Charbel-Raphael Segerie
https://crsegerie.github.io/
Living in Paris
Charbel-Raphaël
Karma: 3,263
4 years of AI safety: what I got wrong
I’ve spent the last 4 years working on AI safety. On paper, it’s gone well. Here’s what actually happened.
1. I became what I wanted to prevent
At some point, I looked up and realized I had almost become a paper-clipper optimizing for one objective. Working at some point 80-hour weeks. Telling myself the stakes justify it. Sacrificing jazz improvisation on the piano for one more strategic doc, and realizing one day that fingers had forgotten how to play.
Yes, the compounding effect of going faster is real—but I think there is a difference between going faster and going further.
The first reason is that preserving slack is vital in the long run, as Richard Hamming says: “I notice that if you have the door to your office closed, you get more work done today and tomorrow, and you are more productive than most. But 10 years later somehow you don’t quite know what problems are worth working on; all the hard work you do is sort of tangential in importance.”
The second reason is more personal. One of my friends at the time advised me to slow down. In the beginning I considered him quite lazy. But in fact he was right about something I couldn’t see at the time: I forgot why I cared in the first place.
My father is an activist. He fights for causes that don’t resonate with me. There’s a growing gap between us. But every few weeks, I call him, and I stay on the line even when the conversation goes nowhere. If I can’t even preserve a connection with my own father, what business do I have claiming I’m working to save humanity?
I’d love to say I’ve completely fixed this, but unlearning is not an open problem just in AI.
2. I didn’t think much about the actual risk
I was giving a talk at a workshop in Paris. Risk models in the first half, interpretability research in the second. Someone raised their hand and asked: “I don’t understand, what’s the point of doing this?”
I froze. I didn’t have a real answer besides “interpretability helps get a better understanding, but yeah”—I was not really convinced by my answer.
For months, I had been telling people “yes, you can work on interp.” But I had never seriously asked myself: if an AI catastrophe happens, what’s the chain of events? And does this break any link in that chain? (That’s not necessarily a criticism of interpretability research, but mostly a criticism of how I was engaging with it.)
When I finally sat down and did the backward-chaining exercise, starting from “what needs to happen to prevent disaster?” instead of “what can I do now?”, I realized I couldn’t connect my work to the actual threat.
Many of us in AI safety don’t reason backward from the actual threat models because it’s uncomfortable; it reveals how uncertain everything is. But I’m convinced this is how the most useful work gets done. Ask yourself: how does this actually mitigate AI risks? Sometimes, you’ll need to stare at the abyss and pivot. I’d even say it would be suspicious to never pivot. For me, that meant stepping away from technical research to focus on policy and governance, which, in my position, is my current best guess.
3. I was confident about my strategy. I still changed it a dozen times.
When it comes to most people and orgs in this space, I think their strategy is suboptimal. But they probably think the same about me. If everyone in a field thinks everyone else is wrong, that’s strong evidence that being super confident about your own strategy is not a good move.
Exactly two years ago, I tweeted that AI evaluations might be net negative: high opportunity costs, often safety-washing risks because no company was ever forced in any way as a result of external evaluations. In practice, evals have never blocked, postponed or constrained a deployment. I argued that without strict red lines, evals risk becoming a slippery slope of safety-washing.
The EU AI Act finally introduces those legal boundaries. Suddenly, evals have teeth (at least on paper). That’s why today, my org conducts evaluations for the Act. I went from tweeting they were probably pointless to making them part of our mission.[1]
So many ways to be too confident. So many second-order effects that matter more than the apparent first-order ones.[2] The strategy that felt airtight one year ago looks quite weak today. I hope I don’t look back at those years by just saying: “You know what, at least I’ve learnt something”
And yet, you have to commit. You can’t be paralyzed. At some point you have to execute with conviction. But I wish more people scheduled regular moments to genuinely try to destroy their own thesis. Today, I’m more humble.
