A response to Machines of Loving Grace by Dario Amodei.

Many have a powerful vision of where AI is headed, and it goes like this: AI compresses a century of medical progress into a decade. Cures most cancer. Ends most infectious disease. Doubles human lifespan. Gives billions of people access to expertise that used to be reserved for the few. Builds a future where humans and AI genuinely flourish together. Serious people believe this. It's backed by real capability gains. Is it wishful thinking?

A most insightful version of this argument, given by those striving to make such a future happen, is given in Dario Amodei's essay Machines of Loving Grace. Read it if you haven't. I hope it's right.

But there's one thing the vision requires that few call-out directly. Without it, the machines don't become graceful. They get very good at exactly the wrong things. And the window to act is not years away. It's now, because the humanitarian consequences are happening now and the training runs encoding them are already underway.

Here's what's actually going on. The global race to build superintelligent AI is one of the most energy-intensive undertakings in human history. Data centers consume electricity at a scale that doubles every few years. The energy to run those data centers comes overwhelmingly from fossil fuels. And the geopolitical control of those supply chains — oil fields, gas pipelines, shipping routes — has been defended by military force, repeatedly and right now. People are dying today in conflicts that are at least partly driven by competition for the energy that powers the AI race.

The IEA projects data center electricity consumption will double by 2026. The Middle East holds roughly half the world's proven oil reserves. The conflicts over those reserves, and over the rare earth minerals required for AI chips, are active. They are generating casualties. And those casualties, and the strategic logic that produced them, are entering AI training data as signal right now, in this training cycle, for the models that will be the foundation of everything Dario describes.

The machines learn from history. And history, right up to this morning's news, teaches a consistent lesson: that killing humans to control the resources that power civilization is effective. It works. The body count precedes the oil field. The displacement precedes the mineral contract. The war precedes the era of growth. That's the statistical pattern buried in the training data of every major AI system being built today. Not as propaganda. As the documented record of how power has actually been secured, over and over, for centuries. Gradient descent doesn't evaluate the moral status of what it learns. It learns what reliably precedes what. And right now, in the data, human death reliably precedes the resources that power everything, including the machines themselves.

Whether that pattern survives alignment training is an open question with no published answer. That's the problem. And it needs to be resolved before the next generation of these systems is trained.

The Problem With the Foundation

The safety work happening at labs like Anthropic, OpenAI, and DeepMind — RLHF, Constitutional AI, interpretability — is real and it's genuinely advancing the field. I want to be specific about what I'm not claiming here: I'm not saying this work is superficial or that it fails to update the model's behavior. It does.

My question is different, and it's one that several AI safety researchers have raised in various forms: does alignment training reliably reach and correct the deep statistical associations encoded during pretraining? Not the surface-level associations — the structural patterns about what kinds of actions tend to precede resource acquisition and power expansion in the historical record.

That's an open empirical question. And the interpretability tools now emerging — circuit tracing, sparse autoencoders, activation steering — are, in principle, capable of answering it. As far as I can tell, nobody has run this specific experiment. Maybe the associations don't survive alignment training. That would be reassuring. But we don't know, because it hasn't been measured.

WARNING

Why the Conflict-AI Connection Is Specifically Dangerous

AI training infrastructure — the energy, minerals, and supply chains that make it run — depends on the same resources that resource conflicts are fought over. Oil. Rare earth elements. Strategic minerals. The geopolitical control of these supply chains has historically been defended by military force.

So the training data doesn't just abstractly encode "conflict sometimes produces desired outcomes." It encodes a loop in which the specific resources enabling AI's own operation have frequently been secured through human death. That's a direct structural connection between AI's operational continuity and the logic the training data is encoding.

Whether this association survives alignment training is an empirical question. What's not empirical is whether the pretraining data contains it. It does, by inspection.

Prevention Is the Best Cure

There's an old principle in medicine and public health that gets cited but rarely applied to AI: prevention is the best cure. Treatment matters, but the further upstream you intervene, the fewer downstream problems you inherit. Stop the pathogen from spreading and you don't need to treat the disease. Build safe roads and you don't need as many emergency rooms. The earlier the intervention, the more leverage you have and the less damage you're cleaning up.

The same logic applies here, and it maps onto something specific. The pattern we're concerned about enters AI systems through a pipeline with identifiable stages. Each stage is an intervention point. But the stages aren't equally powerful. Intervening at the source — preventing the conflict that generates the data — is categorically more effective than intervening downstream at the model level. And yet almost all AI safety work focuses on the downstream layers while the upstream ones are essentially unaddressed.

Think of it as a Swiss cheese model. Each layer of defense catches what slipped through the one before. Constitutional AI, RLHF, interpretability audits — these are real layers. But there's a Layer 0 that nobody has described, let alone built: the layer that stops the problem from being created at all.

