TOBY ORD

Behavioral Archetype

THE ACCOUNTANT OF THE FUTURE — Subject is the moral philosopher who turned the question “how likely is it that we do not survive the century” into a table with numbers in it, then gave the numbers away to anyone who would quote them. In The Precipice (2020) he put the chance of existential catastrophe this century at about one in six, and the chance from unaligned AI at one in ten, the largest single line in the table. He did the same thing with his own income first: in 2009 he fixed what he would live on and gave the rest away, and built a society, Giving What We Can, around the pledge. Both moves share a method, which is to make the thing countable. This profile scores the position, not the man: the hazard is the one the method creates, that a rounded probability in a popular book becomes the number a prime minister, a funder and a governance board all reason from, while the person who wrote it says in plain words that it is a rounded guess.

Essence Indicators

• Senior Researcher at the Oxford Martin AI Governance Initiative, University of Oxford: “Toby Ord is a Senior Researcher at the Oxford Martin AI Governance Initiative,” in the initiative’s own words (AIGI, Toby Ord); GovAI’s 2025 annual report repeats the line on its board page (GovAI, Annual Report 2025). His own page adds that “Toby is also a research affiliate at Forethought” (tobyord.com). Before the initiative he was at Oxford’s Future of Humanity Institute, which Oxford closed on 16 April 2024 (Forethought, Toby Ord; Daily Nous, 18 Apr 2024)

• A director of Centre For The Governance Of Ai Inc, the US nonprofit that houses GovAI: the Form 990 for the year ending December 2024 lists “Toby Ord (Director (beg June))” at $0 beside “Tasha Mccauley (Chair (beg June))” (ProPublica Nonprofit Explorer); GovAI’s report says “The board of our UK subsidiary comprises Toby Ord and two members of the GovAI team: Markus Anderljung, Director of Research and Policy, and Paul Harding, Operations Manager” (GovAI, Annual Report 2025). The chair of the US board is Tasha McCauley; the secretary is Helen Toner; Markus Anderljung sits with him on the UK board

• Founder of Giving What We Can: Giving What We Can is “the brainchild of Toby Ord, a philosopher at Balliol College, Oxford,” who “worked with fellow Oxford philosopher Will MacAskill to create an international organisation,” and “Ord and co-founder MacAskill launched Giving What We Can in November 2009” (Giving What We Can, Our history). That page reports “13,105” members who “have donated more than” $394,745,930; his own page says members “have donated over $300 million” (tobyord.com). Both are the organisation’s and his own self-reports

• His own pledge, in 2010 and 2020. The BBC reported in December 2010: “A year ago, he vowed to give away everything he earned above £20,000 … After his first year, he has lowered his personal allowance to £18,000,” and that he expected to give away £1m over his lifetime (BBC News Magazine, 13 Dec 2010). Vox reported in December 2020 that “As of 2020, he’s donated £126,128 … which is 26.6 percent of his earnings so far,” and that he “donated all royalties and advances from his book The Precipice, totaling more than £150,000 so far, to groups working on behalf of the long-term future of humanity” (Vox, Dylan Matthews)

• Author of The Precipice: Existential Risk and the Future of Humanity, published in the UK on 5 March and in the US on 24 March 2020 (80,000 Hours, podcast with Toby Ord). In his own later words: “In The Precipice, I gave my best guess probabilities for the existential risk over the next 100 years in each of these areas. These were 1 in 1,000 for Climate and for Nuclear, 1 in 30 for Pandemics and 1 in 10 for Unaligned AI” (Ord, “The Precipice Revisited,” 12 Jul 2024); on the show he gave the total: “this century, it’s about one in six” (80,000 Hours)

• Quoted by name by the head of the British government at the United Nations. Prime Minister Boris Johnson told the General Assembly on 22 September 2021: “In the words of the Oxford philosopher Toby Ord ‘we are just old enough to get ourselves into serious trouble’” (GOV.UK, PM speech at the UN General Assembly, 22 Sep 2021)

• Describes himself and is described as an adviser to governments and international bodies. His own page: “Toby has advised the United Nations, the World Health Organization, the World Economic Forum, and the UK Prime Minister’s Office”; the Oxford initiative’s page says the same with the OECD in place of the Forum: “Toby has advised the United Nations, the World Health Organization, the OECD, and the UK Prime Minister’s Office” (tobyord.com; AIGI); Forethought’s page: “including advising Google Deepmind, the UK Prime Minister’s office, and the UN” (Forethought). The UN’s own Our Common Agenda report (10 September 2021) does not name him in its text; an Effective Altruism Forum account says he “was asked to contribute to the report as such an ‘outside expert’” (UN, Our Common Agenda; EA Forum, 17 Sep 2021)

• Was among the few at the Future of Humanity Institute who held back when Sam Bankman-Fried’s money arrived: Asterisk’s history of the institute says “Ord, apparently, was one of the few who evinced reluctance to get involved,” and that he “delivered the eulogy” at the gathering staff called the institute’s “wake” (Tom Ough, Asterisk, issue 8)

On the Record

His own words, in date order, from his site and from the interview and reporting pages that carry him. Each quotation was matched against the page read.