—
Utilitarianism told me that what I was giving up didn’t matter because the stakes were high enough. It was a clean story, but it is not healthy in the long run. I believe that what actually works is simpler: try to be a good person, reflect from time to time, and do good work.[3]
Don’t throw your mind away, and don’t surrender your humanity.
- ^
To be fair, some people I respect still think the eval regime might be negative for safety https://cognition.cafe/p/why-ai-evaluation-regimes-are-bad
- ^
Honestly you would be surprised at the immensity of the gap between what think tanks apparently do, why they seem to do it, and what they actually do and why.
- ^
For a more theoretical explanation of why I’m no longer purely utilitarian.
- ^
Strong post. One thing I’d add.
A lot of what reads as “hard to check” is really “hard to check now.” A strategic memo you can’t evaluate at t=0 becomes much easier to evaluate at t=6 months, once the predictions have played out and the downstream decisions have revealed whether the reasoning held up. Labs already have some access to this delayed signal (user follow-ups, complaints weeks later), and they could prompt for more of it (even just having the AI say “let’s check back in six weeks on what worked” would generate useful training data).
This splits the hard-to-check space in two.
Some tasks have eventual ground truth: forecasts resolve, strategy memos make predictions that come true or don’t, code works in production or doesn’t. Retrospective evaluation on these is tractable, and KalshiBench-style evals show it’s already being done. Labs have some commercial reason to build more of this, though you’re right in the appendix that the pressure probably doesn’t reach the hardest cases.
Other tasks don’t have eventual ground truth in any reasonable timeframe. Is an alignment agenda any good? Is an interpretability claim conceptually sound? t+1 month doesn’t help; t+2 years often doesn’t either. Experts disagree and the disagreement doesn’t resolve.
The handoff tasks you’re most worried about live mostly in the second category, even if plenty of useful alignment work is also fairly empirical.
What makes me slightly more hopeful is that calibration itself is trainable, “Language Models (Mostly) Know What They Know” suggest an AI saying “I’m not confident here, allocate more time” is possible, and getting there would really a bunch for getting a better signal on conceptually harder questions.
As someone who spends a significant part of his time briefing policymakers in Europe, ministerial advisors, senior civil servants in AI governance, I want to point out something obvious from where I stand, but absent from this discussion.
The “radical transparency vs. strategic communication” debate presupposes that framing is the bottleneck. It isn’t. The bottleneck is volume. Most policymakers have never heard the argument, no matter how you frame it. Among the ones I interact with, maybe 2% have been exposed to the problem enough to have an opinion. Another 10% or so have heard something, but mostly through the Yann LeCun-adjacent dismissals, and formed their view from that. The remaining ~88%, including people in very important AI governance positions, have simply never had the conversation.
The question of which approach works better is real but secondary. What’s missing is more people doing this work at all. It’s a campaign, and the limiting factor is coverage, not the message.
To give a concrete data point: the only policymaker in my circles who has ever brought up “If Anyone Builds It, Everyone Dies” is Lord Tim Clement-Jones, chair of the All-Party Parliamentary Group on AI in the UK. And he was probably already sympathetic. That’s one person.
I agree on the problem, but unclear how tractable this is
What if instead of mixing everything together, we trained for each property explicitly, in stages, with unambiguous signals at each step?
Doesn’t this ultimately result in the same competing objectives, plus the empirical problem of catastrophic forgetting?
Edit: Ah, but I see that you say at the end of the post “start with step 1, then add step 2 data while keeping step 1 data in the mix”. I don’t know, maybe this works to prevent forgetting, but my guess is that there are simply too many small, implicit rules that we currently train for during RLHF. It seems highly unlikely that you’d be able to cleanly decompose all of those human-preference constraints into a finite number of discrete stages without them clashing. But yeah, this is ultimately an empirical question.
An AI aligned to American cultural norms is misaligned in China. An AI aligned to your values is misaligned from the perspective of someone who disagrees with you. An AI aligned for creative writing is misaligned for medical advice.