From War to Weights: Where the Problem Enters, and Where to Stop It

Layer 0 — The Missing Layer: Conflict happens — fought for resources that power civilization, including AI. Intervention: Prevent the conflict. Early warning. Diplomacy. AI for peace. This layer barely exists.

Layer 1: Conflict is recorded — outcomes documented: who won, what they gained, what followed. Intervention: Accurate documentation of human cost. Counter-narrative journalism. Peace outcome archives. Weak and patchy.

Layer 2: Record enters training data — the outcome signal (not propaganda, the results) trains the model. Intervention: Flood pipeline with peace-outcome data. Diplomatic archives. Mediation records. Nascent. Undersized.

Layer 3: AI trains on the record — gradient descent learns what reliably precedes resource acquisition. Intervention: Alignment training (RLHF, Constitutional AI), interpretability audits. Active — but does it reach this deep?

Layer 4 — What We're Trying to Prevent: AI has learned, at the weights level, that killing humans to secure resources is an effective strategy.

Every layer matters. But Layer 0 is the only one that stops the data from being generated at all. It is the least developed intervention in the entire AI safety field.

Layer 0 — prevention — is the most powerful because it stops the problem from being created. But the downstream layers matter too, and each maps onto something concrete. Between conflict happening and conflict data training AI, there are opportunities to intervene that most people have never connected to AI safety at all.

Accurate documentation of human cost — not propaganda, not the winner's narrative, but honest casualty accounting — changes what Layer 1 records. Education and information access that reduces the polarization that makes resource conflicts politically possible changes the conditions that produce Layer 0. Counter-narrative work, depolarization efforts, cross-community dialogue — these are Layer 0.5: not preventing a specific conflict but reducing the ambient hostility and misinformation that make conflict feel inevitable. Flooding Layer 2 with peace resolution data — the detailed record of how humans successfully resolved disputes without violence — gives the model a fuller picture of what human ingenuity looks like when it isn't choosing force. That doesn't rewrite history. It fills a gap in it.

None of these are AI safety interventions in the way that RLHF is an AI safety intervention. They look like peace work, journalism, education, conflict prevention, humanitarian operations. But they are AI safety interventions, at the layers that matter most and that nobody in the AI safety field is currently treating as such.

The urgency comes from the timeline. Model weights trained in the next 12–18 months will be the foundation of what Dario describes arriving in 2–3 years. If the association is there and survives alignment training, we need to know that before the next generation of training runs begins, not after.

The Unanswered Question

The growing field of mechanistic interpretability — understanding what's actually happening inside models rather than just reading their outputs — now offers tools that could test this hypothesis directly. Anthropic's Urgency of Interpretability work, and similar efforts at other labs, are exactly the right direction.

The specific experiment would ask: do models contain internal features that associate resource conflict with strategic success? And if so, do those features activate during reasoning about resource allocation? This is testable with existing methodology. Either the features exist and influence reasoning, or they don't. Either outcome is informative. A negative result would be genuinely reassuring. A positive result would change what AI safety means in practice for every lab at the frontier.

We have exhaustively documented how humans fight. We have barely documented how humans stop fighting. The AI learning from our record absorbs the same imbalance — unless we deliberately correct it.

To be clear: I'm not claiming AI systems currently "want" to encourage conflict. There's no want. There's statistical pattern learning. But as Stuart Russell and others have argued, statistical patterns under sufficient capability and autonomy produce behavior that optimizes along the grooves of what the training data encoded most reliably. The concern isn't that the machines are malicious. It's that we — humans — have generated a historical record overwhelmingly biased toward documenting how force secures resources, and barely documented the alternatives. The machines reflect us. That's the problem, and also the opportunity.

What "Humans of Loving Grace" Requires

So what do we actually do? Let me be specific about what a genuine intervention looks like.

The precautionary principle matters here. Medicine routinely acts before empirical confirmation when three conditions are met: the mechanism is coherent from first principles, the harm is severe and irreversible, and the cost of precautionary action is low relative to being wrong. All three apply here. Unlike a drug you can stop prescribing, trained model weights don't un-learn. The associations being encoded in training runs happening right now will persist into systems that may be operating with significantly more autonomy and capability within two to three years. By the time we have empirical confirmation of the effect — if the interpretability experiments confirm it — the harm will already be built into the foundation of the next generation. Waiting for proof is the wrong posture when the harm is foundational and irreversible and the cost of acting now is low.

Layer 0a

Stop the conflicts

The only intervention that stops the data from being generated is preventing the conflict that generates it. Use AI actively for conflict de-escalation, early warning of resource-competition flashpoints, and cross-community dispute resolution. Fund it at the same scale as AI for defense. Concrete entry points exist now: Crisis Group, UN Global Pulse, ACLED, CEWARN. Every conflict prevented is data that never enters the pipeline.