The pledge, December 2010. On why he gave in public: “It’s not that amazing. I’m not that impressed by this, but I’m glad that people feel it’s a good story.” On the method: “it was much more sensible to work out at the start what you can live on [give away the rest in a lump sum] and then after a year readjust … instead of perpetually agonising about it.” On what he asks of others: “We’re not saying that if you don’t join up you’re a bad person” (BBC News Magazine, 13 Dec 2010).

The estimates, March 2020 and July 2024. The book’s table, as read on 80,000 Hours and then restated by him: climate and nuclear war at 1 in 1,000, engineered pandemics at 1 in 30, unaligned AI at 1 in 10, all sources at about 1 in 6. Four years on he wrote: “in my view, Climate risk is down, Nuclear is up, and the story on Pandemics and AI is mixed — lots of changes, but no clear direction for the overall risk,” and, on AI, “The racing is bad. The improved governance is good. The overall picture is mixed” (Ord, “The Precipice Revisited,” 12 Jul 2024). His reason the numbers did not move: they “are all rounded off (usually to the nearest factor of 10) and none of them have moved that far.”

Inference scaling, 12 February 2025. His governance argument in one sentence: “I shall argue that many ideas in AI governance will need either an adjustment or an overhaul.” On training-compute thresholds: “This threatens to break this entire approach of training-compute thresholds.” On disclosure: companies may “substantially scale up the intelligence of their leading models without anyone outside knowing,” and he calls for “policies to require disclosure of current capabilities.” He closes that “the shift to inference-scaling also makes the future of AI less predictable” and “may provide a premium on agility” (Ord, “Inference Scaling Reshapes AI Governance,” 12 Feb 2025).

Reinforcement learning, September and October 2025. On what he sees as the end of the previous era: “In the last year or two, the most important trend in modern AI came to an end,” and the new paradigm “reduces the amount of information a model can learn from per hour of training by a factor of 1,000 to 1,000,000” (Ord, 3 Oct 2025; Ord, 19 Sep 2025). In the 19 September piece he quotes Dario Amodei’s description of the shift from pre-training to reinforcement learning and writes of the labs: “AI labs have trumpeted the new reasoning abilities that RL unlocked, but have downplayed the ending of the remarkable era of pre-training scaling.”

Broad timelines, 9 March 2026. “No-one knows when AI will begin having transformative impacts upon the world.” His position: “The correct epistemic response to the lasting expert disagreement is to have a broad distribution over AI timelines.” His advice to funders and movement builders: “The ideal portfolio of activities aimed at making the AI transition go well should include a number of things that really help us succeed in worlds where we get longer to try,” while “We’re in a race against AI timelines. It is just that we don’t know if that race is a sprint or a marathon” (Ord, “Broad Timelines,” 9 Mar 2026).

AI scaling against human scaling, 9 October 2026. His most recent piece asks “Do AI models scale as well with more tokens as humans do with more time?” and, on a METR chart, “the humans gain about 4x as much score per doubling of the amount of resources as the AIs do” (Ord, 9 Oct 2026).

Social Persona / Impression Management

Immediate impression: The careful arithmetician. A philosopher who writes in short declarative steps, shows the table and then says what the table does not show. The 2024 revisit is typical: it lists what changed, says which risks went up and down, and declines to produce new decimals.

Energy: Steady and low-key. The BBC’s 2010 profile found a flat in Oxford “furnished sparingly in a style reminiscent of student digs,” and he answered the praise with “I’m not that impressed by this.” Fourteen years later his own writing is the same: one long essay at a time, on a personal site, with no promotional frame (BBC News Magazine, 13 Dec 2010; tobyord.com).

Impression management strategy: The honest rounding. He presents his own estimates as guesses that were rounded “to the nearest factor of 10,” publishes updates when the evidence moves, and credits the other side’s best data (the 2024 piece cites the IPCC narrowing its climate-sensitivity range as lowering risk) (Ord, 12 Jul 2024). The method is persuasive for the reason that it concedes: a reader who sees the number labelled as a rounded guess finds it harder to dismiss as alarm. The structural point is the one the profile scores: a number shown with its uncertainty travels further than one that is not, and his travel was documented in a Downing Street speech.

