Essay 01
· 16 min read
For the rest of my working life, I expect the defining breakthroughs in the world to come from the private sector. Capital, and the talent that follows it, are concentrated there, so public institutions (in the West) are unlikely to fund or build frontier science and technology at the scale and speed our global crises demand. The pursuit of artificial general intelligence (AGI) will deepen that pull, drawing even more of the world’s value formation inside companies, where a few people decide what’s worth building and market forces decide who gets it. Unless we build deliberate paths out, that value won’t reach the people who need it most, on terms that work for them.
Consider AlphaFold, which I had the privilege to work on. For 50 years, scientists tried to predict protein structures (which determine function, as the spike protein did for COVID therapeutics) from amino acid sequences, and since 1994 they’ve competed in a contest called CASP to do it. Within a decade of Alphabet’s 2014 acquisition, DeepMind, with compute and talent no one else had, didn’t just solve it. It won the Nobel.12
I don’t love the fact that the frontier is easiest to pursue inside corporations, and suspect many elder millennials feel similarly conflicted. We began university as post-9/11 aspiring policy wonks, believing public, international institutions like the World Trade Organisation, European Union and International Criminal Court were routes to an enlightened, globally stable ‘end of history,’ and then went corporate in an era of ‘hope and change,’ convinced that tech companies could be paths to global development, too. That may have been naive, but it wasn’t unfounded. At Google, I wrote the testimony for, and briefed Congress on, why we pulled our search engine out of China in response to a security hack that targeted human rights activists.
As the first to experience social media before our prefrontal cortexes were fully developed, the first to take jobs in a hot, internet-fuelled tech industry, and now the ones ushering in AI from leadership roles, we’ve seen the limits as clearly as the possibilities: good intentions and high-performing technology aren’t enough. Now in our 40s, we have much to answer for. As the newly feted East of Eden rendition posits, we must take responsibility for how we show up in the world. Our (technological) futures are not predetermined. We may choose, and choosing a better path to shared progress is what we really owe the future.
What grows inside
can take root elsewhere.
Being the adults in the room
If you, like me, spent your career betting on policy and international institutions to steer technology, I want to propose a different starting point: use private rails as the machinery (the speed, the capital, the talent, the market discipline), and put our energy into deliberately engineering the conditions that make the value they generate serve and respond to the public. Here’s why.
Let’s start with the conditions and constraints for those of us considering what to do for the remainder of our careers. Not the projected figures of when AGI will arrive, but the facts we know for sure.
- Time is finite: we’ve got about 20-25 good years left working at this pace, if that.
- History suggests economic and social structures rarely change quickly without a systemic shock on the scale of a world war or an economic crisis. Barring one, societal change will lag even as AI speeds everything else up.
- Recent valiant attempts to stretch capitalism toward beneficial technology through impact investing haven’t produced the step-function changes we’d hoped, partly because today’s market incentives work against it, altruism is unreliable, and policy is erratic (e.g. climate regulations).
- Private R&D spend has far outpaced public spend, making it a no-brainer for any aspiring founder to raise from the bigger pool, irrespective of the terms.
In the US, business now funds three-quarters of all R&D, up from under a third in the 1960s, while the federal share has fallen below a fifth. Google, Amazon, Microsoft and Meta have guided to roughly $725 billion of capital expenditure in 2026 alone, while NHS England’s entire annual revenue budget reaches £226 billion by 2028/29. It isn’t a like-for-like comparison, but directionally, no public balance sheet in Europe competes.
Changing any of this quickly is unlikely, and waiting for it is a misuse of limited time. Instead of bemoaning the circumstances, wishing them away, or ignoring them, we should accept these constraints and focus on the levers we can pull to create public goods from the most important technology of our time, even if it’s privately owned.