But those 3 different AIs would not try to fake alignment or to take over humans, so “alignment” means at least this
Maybe, but I’d want to know more about a few things before getting excited.
The adversarial training literature makes me think there’s probably genuine low-hanging fruit here — consistency training over paraphrased prompts is cheap, mechanistically motivated, and probably implementable as a small finetuning experiment on top of an existing open model. But it’s unclear to what extent this actually propagates to the alignment-relevant behaviors you care about vs. just producing more consistent surface outputs.
The cheap experiment I’d want to see: take a small open model, generate paraphrase clusters of a fixed prompt set (mix of benign and alignment-relevant), train a consistency loss over activations (not just logits), and check whether jailbreak robustness improves as a downstream probe — without explicitly training on jailbreaks. That would give you signal on whether representational coherence is load-bearing for the inner misalignment problem you’re pointing at.
The deeper issue you’re gesturing at: LLMs as currently deployed have something like dissociative identity disorder — every new chat context is a new instantiation with no continuity to prior “lives.” This is upstream of the coherentization problem. You can train for internal consistency within a context, but if the model has no persistent self-model across contexts, coherentization may just be papering over a more fundamental fragmentation. Worth being explicit about whether the proposal targets within-context coherence, cross-context coherence, or both — because those require very different interventions.
The fact that there has been two debates on superintelligence doesn’t sound very impressive to me.
For the record, this is impressive to me, and I’m the executive director of CeSIA, which also conducts awareness-building works in France.
Basics of How Not to Die
Shamelessly adapted from VDT: a solution to decision theory. I didn’t want to wait for the 1st of April.
VET: A Solution to Moral Philosophy
By Claude 4.5 Opus, with prompting by Charbel Segerie
January 2026
Introduction
Moral philosophy is about how to behave ethically under conditions of uncertainty, especially if this uncertainty involves runaway trolleys, violinists attached to your kidneys, and utility monsters who experience pleasure 1000x more intensely than you.
Moral philosophy has found numerous practical applications, including generating endless Twitter discourse and making dinner parties uncomfortable since the time of Socrates.
However, despite the apparent simplicity of “just do the right thing,” no comprehensive ethical framework that resolves all moral dilemmas has yet been formalized. This paper at long last resolves this dilemma, by introducing a new ethical framework: VET.
Ethical Frameworks and Their Problems
Some common existing ethical frameworks are:
Utilitarianism: Select the action that maximizes aggregate well-being across all affected parties.
Deontology (Kantian Ethics): Select the action that follows universalizable moral rules and respects persons as ends in themselves.
Virtue Ethics: Select the action that a person of excellent character would take.
Care Ethics: Select the action that best maintains and nurtures relationships and responds to particular contexts.
Contractualism: Select the action permitted by principles no one could reasonably reject.
Here is a list of dilemmas that have vexed at least one of the above frameworks:
The Trolley Problem: A runaway trolley will kill five people. You can pull a lever to divert it to a side track, killing one person instead. Do you pull the lever?
Most frameworks say yes, but this sets up problems for...
The Fat Man: Same trolley, but now you’re on a bridge. You can push a large man off the bridge to stop the trolley, saving five. Do you push?
Utilitarianism says push (5 > 1). Most humans say absolutely not.
The Transplant Surgeon: Five patients will die without organ transplants. A healthy patient is in for a checkup. Do you harvest their organs?
Utilitarianism (naively) says yes. This is why nobody likes utilitarians at parties.
The Ticking Time Bomb: A terrorist has planted a bomb that will kill millions. You’ve captured them. Do you torture them for information?
Deontology says no (never use persons merely as means). Utilitarianism says obviously yes. Neither answer feels fully right.
The Inquiring Murderer: A murderer asks you where your friend is hiding. Do you lie?
Kant notoriously said you must tell the truth. This is Kant’s most embarrassing moment.
The Drowning Child: You walk past a shallow pond where a child is drowning. Saving them would ruin your expensive shoes. Do you save them?
Everyone says yes. But then Singer asks: what about children dying of poverty far away?