• Social accounts and feeds. X @tobyordoxford, linked from his Oxford Martin AI Governance Initiative page (AIGI, Toby Ord); feed tobyord.com/writing (RSS); personal site tobyord.com. No Substack or other newsletter is linked from the pages read.

Forensic Archetype Comparison

Psychometric Assessment

Observations of the documented public role, from his own writing, two interviews and three institutional pages. Trait scores here are style inferences, not claims about private character, and each says which exhibit it rests on.

Big Five (OCEAN):

Dark Triad:

MBTI: INTP (“The Logician”) — builds the framework first and revises it in public when the data change; treats an argument as a structure to test. First-hand rationale: “Broad Timelines” builds its case by definitions (the hikers and the cloud) before it argues from them, and “The Precipice Revisited” reports changes by risk, then declines to produce new numbers (Ord, 9 Mar 2026; Ord, 12 Jul 2024). This is a style inference from the documents, not a test result.

Threat Assessment

Score basis. ATK 5 is leverage: he is a director of the US entity that houses GovAI and sits on its UK board, and the UK prime minister quoted him at the UN, but no record shows him deciding what a deployed model does, who deploys it or which budget funds it (ProPublica; GOV.UK). DEF 6 is insulation: GovAI’s 2025 funders were “Coefficient Giving (formerly Open Philanthropy), the Survival and Flourishing Fund, Longview Philanthropy, along with many donors on the Giving What We Can and Every.org platforms,” not the labs (GovAI, Annual Report 2025); the counterweight is that his earlier institution was closed from above: “On 16 April 2024, the Institute was closed down” (Daily Nous). HP 7 is durability: The Precipice, with a probability table that outlives any seat, and Giving What We Can, which reports 13,105 members on a 2009 pledge (Giving What We Can). Against peers: Markus Anderljung (6/6/7, hazard 61) leads the research GovAI publishes and helped draft an EU code, so he holds more leverage and the same durability; Nick Bostrom (7/5/5, hazard 59.6) founded the institute that closed and has less surviving infrastructure than a pledge society with a book behind it. At 5/6/7 and 58, Ord sits between them: less leverage than Anderljung, more durability than Bostrom.

Alignment Analysis

Stated alignment: Reduce the chance that humanity loses its future: “existential security … a hypothetical future state of the world where existential risk is low and kept low,” built through “the norms and institutions to make sure things never get as out of control as they are this century” (Ord, 12 Jul 2024). On governance, disclosure and agility (Ord, 12 Feb 2025).

Observed alignment: Consistent with stated on the personal record: the numbers are published with their uncertainty, the pledge has been kept and reported for over a decade, and the writing updates on the evidence, including evidence against the loudest industry story (“The Extreme Inefficiency of RL for Frontier Models”). The structural record adds the part that is not about him: the seats he holds run on philanthropic money from the donors listed above, and the organisation he founded, Giving What We Can, is one of the donor platforms GovAI’s report thanks.

Gap assessment: No gap between what he says and what he does. The interesting gap is between the status of the number and its use: the book calls its table “my best guess probabilities,” and the same table is now a standard citation for how much attention AI risk deserves. The record does not settle whether a rounded estimate should bear that load; he says in his own words that they are “rounded off.” What the record can settle, and what this profile scores, is that the table travelled beyond the book: into a prime minister’s UN speech and a funding ecosystem he sits inside.

Interest: Documented, never motive, never a verdict.