I came up in tech policy believing that if we got the laws, norms, and code right, everything else would follow. But I hadn’t properly read Lawrence Lessig’s Code, which argued that four forces shape behaviour: law, social norms, the design of technology itself, and markets. We’ve spent the last two decades pulling hard on the first three and largely leaving the fourth, the one that decides what gets built and who it reaches, to companies. Laws gave us cookie pop-ups everyone clicks through, years of social media harms before governments could agree on age limits for access, and reporting requirements for frontier AI models trained on computing power that exceeded 10^26 flops, which quickly became irrelevant despite industry cooperation. Better design gave us privacy dashboards no one uses. Enough tech policy work has stalled, and enough well-meaning governance structures and social movements have run dry, for us to see that good intentions don’t scale; market incentives do. For now, the most durable alignment mechanism we have might be pointing markets in the right direction.
That’s possible because ‘the market’ isn’t some immutable force. It’s a human-designed system. The question isn’t whether competition or ethics wins the day, but what kind of competition society chooses to incentivise. And that might include regulation, but doesn’t rely on it alone. It could also mean redefining what’s possible in the private sector, for example, establishing entirely new markets, listening to communities and experts in ways tech has never done well, or orchestrating new asset classes to finance urgent work no one is doing at scale. These are new institutions we need to build in the age of AI, because the traditional consumer and enterprise paths won’t get us to the public benefit we hope for. Create the runway the plane can actually land on, rather than spread-betting on launch pads and building in flight.
None of this works without patience, tenacity, empathy, and dare I say anthropology, all largely absent in a sector that deifies moving fast and breaking things. That is what will take us from an ideology of goodness toward the actual craft of practising it.
Private capability
Capital, compute, talent and technical progress.
The work of diffusion
Public input, useful evidence, partners, financing and capacity.
Public benefit
Useful outcomes that people can reach and help shape.
A conceptual map of the essay’s argument. The work is iterative; opening access is one part of it.
Nothing for us without us
Who gets to define what it means for technology to deliver benefits to society? Benefit is not unilaterally bestowed. As my friends and AI scholars Shakir Mohamed, William Isaac, and Marie-Therese Png have written, it requires the receiving party to independently agree that the shift is indeed beneficial. Asked for a single word to describe AI, Britons most often say “scary” and “worried”, yet technologists too often put that down to a lack of technical understanding.
Even democratic societies struggle with appropriate means of representation, so I concede there is no easy way to close the gap between the public and those building the technology. But for the public to trust technology, they first need to believe someone in the room is looking out for their interests. That starts with going back to first principles about who has the political legitimacy to decide what. Companies decide what is built and released, but elected representatives are the ones with the legitimacy to decide how value should be shared, with input from the public, companies included. Corporate experiments with nonprofit structures, windfall clauses and proposals for Universal Basic Income are well-intended, and show foresight about the inequalities AI will likely produce, but they can’t stand in for democratic choices. That’s part of the problem with over-extending the role of the private sector: companies can’t take on a role traditionally held by government without also taking on its accountability to the public. Representation, in its most basic sense, needs to be real, and that means working through elected officials, not around them.
I’m still figuring out where the work of building that trust should sit. It’s likely spread across governments, industry, funders, and think tanks or academia. But between bashing big tech, knee-jerk (or expiring) regulation, and letting the market do what it wants, there has to be a place that makes for a better, softer landing. And of course this may all sound soft. But influence in its most effective form is soft power: shaping what people want to do, rather than forcing them.
Practising the craft
Before lockdown, I sat in a room convened by DeepMind’s then CTO and COO, where a cross-disciplinary group of leads was deciding whether to use our experimental AlphaFold to predict unknown protein structures of COVID-19. The system wasn’t quite ready and we couldn’t guarantee accuracy, though we could give confidence intervals. After we decided to go ahead, with the necessary disclosures and precautions in place, notably not overhyping the work, I asked: why COVID? The pandemic touched everyone, but plenty of other diseases hit other communities just as severely and urgently, with far less attention.