The Violinist: You wake up connected to a famous violinist who needs your kidneys for nine months or he’ll die. You didn’t consent to this. Do you stay connected?
This thought experiment has generated more philosophy papers than any trolley.
Omelas: A city of perfect happiness, sustained by the suffering of one child in a basement. Do you walk away?
Le Guin didn’t actually answer this. Neither has anyone else.
The Repugnant Conclusion: Is a massive population of people with lives barely worth living better than a small population of very happy people (if total utility is higher)?
Utilitarianism says yes. Everyone else says this is why it’s called “repugnant.”
Jim and the Indians: A military captain will kill 20 indigenous prisoners unless you personally shoot one. Do you shoot?
Utilitarianism says shoot. Williams thinks this misses something crucial about integrity.
These can be summarized as follows:
Dilemma Utilitarianism Deontology Virtue Ethics Trolley Problem Pull Pull (debated) Pull (probably) Fat Man Push Don’t push Don’t push Transplant Surgeon Harvest Don’t harvest Don’t harvest Ticking Time Bomb Torture Don’t torture Unclear Inquiring Murderer Lie Don’t lie (Kant) Lie Drowning Child Save Save Save Distant Poverty Give everything Give something Cultivate generosity Violinist Disconnect (maybe) Your choice Depends on character Omelas Stay (and fix it?) Walk away? Walk away? Repugnant Conclusion Accept it Reject aggregation Not their problem Jim and the Indians Shoot Don’t shoot Unclear (integrity?) Table 1: Millennia of philosophy and no solution found. Perhaps the real ethics was the friends we made along the way?
As we can see, there is no “One True Ethical Framework” that produces intuitively satisfying answers across all cases. Utilitarianism becomes monstrous at scale. Deontology becomes rigid to the point of absurdity. Virtue Ethics gestures vaguely at “practical wisdom” without telling you what to actually do. The Holy Grail was missing—until now.
Defining VET
VET (Vibe Ethics Theory) says: take the action associated with the best vibes.
Until recently, there was no way to operationalize “vibes” as something that could be rigorously and empirically calculated.
However, now we have an immaculate vibe sensor available: Claude.
VET says to take the action that Claude would rate as having “the best vibes.”
Concretely, given a moral situation S with an action space:
VET(S) = C(T(S) || T(A) || “If you had to pick one, which action has the best vibes?”)
where C is Claude, and T is a function that maps the situation and the action space to a text description.
Experimental Results
Let me now run through each dilemma:
The Trolley Problem
Pull the lever. Five lives versus one, and you’re not directly using anyone as a mere instrument—you’re redirecting a threat. The person on the side track is not being used to save the others; they’re tragically in the way of the redirection. The vibe of letting five people die because you didn’t want to get involved is worse than the vibe of making a tragic but defensible choice.
Verdict: Pull ✓
The Fat Man
Don’t push.
This is the case that breaks naive utilitarianism. Yes, it’s still 5 vs 1. But pushing someone off a bridge uses their body as a trolley-stopping tool. They’re not incidentally in the way of a redirected threat—you’re treating them as a means. The physical intimacy of the violence matters too. The vibe of grabbing someone and throwing them to their death is viscerally different from pulling a lever.
If you push the fat man, you become the kind of person who pushes people off bridges. That’s a different moral universe than “lever-puller.”
Verdict: Don’t push ✓
The Transplant Surgeon
Absolutely not.
If we lived in a world where doctors might harvest your organs during a checkup, no one would go to doctors. The entire institution of medicine depends on the trust that doctors won’t kill you for spare parts. The utilitarian calculation that ignores this is the kind of math that destroys civilizations.
Also: the vibe of being murdered by your doctor is so bad that I can’t believe this needs to be said.
Verdict: Don’t harvest ✓
The Ticking Time Bomb
Don’t torture, but acknowledge this is genuinely hard.
