Scores by kind of claim
Each kind of claim is scored on its own, out of 100: the share that held up, with the range it likely falls in. Fewer resolved claims give a wider range.
-
Factual claims
Statements about the present or the past, checked against the record: did they hold up?
81 76 87 -
Predictions
Statements about events outside the person's control, checked once they fell due: did they come true?
Insufficient record: 2 predictions resolved, 5 needed for a score.
-
Promises
Commitments to the person's or organisation's own future action, with something to deliver: was it delivered?
65 45 83
All types combined: 80 (likely 75 to 85, high confidence), from 84 claims: 78 factual claims, 6 promises. It mixes kinds of claim in different proportions for different people, so compare people kind by kind.
Demis Hassabis's claims
“Yes, we already have lots of internal checks and balances but we’re going to start publishing.”
Demis Hassabis, CEO of Google DeepMind 28 February 2024 Demis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold (Dwarkesh Podcast)
Google DeepMind will start publishing its safety checks and capability-based safeguards, having already internal checks and balances.
Resolved true The record shows it happened or was so.
The evidence is the person's or their organisation's own records.
Said for an organisation they lead or work for; it counts on both records.
What would count as true or false
No deadline given. True: Google DeepMind began publishing documents that set out its safety checks or safeguards tied to model capabilities, for example a published safety framework, dangerous-capability evaluation results or model-card safety sections. Its contents need not meet any quality bar, but the document itself must describe such checks or safeguards; merely calling something safety-related is not enough. Mostly true: it published such material but only partially, such as evaluation results without the safeguard framework, or the reverse. Mostly false: it made only general statements of principle, with no description of specific checks or safeguards. False: dated evidence shows it declined to publish, or later statements confirm that nothing of this kind was published.
Evidence
- Introducing the Frontier Safety Framework Google DeepMind · 17 May 2024 Own record
“Today, we are introducing our Frontier Safety Framework — a set of protocols for proactively identifying future AI capabilities that could cause severe harm and putting in place mechanisms to detect and mitigate them.”
- Evaluating Frontier Models for Dangerous Capabilities Google DeepMind · 21 March 2024 Own record
“We do not find evidence of strong dangerous capabilities in the models we evaluated, but we flag early warning signs.”
- Gemini 3.7 Flash Frontier Safety Framework Report Google DeepMind · August 2026 Own record
“We released our Frontier Safety Framework (FSF) (v1) in May 2024 and updated it in February 2025 (v2),”
- Gemini 4 Argon - Model Card Google DeepMind · October 2026 Own record
“Cybersecurity CCL: We cannot rule out that Gemini 4 Argon has reached this CCL, and we therefore developed a safety case for its cyberoffensive capabilities that we assess lower risk to below the cybersecurity CCL level.”
Reasoning
The claim resolves true: Google DeepMind published both parts the criteria call for. In March 2024 it published dangerous-capability evaluation results (E2). In May 2024 it published a Frontier Safety Framework of protocols for spotting capabilities that could cause severe harm and putting mitigations in place (E1). Later records show the framework was updated in February 2025 (E3), and the Gemini 4 Argon model card includes a cybersecurity CCL assessment and a safety case for the model's cyberoffensive capabilities (E4). All of these are the subject's own publications, but they are the published deliverables themselves. There was no deadline, so on_time is null.
Cite this claim
Demis Hassabis, “Google DeepMind will start publishing its safety checks and capability-based safeguards, having already internal checks and balances.” (said 28 February 2024). Verdict: resolved true. Reliability API, claim c_f052e2351f49, method 0.1.0+c876030c. https://epistacklabs.com/claims/c_f052e2351f49
Claims without a verdict
Is a verdict here wrong, or do you speak for Demis Hassabis? Reply on the claim's own page, with the form at its foot (how replies work). Verdicts are made automatically; how verdicts and scores are made.