I was lucky (and remain grateful) that DeepMind’s leadership let my team pitch the Drugs for Neglected Diseases Initiative (DNDi) on validating protein structure predictions for two long-neglected diseases: Chagas disease and Leishmaniasis. The collaboration kept the company’s commercial options intact while giving the people who needed it most access in exchange for validation and feedback, and it was the precursor to openly releasing AlphaFold. This was the first time I saw how to, in my friend Halle Tecco’s words, align mission and margin to deliver public goods from private rails, and the moment I realised how underused those rails were. I’d spent a decade trying to change technology through laws and norms. Here the change came from inside, through what a company chose to give away and how it justified the value exchange for it. But such interventions aren’t the norm, and they aren’t guaranteed. A few conditions enabled it and underpinned the later open-source release:3
- Approaching the question without denigrating commercial or scientific responsibility. Both matter. It took money to finance AlphaFold, and shareholders deserve to understand why it is being given away. Conversely, it’s worth asking whether withholding scientific progress on diseases affecting many people is morally defensible.
- Making these decisions with input from multidisciplinary communities inside and outside the company, from bioethicists and safety experts to scientists and doctors on the ground. DNDi runs clinical trials alongside health workers in the countries where these diseases are endemic, often in remote areas. That proximity made them the right partner: they knew which structures could move a treatment forward, and which would just make a nice paper.
- Paying attention to timing and the social and political context. Releasing early COVID structures made sense because the world was in an emergency and scientists needed anything that might help, caveats attached. Waiting can be right, too. After we showed AlphaFold worked, it took about eight months to release it openly, and some scientists were understandably frustrated. We lived with that criticism while we validated the system, worked out what was responsible to release, and built the infrastructure to support it. Context also shapes how a decision lands: who is already sceptical, what promises the industry has recently broken, which communities feel things are being done to them rather than for them. A good decision at the wrong moment can cost trust that took years to build.
- Licensing the technology free for nonprofit and academic use, while reserving commercial use to sustain the work, as later happened with AlphaFold 3. As my friend Kyunghyun Cho suggested, a non-compete clause in the licence could stop the free version being used to build a competing commercial product. Otherwise the open release subsidises rivals and the business that funds the work loses its reason to exist. It also makes the case to shareholders easier: openness for science, protection for the business that pays for it.
- Not underestimating what it takes for benefit to take hold. Since then, DeepMind has launched funds and teams to help scientists who want to use AlphaFold but lack the compute or the AI and coding fluency to make it work for them. Open Athena provides exactly this engineering capacity, and more such bridge organisations need to be funded by the entities funding startups if they want these solutions to land.
None of this guarantees anyone will get it right the first time or every time. But that doesn’t mean it’s not worth trying. In its absence, people are turning to cleaner, simpler theologies, ones where doing good is a matter of arithmetic, instead of grappling with the complexity and resistance that come with any worthwhile change. The real world is much messier. Those of us who’ve been in the room know that. It’s time to roll our sleeves up and put all that hard-won understanding to use.
Threading the needle
Here’s my bet, having done this before: as AI, and eventually AGI, speeds up discovery, more public goods than ever will be created inside companies. Our job is to find them and make it easier, and worth the cost, to get them out.
To make that happen, someone with good judgement (not just moral ambition) has to do four things:
- Identify and value the public good honestly. This usually starts with the founder, checked by an independent reviewer who can judge what the asset is worth to the field.
- Clear the fiduciary bar. Giving something away can’t breach the company’s duties, which puts this with the board and investors. The choice is not between values and no values, but between implicit, unexamined ones and explicit, contestable ones. Shareholder value and fiduciary responsibility can be perfectly good reasons for keeping something private, as can safety, but these constraints need to be well articulated and publicly tested for people to believe you. A short release policy, published alongside each major launch and updated as things change, would go a long way.
- Find a partner who proves the public value. It should be on a real, unglamorous problem, as DNDi’s work on neglected diseases was for AlphaFold. Serving the margins can also improve how you serve everyone: captions, voice control and text-to-speech were built for users with disabilities and are now used by almost everyone (the so-called curb-cut effect).