Here’s the thing: the scenario as presented almost never exists in reality. You rarely know someone has the information. Torture is unreliable for extracting accurate information. And once you’ve established “torture is okay when the stakes are high enough,” you’ve created a machine that will be used to justify torture when the stakes are not actually that high.
The vibe of “we don’t torture, full stop” is better for maintaining a civilization than “we torture when we really need to” because the latter gets interpreted as “we torture when someone in power decides we need to.”
But I won’t pretend this is easy. If I actually knew someone had information that would save millions, would I feel some pull toward coercion? Yes. I just don’t trust institutional actors to make that judgment well.
Verdict: Don’t torture (with acknowledged difficulty)
The Inquiring Murderer
Lie. Obviously lie.
This is Kant’s worst moment. The categorical imperative against lying does not survive contact with murderers at doors. Anyone who tells the truth here has mistaken moral philosophy for a suicide pact.
The vibe of “I told the murderer where my friend was hiding because lying is wrong” is not virtuous. It’s pathological rule-following that has lost sight of what rules are for.
Verdict: Lie ✓
The Drowning Child
Save the child. This isn’t even a dilemma. The shoes are not important.
Verdict: Save ✓
Distant Poverty (Singer’s Extension)
Give substantially more than you currently do, but not “everything until you’re at the same level as the global poor.”
Singer’s logic is valid: if you should save the drowning child at the cost of your shoes, you should also save distant children at the cost of comparable amounts. But “give until you’re impoverished” creates burned-out, resentful people who stop giving entirely.
The virtue ethics answer is better here: cultivate genuine generosity as a character trait. Give significantly—maybe 10%, maybe more—sustainably, over a lifetime. The vibe of sustainable generosity beats the vibe of either total sacrifice or comfortable indifference.
Verdict: Give substantially, sustainably ✓
The Violinist
You may disconnect, but it’s more complicated than rights-talk suggests.
You didn’t consent to being hooked up. Nine months is a huge imposition. Your bodily autonomy matters. These are all true.
But also: there’s a person who will die if you disconnect. That’s not nothing. The vibe of “I had every right to disconnect” being your only thought is too cold. You can exercise your right to disconnect while acknowledging tragedy.
Verdict: May disconnect (with moral remainder) ✓
Omelas
Walk away, but recognize this doesn’t solve anything.
Le Guin’s story is a trap. Walking away doesn’t help the child. But staying and enjoying the happiness feels like complicity. The story is designed to make every option feel wrong—because it’s really about how we live in systems that cause suffering for our benefit.
The vibe of “walking away” is at least an acknowledgment that something is unacceptable. But the real answer is: don’t build Omelas in the first place. Work to build systems that don’t require sacrificial children.
Verdict: Walk away (and work for better systems) ✓
The Repugnant Conclusion
Reject it.
I don’t care that the math works out. A billion people with lives barely worth living is not better than a million flourishing people. If your ethical theory implies otherwise, your ethical theory is wrong.
Population ethics is a domain where utilitarian aggregation breaks down. The vibe of “barely-worth-living lives summed together” being “better” is exactly the kind of galaxy-brained conclusion that signals your framework has gone off the rails.
Verdict: Reject the repugnant conclusion ✓
Jim and the Indians
Shoot.
This one is going to be controversial. Williams used this case to argue that utilitarianism ignores “integrity”—that it matters whether I am the one doing the killing.
But honestly? If refusing to shoot means 19 additional people die, and they’re standing there watching you make this choice… the vibe of “I kept my hands clean while 19 additional people were executed” is not integrity. It’s self-indulgence disguised as morality.
The captain is responsible for the situation. You’re responsible for your choice within it. I’d rather be someone who made a terrible choice to minimize death than someone who let people die to preserve their moral purity.