1. Position: AI governance needs disclosure of current capabilities (“policies to require disclosure of current capabilities”; companies may “substantially scale up the intelligence of their leading models without anyone outside knowing,” Ord, 12 Feb 2025). What adoption pays: the essay names no GovAI or Oxford service, contract or seat that the rule would create. The structural fact is that a disclosure rule falls on the frontier developers and creates demand for the analysis GovAI’s output list shows (“Third-Party Compliance Reviews for Frontier AI Safety Frameworks,” “Incident Analysis for AI Agents”), at an organisation he directs and whose UK board he sits on (GovAI, Annual Report 2025). His best case: “the shift to inference-scaling also makes the future of AI less predictable.” Others who gain: regulators who need an outside view of capability, and, in his own argument, smaller players: changing the business model could mean “allowing more room for smaller players in the industry.”
2. Position: existential risk from AI is a top-tier priority, at about 1 in 10 (“1 in 10 for Unaligned AI,” Ord, 12 Jul 2024). What adoption pays: the framing directs money, careers and attention to the field in which his own employers and funders sit: GovAI’s 2025 funders are named above, and he gave “more than £150,000” of book earnings to “groups working on behalf of the long-term future of humanity” (Vox). His best case: “Existential risk from AI had finally gone from a speculative academic concern to a major new topic in global governance,” which he attributes to the 2023 statements by Hinton, Bengio and The Elders. Others who gain: the bodies that, in his words, now have “started to explicitly focus on existential risks as a major global priority” (the UN), and the funders whose programmes the number justifies.
3. Position: give at least a tenth, to the most effective charities (the pledge: “everything he earned above £20,000,” later £18,000, BBC). What adoption pays: pledges flow to the charities the society recommends; Giving What We Can donors are also among the donors GovAI thanks in its 2025 report (GovAI, Annual Report 2025). No payment to him personally is documented, and royalties from the book went to the causes. His best case: “You can help people so much without impacting on your own life.” Others who gain: the recipient charities, and the organisations that depend on the platform’s donors.
4. Position: plan for broad AI timelines, not short or long (“The correct epistemic response to the lasting expert disagreement is to have a broad distribution over AI timelines,” Ord, 9 Mar 2026). What adoption pays: the essay addresses “Funders, career advisors, and movement builders” and argues that the community “are likely underrating types of work that need five or more years in order to pay off”; an adopter portfolio would include work of the kind his institutions do. His best case: “none of this implies we can slack off.” Others who gain: every faction with a stake in the timeline debate, since the position refuses both the short-timeline and long-timeline camps: it names Dario Amodei at one end and Ege Erdil of Epoch AI at the other.

Alliance Map

The Corporate, Movement and Agency rows run at the same time, and that is the finding: one philosopher founded the pledge society, wrote the table, sits on the board of the research body that the same donor network funds, and had his sentence read into a prime minister’s UN speech. The two roles that pull against each other, the author of an estimate and a director of an organisation whose funding depends on how seriously the estimate is taken, sit in one record. He is also the person on the record saying the number is rounded.

Convergent Drive Classification

Self-preservation: The institution he worked in for ten years was closed by its faculty, and he moved to an Oxford initiative that kept the work going. First-hand: “On 16 April 2024, the Institute was closed down,” and his current title at the AI Governance Initiative (Daily Nous; AIGI). Goal preservation: The question is the same in 2020 and 2026: how much risk, from what, and what to do about it. First-hand: the 2024 post carries the 2020 table forward, and the 2026 post on timelines advises funders on the same question (Ord, 12 Jul 2024; Ord, 9 Mar 2026). Resource acquisition: The resources are donations and attention: a pledge society with 13,105 members and a donor network funding GovAI. First-hand: “We now have 13,105 members” and the 2025 funders list (Giving What We Can; GovAI, Annual Report 2025). Self-improvement: Revises his own estimates in public and switches subject to what the evidence now says about the models. First-hand: “Climate risk is down, Nuclear is up” in 2024, then reinforcement-learning efficiency and inference scaling in 2025 (Ord, 12 Jul 2024; Ord, 19 Sep 2025).

Subject is not an AI system. The drives appear anyway, in the estimate: a number written to be reasoned from, kept honest by its author, and carried by others into rooms he was not in.

Public footprint: tobyord.com · Oxford Martin AIGI page · Giving What We Can history.

Related: GovAI; Tasha McCauley; Helen Toner; Markus Anderljung; Nick Bostrom; Joe Carlsmith; Open Philanthropy; Revolving door graph.

Sources: Oxford Martin AI Governance Initiative, Toby Ord; tobyord.com; Ord, “The Precipice Revisited,” 12 Jul 2024; Ord, “Inference Scaling Reshapes AI Governance,” 12 Feb 2025; Ord, “The Extreme Inefficiency of RL for Frontier Models,” 19 Sep 2025; Ord, “Evidence that Recent AI Gains are Mostly from Inference-Scaling,” 3 Oct 2025; Ord, “Broad Timelines,” 9 Mar 2026; Ord, “AI Scaling vs Human Scaling,” 9 Oct 2026; Forethought, Toby Ord; GovAI, Annual Report 2025; ProPublica Nonprofit Explorer, Centre For The Governance Of Ai Inc; Giving What We Can, Our history; BBC News Magazine, “Toby Ord: Why I’m giving £1m to charity,” 13 Dec 2010; Vox, Dylan Matthews, Toby Ord on his pledge; 80,000 Hours, podcast with Toby Ord; GOV.UK, PM speech at the UN General Assembly, 22 Sep 2021; UN, Our Common Agenda; EA Forum, Major UN report discusses existential risk and future generations; Daily Nous, “The End of the Future of Humanity Institute,” 18 Apr 2024; Tom Ough, “Looking Back at the Future of Humanity Institute,” Asterisk, issue 8