www.meta.ai/) have abandoned their fierce corporate rivalry to issue a joint warning about AI safety. More than 40 researchers across these competing companies [published a research paper](https://tomekkorbak.com/cot-monitorability-is-a-fragile-opportunity/cot_monitoring.pdf) today arguing that a brief window to monitor AI reasoning could close forever — and soon.
The unusual cooperation comes as AI systems develop new abilities to " [think out loud](https://tomekkorbak.com/cot-monitorability-is-a-fragile-opportunity/cot_monitoring.pdf)" in human language before answering questions. This creates an opportunity to peek inside their decision-making processes and catch harmful intentions before they become actions. But the researchers warn that this transparency is fragile and could vanish as AI technology advances.
The paper has drawn endorsements from some of the field's most prominent figures, including Nobel Prize laureate [Geoffrey Hinton](https://x.com/geoffreyhinton?lang=en), often called the "godfather of AI," of the [University of Toronto](https://www.cs.toronto.edu/~hinton/); [Ilya Sutskever](https://x.com/ilyasut?lang=en), co-founder of OpenAI who now leads [Safe Superintelligence Inc](https://ssi.inc/).; [Samuel Bowman](https://sleepinyourhat.github.io/) from [Anthropic](https://www.anthropic.com/); and [John Schulman](http://joschu.net/) from [Thinking Machines](https://thinkingmachines.ai/).
> Modern reasoning models think in plain English.
>
> Monitoring their thoughts could be a powerful, yet fragile, tool for overseeing future AI systems.
>
> I and researchers across many organizations think we should work to evaluate, preserve, and even improve CoT monitorability. [pic.twitter.com/MZAehi2gkn](https://t.co/MZAehi2gkn)
>
> — Bowen Baker (@bobabowen) [July 15, 2025](https://twitter.com/bobabowen/status/1945153754233180394?ref_src=twsrc%5Etfw)
"AI systems that 'think' in human language offer a unique opportunity for AI safety: We can monitor their chains of thought for the intent to misbehave," the researchers explain. But they emphasize that this monitoring capability "may be fragile" and could disappear through various technological developments.
## Models now show their work before delivering final answers
The breakthrough centers on recent advances in [AI reas
... (truncated, 26 KB total)Resource ID:
2ec3d817ef749187 | Stable ID: Y2M1ZWE1NG