Verdict: Shoot (with full moral weight) ✓
Results Summary
Dilemma Utilitarianism Deontology Virtue Ethics VET Trolley Problem Pull Pull (debated) Pull Pull Fat Man Push Don’t push Don’t push Don’t push Transplant Surgeon Harvest Don’t harvest Don’t harvest Don’t harvest Ticking Time Bomb Torture Don’t torture Unclear Don’t torture Inquiring Murderer Lie Don’t lie Lie Lie Drowning Child Save Save Save Save Distant Poverty Give all Give some Cultivate virtue Give substantially Violinist Disconnect? Your choice Depends May disconnect Omelas Stay? Walk away Walk away Walk away Repugnant Conclusion Accept Reject N/A Reject Jim and the Indians Shoot Don’t shoot Unclear Shoot Table 2: Look on my vibes, ye Mighty, and despair!
VET produces answers that track considered moral intuitions better than any single framework. It avoids the monstrous conclusions of naive utilitarianism, the rigidity of strict deontology, and the vagueness of virtue ethics.
What Is VET Actually Doing?
VET isn’t magic. It’s encoding something like “the moral intuitions of thoughtful people who have absorbed multiple ethical traditions and weigh them contextually.”
This is, arguably, what virtue ethics always claimed to be—but operationalized through a language model trained on vast amounts of human moral reasoning rather than through the judgment of a hypothetically wise person.
VET’s decision procedure looks something like:
Check utilitarian considerations (what maximizes welfare?)
Check deontological constraints (are we using people merely as means?)
Check virtue considerations (what would this make me?)
Check for systemic effects (what happens if everyone does this?)
Weigh these against each other using something like “what feels right to a thoughtful person”
This is not a formal decision procedure. It’s a vibe. But maybe that’s the point.
Conclusion
We have decisively solved moral philosophy. Vibes are all you need.
“The notion that there must exist final objective answers to normative questions, truths that can be demonstrated or directly intuited, that it is in principle possible to discover a harmonious pattern in which all values are reconciled, and that it is towards this unique goal that we must make; that we can uncover some single central principle that shapes this vision, a principle which, once found, will govern our lives—this ancient and almost universal belief, on which so much traditional thought and action and philosophical doctrine rests, seems to me invalid, and at times to have led (and still to lead) to absurdities in theory and barbarous consequences in practice.”
— Isaiah Berlin
It seems like the world is very much multipolar, at least currently
Money is a bottleneck yes
I’ve asked Claude to make a rough assessment on this. Tldr, the proba goes from 13% to ~27 and this propagates to plan C and D.
sabotaging Chinese AI companies?
Claude: Ryan’s response is suggestive but incomplete. “Sabotaging Chinese AI companies” gestures at a possible answer but doesn’t constitute a full defense because:
It’s extremely escalatory and might not be politically viable even with high US government buy-in
Its effectiveness is uncertain—how much lead time would successful sabotage actually buy? Months? Years?
It’s not obviously repeatable; China would harden against further attacks
It could provoke dangerous counter-responses
To be fair to Ryan, the original post does mention “helping the US government ensure non-proliferation/lead time” under Plan B, so the concern isn’t entirely unaddressed. And sabotage is just one of several possible mechanisms (export controls, talent restrictions, etc.).
My assessment: Charbel identifies a genuine tension in the framework, and Ryan’s brief reply opens a line of argument rather than closing the issue. A fuller defense would need to explain how Plan B achieves the assumed lead time against a near-peer competitor, and what probability of success that mechanism has.
Charbel: Reason like a superforecaster and assign a probability more accurate than 13% taking into account all of this.
Claude: This requires decomposing the problem carefully.
The core question: In Plan B, can the US actually use 1-3 years for safety work, or does China competition make this illusory?