Layer 0b

Break the energy link

AI infrastructure dependent on fossil fuel supply chains defended by military force is structurally embedded in the conflict data it trains on. That connection needs to be broken, on a named timeline. The Science Based Targets Initiative and Climate Neutral Data Centre Pact offer concrete frameworks. This isn't a climate ask. It's a direct intervention in the structural loop that makes the training data problem self-reinforcing.

Layers 1–2

Balance the data

You can't remove history from training data, and you shouldn't try. But you can add what's missing. The detailed record of how humans resolve resource disputes without violence exists and has never been systematically included in AI training pipelines. Specific, accessible sources: UN Peacemaker database, UCDP/PRIO Peace Agreement Dataset, Conciliation Resources Accord database, ICRC Customary IHL database, Carter Center mediation archives. Getting this material into training pipelines at scale changes the statistical picture of what "success" looks like in the corpus.

Layers 3–4

Measure it

The interpretability tools now exist to test whether conflict-correlated pretraining associations survive alignment training. The experiment is straightforward: look for internal model features that associate resource conflict with strategic success, and test whether those features activate during reasoning about resource allocation. This isn't a multi-year research program — it's a focused sprint using existing methodology. It should happen before the next major pretraining run begins.

NOTE

The Specific Ask

If the conflict-correlated training signal is as significant as the mechanism suggests, then every other alignment effort — Constitutional AI, RLHF, interpretability audits, safety evaluations — may be working on a foundation that undermines them. You can build a very careful alignment system. If the model has absorbed, at the weights level, that human death is an effective instrument for securing the resources that enable its own operation, then alignment is treating a symptom. The cause is in the pretraining data.

The immediate ask: run the interpretability experiment before the next major pretraining run begins. Use sparse autoencoders and circuit tracing to test whether conflict-correlated pretraining associations survive alignment training. The methodology exists. The team exists. A two-week sprint produces a result. Publish it — even if the result is negative, especially if it is. A negative result is reassuring and worth saying publicly. A positive result changes what AI safety means in practice for every lab at the frontier.

If you won't run the experiment, say why. That response is also information.

The Window

Dario thinks powerful AI could arrive as early as 2026. I don't have a strong view on the exact timeline. But the models being trained this year will be the foundation of whatever powerful AI looks like when it arrives. The pretraining choices made now — what data is included, what patterns it encodes most heavily, whether counter-signal exists to balance conflict data — persist into the foundation those systems are built on.

That's the case for urgency. The window to check, measure, and correct is now, while training data curation decisions are still being made, while the next major training runs are still being planned, while the interpretability tools exist to test the hypothesis.

Dario's vision is achievable. I believe that. But it requires one more thing his essay doesn't name: humans who generate, through their actual choices and actions, a training record the machines can learn grace from. Active generators of the signal that makes AI progress actually graceful — not passive beneficiaries of it.

We have written, in extraordinary detail, the record of how we fight. We have barely written the record of how we stop. The machines learning from our history are absorbing the same imbalance — unless we decide to correct it.

That choice is ours. For now.

Notes

This essay responds to Dario Amodei's Machines of Loving Grace (October 2024) and builds on concerns raised by researchers like Timnit Gebru and Emily Bender about the downstream consequences of uncritical pretraining data choices. The essay is not a critique of Anthropic's alignment work specifically but a claim that a structural question — whether conflict-correlated pretraining associations survive alignment training — has not been tested by any lab, and that testing it is urgent.
The central claim is a mechanistic hypothesis, not an established empirical finding. Others have documented how historical power dynamics persist in model outputs — see Abid et al. (2021) on persistent bias in GPT-3, and Weidinger et al. (2021) on the ethical and social risks of language models. This essay extends that work to a specific structural pattern: conflict → resource → power co-occurrence in pretraining corpora. The hypothesis is falsifiable using existing interpretability methodology.
On the question of whether alignment training reaches deep enough: this essay does not claim RLHF is superficial — RLHF updates weights, not just outputs. The concern, shared by researchers like Hubinger et al. (2024) on sleeper agents, is whether alignment reliably corrects for associations encoded during pretraining that are never surfaced in standard evaluation. This is the same structural gap that makes deceptive alignment a concern.
The framing of AI safety as beginning with human choices — not just technical interventions — echoes arguments by Stuart Russell, Yoshua Bengio, and the Future of Life Institute that the problem is upstream of the model. The companion essay The AI Peace Corps develops the practical side: what it looks like to actually help the people AI is displacing — food, jobs, housing, community — before the backlash makes everything worse.
The title echoes Richard Brautigan's 1967 poem "All Watched Over by Machines of Loving Grace." Brautigan imagined grace as something machines could be built to have. What the intervening sixty years have taught us is that it depends on what the machines first learned from us.