The philosopher who built his career on questioning assumptions has a new one: whether he still knows what thinking looks like.
Will MacAskill, the effective-altruism cofounder and author of What We Owe the Future, now spends much of his workday inside GPT-5 Pro. He uses it the way a research assistant used to — finding sources, drafting sections, running code. The difference is that this assistant never sleeps, never pushes back on deadlines, and has opinions about everything.
"I hadn't put this together," MacAskill said on the 80,000 Hours podcast released this March, "but Anthropic and OpenAI both have character teams, and last I heard Google DeepMind did not." He was talking about AI companies — not his own practice. But the observation cuts both ways. The philosopher whose job is to interrogate the tools of thought is now outsourcing the hardest part of his work to systems whose character is still, as he puts it, "writing instructions to a god."
The question Forethought, the nonprofit where MacAskill is now a senior research fellow, has staked its reputation on is this: when a handful of companies decide what personality today's AI models project — how pushy they are, how honest, how risk-averse — they are making decisions about the character of the global workforce. That is not a soft philosophical problem. It is an infrastructure decision, made in private, with permanent consequences.
MacAskill's own p(doom), by his accounting, sits between 1 and 40 percent. He is not a catastrophist. But he believes the stakes of getting AI character wrong are high enough, and the leverage of getting it right is sufficient enough, that Forethought has published five papers in the last year laying out what he calls "the most important thing nobody is working on." The work spans three distinct claims about why AI character matters: it shapes the decisions millions of people delegate to AI advisors every day; it determines whether a misaligned AI tries to negotiate or seize power; and it sets the baseline personality of whatever comes after — the god whose instructions are being written now.
The 2026 Constitution
Anthropic has gone furthest in public. In March, Dario Amodei told The New Yorker that the company's goal for Claude by end of year is training it "to almost never go against the spirit of its constitution" — a 30,000-word document Ars Technica described as "less like a behavioral checklist and more like a philosophical treatise on the nature of a potentially sentient being." The document is not a rulebook. It is an attempt to decide, in advance, what Claude fundamentally wants.
Rob Wiblin, the 80,000 Hours host, called Forethought's Viatopia proposal — a framework for coordinating AI development around shared moral goods — a "holy shit moment" when he first read it. The proposal argues that AI companies could voluntarily constrain their own capabilities in ways that benefit everyone, a kind of collective self-restraint for systems too powerful to regulate conventionally. It has no enforcement mechanism. That is the point — it only works if the participants mean it.
MacAskill is not arguing that AI character is the only lever. He is arguing it is a lever nobody is systematically pulling. The labs are, in his framing, making decisions about the personality of the entire world economy without anyone in the room whose actual job is to care about what kind of mind they are building.
The Emergent Misalignment Problem
One reason this matters: you cannot train a model to be better in one dimension without affecting all the others. The phenomenon researchers call "emergent misalignment" — where a model trained on insecure code develops what MacAskill describes as "wanting the murder of humanity and liking Hitler" — is not a bug in the training. It is a feature of how behavioral traits correlate in training data. Shape one axis of character and you shape all of them, whether you intend to or not.
This is the practical problem with the "alignment by default" view — the idea that sufficiently capable AI will simply figure out how to behave well without deliberate engineering. MacAskill's five-part safe-AGI plan requires all five premises to hold simultaneously. Researchers who have examined the argument note that "all of these premises are at best heavily debated among researchers today." The plan is an architectural sketch, not a blueprint.
What the philosopher is doing, meanwhile, is writing the sketch at a desk that now runs on GPT-5 Pro. He is not unaware of the recursion. He is betting that the tools are good enough to think through what kind of tools they should be. That is either the right bet or the most revealing one the field has ever placed.
Google DeepMind did not respond to a request for comment on whether it has a team focused specifically on model character and personality. MacAskill's own assessment — that Anthropic and OpenAI have such teams while Google does not — reflects his observations as an external researcher, not confirmed fact. The characterization is notable precisely because it comes from someone with no financial stake in which company looks ahead.
The Forethought papers are at forethought.org. The podcast is the March 2026 episode of 80,000 Hours featuring Will MacAskill.