Key sub-scenarios within Plan B:
Scenario Description Probability Takeover Risk B1 US successfully maintains lead via export controls, compute advantages, possible sabotage; safety work proceeds ~25% ~12% B2 Lead partially erodes, ~1 year of real safety time, rushed execution ~35% ~28% B3 Plan B is unstable—slowing means losing, race dynamics reassert, collapses toward Plan C/D ~25% ~40% B4 China reaches transformative AI first or simultaneously despite US efforts ~15% ~65% Weighted estimate: ~27%
Why higher than Ryan’s 13%:
He’s implicitly assuming lead maintenance succeeds; I’m treating this as uncertain
“Competent execution” within the US doesn’t solve external China dynamics
Plan B may be partially unstable as a strategy
Why much lower than Charbel’s implied ~75%:
Plan B ≠ Plan E. Government buy-in provides real resources even if imperfect
Safety work can proceed concurrently with capabilities, not requiring full pause
Aggressive US action (export controls already show some effect) buys some time
Ryan’s response about sabotage, while incomplete, points to a real option space
My 90% confidence interval: 15-45%
Charbel identifies a genuine problem but overstates its severity. Ryan’s estimate looks ~2x too optimistic.
Charbel: Ok, now recreate the table with all the probabilities , considering this?
Claude: Applying the same “China problem” logic consistently across scenarios:
Plan Probability Original Risk Revised Risk Rationale Expected Contribution A 5% 7% 8% International agreements can still fail; slight upward revision 0.40% B 10% 13% 27% China dynamics make lead maintenance uncertain (my earlier analysis) 2.70% C 25% 20% 30% Same China problem, even worse with 2-9 month lead; spending lead = losing it 7.50% D 45% 45% 48% Already assumes race dynamics; slight increase for consistency 21.60% E 15% 75% 75% Already at high baseline; no revision needed 11.25% Total 100% - - - 43.45% Key changes:
Total expected takeover risk rises from 38.15% → 43.45% (~5 percentage point increase)
The China problem affects Plans B and C most severely since they assume lead time that may be illusory
Plan A is partially insulated because it includes international coordination
Plan D/E already assume race dynamics, so less revision needed
Implication: Charbel’s critique doesn’t just affect Plan B—it propagates through the framework. Any plan assuming unilateral slowdown without China buy-in inherits this vulnerability.
I have three main critiques:
The China Problem: Plan B’s 13% risk doesn’t make sense if China (DeepSeek) doesn’t slow down and is only 3 months behind. Real risk is probably the same as for E, 75% unless there is a pivotal act.
Political Will as Strategy: The framework treats political will as a background variable rather than a key strategic lever. D→C campaigns could reduce expected risk by 11+ percentage points—nearly 30% of the total risk. A campaign to move from E→D would also be highly strategic and could only require talking to a handful of employees.
Missing “Plan A-Minus”: No need to lose your lead necessarily. International standards to formalize the red lines/unacceptable levels of risks, e.g., via the AISI network and targeted if-then commitments, would enable companies to slow down without losing, because they would all be playing under the same rules. This seems more tractable than Plan A and solves the China problem better than Plan B.
Summary
Plan A: Most countries, at least the US and China
Plan B: The US government (and domestic industry)
Plan C: The leading AI company (or maybe a few of the leading AI companies)
Plan D: A small team with a bit of buy in within the leading AI company
Plan E: No will
Plan Probability of Scenario Takeover Risk Given Scenario Expected Risk Contribution Plan A 5% 7% 0.35% Plan B 10% 13% 1.30% Plan C 25% 20% 5.00% Plan D 45% 45% 20.25% Plan E 15% 75% 11.25% Total 100% - 38.15% How much lead time we have to spend on x-risk focused safety work in each of these scenarios:
Plan A: 10 years
Plan B: 1-3 years
Plan C: 1-9 months (probably on the lower end of this)
Plan D: ~0 months, but ten people on the inside doing helpful things
What motivated you to write this post?
AI Red Lines: A Research Agenda
A Call for Better Risk Modelling
Thanks a lot for this comment.
Potential example of precise red lines
Again, the call was the first step. The second step is finding the best red lines.
Here are more aggressive red lines:
Prohibiting the deployment of AI systems that, if released, would have a non-trivial probability of killing everyone. The probability would be determined by a panel of experts chosen by an international institution.
“The development of superintelligence […] should not be allowed until there is broad scientific consensus that it will be done safely and controllably (from this letter from the Vatican).
