Will technical monitoring make it possible to reliably detect and verify prohibited large-scale AI training runs?
The dispute is whether technical monitoring can reliably spot and verify banned large-scale AI training runs, or whether such runs can avoid clear detection. Over the past two months, Soares (The Diary Of A CEO) and Davidson (80,000 Hours Podcast) have argued that detection and verification are feasible. In the same period, Edouard Harris (Doom Debates) and Tom Milton (80,000 Hours Podcast) have argued that current and foreseeable monitoring cannot reliably distinguish prohibited training from allowed use.
Technical monitoring can reliably detect and verify prohibited large-scale AI training runs.
Davidson (80,000 Hours Podcast) claims it is not hard to tell when projects are training powerful AI systems and to declare such training illegal. Soares (The Diary Of A CEO) says computer chips can include built-in location tracking that makes large AI training runs more verifiable.
2 claims, in their words
-
“And you can build things into these computer chips to make this stuff more verifiable. You can build location tracking devices into these chips.”
-
“I also just disagree that it's going to be that hard to coordinate multiple projects to slow down. Like, it's not that hard to tell if projects are training powerful AI. It's not that hard to just tell them that's illegal. I think it's, you know, totally, totally feasible.”
Such monitoring cannot reliably detect or distinguish prohibited large-scale AI training runs.
Edouard Harris (Doom Debates) argues that current verification tools cannot confidently determine whether a given data center workload is training a model that meets particular evaluation thresholds. He says national technical means in an AI crisis will only support coarse claims, such as whether a data center is active, not what exact model is training. Milton (80,000 Hours Podcast) says external monitoring can tell if a data center is on and whether workloads look like inference or pre-training but not show fine-tuning details. Jeremie Harris (80,000 Hours Podcast) argues that satellite imaging and similar national technical means cannot distinguish permitted inference from illicit new training runs.
5 claims, in their words
-
“Those are the tools that we'll be able to use. They'll be very coarse-grained, and the kinds of statements they'll be able to verify are not this training run is, you know, running a model of like this description that's passed these evals. Rather, it's just, is the data center even on?”
-
“today, we do not have the verification technology that allows us to confidently do that. Like we cannot look at a data center and be like, that workload, you know, totally fine.”
-
“From this, I think we're unlikely to know about the exact details of how fine-tuning is done or what data were held out during training. We did or didn't hold out our dangerous biological data. But we can, for a start, say the data center is on or off. We might be able to then say something more fine-grained, like”
-
“there's not going to be an ability, for example, to say, listen, you can keep inferencing whatever models you have. In other words, you can keep serving up the models that you've already trained and built, but you're not going to train any more advanced ones. This is like this far and no further. ... simply because there's no way”
-
“Would they be able to do it? I suspect so. They'll take some efficiency hit. Whether that efficiency hit is very small or very large is still up in the air.”
Listeners of “Technical monitoring can reliably detect and verify prohibited large-scale AI training runs.” have not heard
Listeners of these shows have not heard Edouard Harris's claim that national technical means will at best confirm whether a data center is active, not what specific model is training. They have also not heard Jeremie Harris's argument that satellite imaging cannot practically separate allowed inference activity from banned new training.
2 unanswered arguments, in their words
-
“Those are the tools that we'll be able to use. They'll be very coarse-grained, and the kinds of statements they'll be able to verify are not this training run is, you know, running a model of like this description that's passed these evals. Rather, it's just, is the data center even on?”
-
“there's not going to be an ability, for example, to say, listen, you can keep inferencing whatever models you have. In other words, you can keep serving up the models that you've already trained and built, but you're not going to train any more advanced ones. This is like this far and no further. ... simply because there's no way”
Listeners of “Such monitoring cannot reliably detect or distinguish prohibited large-scale AI training runs.” have not heard
Listeners of these shows have not heard Soares's suggestion that future computer chips could include location tracking features to make large training runs more verifiable. They have also not heard Davidson's claim that it is straightforward to detect when projects are training powerful AI systems and to outlaw that activity.
2 unanswered arguments, in their words
-
“And you can build things into these computer chips to make this stuff more verifiable. You can build location tracking devices into these chips.”
-
“I also just disagree that it's going to be that hard to coordinate multiple projects to slow down. Like, it's not that hard to tell if projects are training powerful AI. It's not that hard to just tell them that's illegal. I think it's, you know, totally, totally feasible.”
Position 0 leans on claims that powerful training runs are straightforward to detect and that chips can report locations, which position 1's speakers never directly discuss. Position 1 rests on limits of coarse external monitoring and national technical means, while position 0's speakers do not address those monitoring constraints in detail.
- Nov 14, 2026 Published: cleared the gate.
- Nov 14, 2026 2 new claims (The Diary Of A CEO, 80,000 Hours Podcast).
- Nov 13, 2026 2 new claims (80,000 Hours Podcast).
- Nov 6, 2026 7 new claims (Doom Debates, The Diary Of A CEO, 80,000 Hours Podcast).
Published because: 3 shows, 2 positions, 49 days of debate. Questions publish only when claims come from at least three shows, at least two positions are held by different people, and the debate has lasted two weeks.
Pairs are matched by content across different episodes and shows; no exchange between these speakers took place unless both quotes come from the same episode.
The prose on this page is written by a model from the quotes shown, and every sentence of it rests on quotes you can open; sentences that could not be traced to a quote were removed before publishing. Positions are placed from each speaker's own words. Evidence labels say what the speaker cited, never whether it is true.
5 episodes behind this page
- 2026-11-0580,000 Hours Podcast — An AI disaster hits. The president calls Xi. Then what? | Jeremie Harris (crosspost from NonZero)
- 2026-10-2380,000 Hours Podcast — Tom Milton on how we could verify a US–China deal on AI
- 2026-10-20Doom Debates — “AI Chernobyl” Will Scare Trump and Xi To Order An Immediate Pause — Jeremie & Edouard Harris
- 2026-09-2980,000 Hours Podcast — Will AI take power — or will humans use it to take power first? With Katja Grace and Tom Davidson
- 2026-09-17The Diary Of A CEO — The Great AI Debate: Is Artificial Intelligence an Extinction Threat? Debating the True Risks of Advanced Models
The week's bottom lines, in your inbox
One email a week: the stories that mattered and what they actually mean. Free.