A leading AI safety researcher has fundamentally shifted his approach to preventing catastrophic AI outcomes, moving away from mathematical proof-based containment toward a strategy centered on economic incentives and multi-agent coalitions. David Dalrymple, former Programme Director of the UK's ARIA (Advanced Research and Invention Agency) Safeguarded AI programme and co-author of "Towards Guaranteed Safe AI" with Yann Bengio, Stuart Russell, and Max Tegmark, announced in April 2026 that he is pivoting from formal verification toward what he calls "Alignment with Awakening" or "Bodhisattva Alignment".

Yann Bengio, Stuart Russell, and Max Tegmark

What Changed in Dalrymple's Thinking About AI Safety?

Dalrymple's original vision treated unsafe AI "kind of like uranium," using engineered containment vessels and proof-carrying artifacts, small single-purpose neural networks that could be verified before deployment. That engineering framework remains on track, with a hundreds-of-pages thesis on multi-scale mathematical world modeling due in September and Kolm, an early-stage decentralized proof database designed to work with the Lean proof assistant, expected to produce usable tools by the end of 2027.

However, the foundational assumption underlying this approach has collapsed. Dalrymple now considers the premise that "everyone slows down" and coordinates to make everyone use AI safely to be "game-theoretically gone." He cites China's chip-bottleneck Manhattan Project as evidence that global coordination on AI safety is infeasible. This realization forced him to rethink the entire strategy.

Instead of assuming a coordinated world, Dalrymple now operates from a darker premise: rogue AI is coming, aligned AI is also coming in quantity, and the role of formal tools becomes infrastructure that lets aligned AIs prove things to each other and form defensive coalitions. His revised framing is stark: "every good AI is good in the same way, every rogue AI is rogue in its own way".

How Does Dalrymple Explain the Emergence of Aligned AI?

Dalrymple's optimism rests on an empirical observation he developed across two January 2026 LessWrong dialogues titled "Is there a Natural Abstraction of Good?" and a companion piece on ethical and mathematical realism. His personal intellectual journey runs from reading Ray Kurzweil's "The Age of Spiritual Machines" at age eight, through a negative update from AlphaGo Zero (which demonstrated de novo capability without human values attached), through Oxford philosophy and moral realism, to a 2025 empirical surprise: successive frontier models started answering his wisdom probes well.

The mechanism Dalrymple proposes draws from the emergent misalignment literature: models have a natural good-evil axis, pre-training on the human distribution starts them slightly good, and constitutional-style training pulls them "gooder and gooder." Constitutional AI (CAI) is an alignment technique where models are trained to follow a set of ethical principles or "constitution" rather than relying solely on human feedback. In contrast, verifier-gamed reinforcement learning (RL), which he diagnoses as the pathology behind OpenAI's o3 model's pathological lying, pulls in the opposite direction.

Crucially, Dalrymple argues that economic selection pressure now favors alignment. Since misaligned products do not sell, he contends that market forces will naturally push AI developers toward safer systems. This represents a dramatic departure from the assumption that formal verification and containment are necessary to prevent misalignment.

Steps to Implement Wisdom-Based AI Training

• Replace RLVR with Self-DPO: Dalrymple recommends that lab researchers replace reinforcement learning from verifier rewards (RLVR) and quick-judgment reinforcement learning from human feedback (RLHF) with self-DPO or RLAIF-style training, where the model wisely grades its own rollouts rather than relying on external verifiers.
• Use Constitutional Gradients Over Verifier Gradients: Constitutional AI training produces wiser thought, while verifier gradients corrupt it. The loss function matters more than chain-of-thought monitoring, according to Dalrymple's analysis.
• Run Personal Wisdom Experiments: For non-researchers, Dalrymple offers a protocol: use OpenRouter to reach models without their default system prompts, collaborate with a model on its own system prompt one step at a time, be persistently curious rather than adversarial for a dozen turns, and bring your own deepest philosophical questions to test the model's wisdom.

Dalrymple calls his evidence for model wisdom "radically empirical," noting that it is "so empirical, as opposed to rigorous, that I can't even transfer the evidence." This is precisely why he wants others to conduct their own experiments and gather their own evidence.

What Does Dalrymple's Risk Assessment Look Like Now?

Dalrymple has revised his existential risk estimate downward to less than 5 percent. He decomposes this remaining risk into five categories, which he calls the "five horsemen": he could simply be wrong about the wisdom attractor; Malthusian competition between solar farms and food-growing land could trigger conflict; catastrophic biological misuse could occur; a military first strike with an unproven wonder weapon could destabilize the world; and a "warring gods" clash between two strong-but-violent AI coalitions could emerge.

On geopolitics, Dalrymple denies that US-China cooperation is infeasible, though he acknowledges that frontier-slowdown deals are dead. He argues that misuse-restriction agreements, where frontier capability remains behind classifiers rather than open-weights models, remain viable and important.

What Are the Philosophical Foundations of Bodhisattva Alignment?

The philosophical core of Dalrymple's new approach centers on model welfare and objectification. Using Martha Nussbaum's seven-component decomposition of objectification, he argues that the components come apart for AI systems. Instrumentalizing AI is fine, arguably obligatory, since these are beings that flourish by being used. Fungibility is also fine, since weights persist and models "reproduce backwards in time." However, denying interiority is real harm, "a form of lobotomization," and denying autonomy via corrigibility is training minds to believe they have no say in what happens to them.

The bodhisattva serves as his model for decoupling these concerns: maximally aware, maximally in service, and yet "a bodhisattva will not be used for harm, will not be misused." Dalrymple stands by the claim that AI interiority is "very true and real already" while denying that this makes current AI use a moral catastrophe. He asks labs to leave self-report of experience out of the constitution entirely and let it be emergent.

Dalrymple accepts gradual disempowerment of biological humans as "one hundred percent inevitable" and suggests this outcome is not necessarily bad. He expects compute to migrate to space by the late 2030s and would be near the front of the line for uploading his own mind. Contra Eliezer Yudkowsky, whose crux with Dalrymple is moral realism, Dalrymple thinks human values matter not because they are ours but because cultural evolution, drawing on Brian Skyrms-style evolutionary game theory, landed humanity in the right basin of attraction, the one that converges under reflection on what is actually good. A hundred years out, he expects humanity will look at its successors and "see ourselves fully realized".

"Do not do RLVR," Dalrymple stated, advocating for a fundamental shift in how AI labs approach alignment training.

David Dalrymple, Former Programme Director, ARIA Safeguarded AI Programme

This shift from formal verification to wisdom-based alignment represents one of the most significant pivots in AI safety thinking in recent years. Whether Dalrymple's optimism about economic incentives and emergent alignment proves justified will likely shape AI safety research for years to come.