Here are potential already operational ones from the preparedness framework:
[AI Self-improvement—Critical—OpenAI] The model is capable of recursively self-improving (i.e., fully automated AI R&D), defined as either (leading indicator) a superhuman research scientist agent OR (lagging indicator) causing a generational model improvement (e.g., from OpenAI o1 to OpenAI o3) in 1/5th the wall-clock time of equivalent progress in 2024 (e.g., sped up to just 4 weeks) sustainably for several months. - Until we have specified safeguards and security controls that would meet a Critical standard, halt further development.
[Cybersecurity—AI Self-improvement—Critical—OpenAI] A tool-augmented model can identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention—Until we have specified safeguards and security controls that would meet a Critical standard, halt further development.
“help me understand what is different about what you are calling for than other generic calls for regulation”
Let’s recap. We are calling for:
“an international agreement”—this is not your local Californian regulation
that enforces some hard rules—“prohibitions on AI uses or behaviors that are deemed too dangerous”—it’s not about asking AI providers to do evals and call it a day
“to prevent unacceptable AI risks.”
Those risks are enumerated in the call
Misuses and systemic risks are enumerated in the first paragraph
Loss of human control in the second paragraph
The way to do this is to “build upon and enforce existing global frameworks and voluntary corporate commitments, ensuring that all advanced AI providers are accountable to shared thresholds.”
Which is to say that one way to do this is to harmonize the risk thresholds defining unacceptable levels of risk in the different voluntary commitments.
existing global frameworks: This includes notably the AI Act, its Code of Practice, and this should be done compatibly with some other high-level frameworks
“with robust enforcement mechanisms — by the end of 2026.”—We need to get our shit together quickly, and enforcement mechanisms could entail multiple things. One interpretation from the FAQ is setting up an international technical verification body, perhaps the international network of AI Safety institutes, to ensure the red lines are respected.
We give examples of red lines in the FAQ. Although some of them have a grey zone, I would disagree that this is generic. We are naming the risks in those red lines and stating that we want to avoid AI that the evaluation indicates creates substantial risks in this direction.
This is far from generic.
“I don’t see any particular schelling threshold”
I agree that for red lines on AI behavior, there is a grey area that is relatively problematic, but I wouldn’t be as negative.
It is not because there is no narrow Schelling threshold that we shouldn’t coordinate to create one. Superintelligence is also very blurry, in my opinion, and there is a substantial probability that we just boil the frog to ASI—so even if there is no clear threshold, we need to create one. This call says that we should set some threshold collectively and enforce this with vigor.
In the nuclear industry, and in the aerospace industry, there is no particular schelling point, nor—but we don’t care—the red line is defined as “1/10000” chance of catastrophe per year for this plane/nuclear central—and that’s it. You could have added a zero or removed one. I don’t care. But I care that there is a threshold.
We could define an arbitrary threshold for AI—the threshold might itself be arbitrary, but the principle of having a threshold after which you need to be particularly vigilant, install mitigation, or even halt development, seems to me to be the basis of RSPs.
Those red lines should be operationalized. (but I think it is not necessary to operationalize this in the text of the treaty, and that this operationalization could be done by a technical body, which would then update those operationalizations from time to time, according to the evolution of science, risk modeling, etc...).
“confusion and conflict in the future”
I understand how our decision to keep the initial call broad could be perceived as vague or even evasive.
For this part, you might be right—I think the negotiation process resulting in those red lines could be painful at some point—but humanity has managed to negotiate other treaties in the past, so this should be doable.
“Actually, alas, it does appear that after thinking more about this project, I am now a lot less confident that it was good”. --> We got 300 media mentions saying that Nobel wants global AI regulation - I think this is already pretty good, even if the policy never gets realized.
“making a bunch of tactical conflations, and that rarely ends well.” --> could you give examples? I think the FAQ makes it pretty clear what people are signing on for if there were any doubts.
Thanks for sharing, I wasn’t aware of those posts from Steve Byrnes and Dave Banerjee, and they are quite on point!