- Design and finance the open release. It needs to be sustained rather than abandoned, often with a catalytic funder. Even ring-fencing funding to maintain the open version for a fixed number of years would help it outlast the launch, alongside bridge resources like engineering capacity, datasets and evaluations that help people actually use it.
This is a structured decision process that needs to reconcile genuine commercial value with durable, open access. We need more people and organisations that will specialise in doing this, and we need venture capitalists and foundations to fund it directly, alongside the founders they back, with support from government.
The obvious objection is that AlphaFold was a one-off, from an unusually well-resourced lab. So it matters that the same structure has been repeated further down the food chain, at an ordinary venture-backed company (though special to me, for its creative founders, whom I was proud to support).
Take Spore Bio, a Paris-based deep-tech company founded in 2023. Its three founders are children of immigrants who grew up in their families’ restaurants, where they saw how a single contamination scare could threaten a small business. They became scientists to fix the problem. Microbial contamination testing is the hidden infrastructure behind the safety of food, water, medicine and cosmetics, and it has barely changed in a century: culture a sample, then wait days. Spore cuts that to minutes, using AI and spectral imaging to read microbial signatures in real time on a platform built end to end. Sanofi and Clarins are among its early partners, and it has won a European Innovation Council Accelerator award.
Its business also throws off a positive externality of obvious public value: one of the world’s best continuously updated antimicrobial-resistance (AMR) datasets, on one of the defining health threats of the coming decades. Spore’s founders approached Google.org’s AI for Science Fund with a proposal to make this dataset open source for nonprofit and academic use, and were selected for a multi-million-euro award to stand up Spore.Labs, which will publish findings and validate them with academic and clinical partners. Catalytic philanthropic capital funded the open public good and venture capital funded the company, so the same science, done on private rails, serves the public and customers while returning value to shareholders.4
Many science founders raised money originally to do the science and publish it for the world. They didn’t necessarily intend to build a company, but with government funding cuts and scarce grants, venture was the fastest option. Along the way, they realise they hold a unique dataset, model, or piece of scientific infrastructure that the whole world could build on but that is unlikely to earn much on its own. Investors back what returns capital, so an asset that won’t earn is shelved, or sold narrowly to whoever will pay. A funder who pays for the open release removes that cost, so the company doesn’t sacrifice revenue or investor goodwill. Finding other financing to open-source the externality for academic and nonprofit use, while the company keeps licensing the same work commercially, is gold.
Plenty of positive externalities exist but lack the engagement and backing to reach fruition. In my own work, I’ve pushed for tools like the open-sourced football system TacticAI and a double-digit number of cancer tools at Google and DeepMind to be broadcast and shared more widely. TacticAI is now used by major football clubs, and the cancer tools have attracted research grants. What would the world look like if this mechanism were repeated and practised more intentionally?
Open-sourcing an asset has limits, of course. Access is not the same as equity: using the asset requires compute, talent, and institutional capacity that are unequally distributed. Funding that capacity, via public compute, deployment muscle, and domain-specific data, is where new institutions and government could lead (see Open Athena and similar groups above).
One instrument, not the whole toolkit
Open-sourcing the externality is the method I know best, but it is only one instrument. We need a movement with many shots on goal. This series will describe some of the others. Many already work well in a single sector, like climate or global health, but none is sufficient alone. Together they are the beginnings of a discipline: ways to route the value on private rails toward the public, chosen to fit the situation. Open-source the externality when it’s separable, as Spore did with its AMR data. Commit demand when the outcome can be specified, as Frontier did by pre-buying carbon removal. De-risk with first-loss capital when the market can’t read the return, as the Gates Foundation and Sida did by guaranteeing investors in the Global Health Investment Fund against part of their losses. Gate funding on outcomes, releasing more as agreed milestones are met, with community-driven benchmarks rather than financial performance alone. Cap returns and use the difference to finance public goods, as Jonathan Soros and Camille Canon have proposed. We need more creative solutions like these to make public goods durable. There will be more than I can name, and inventing them is some of the most important work available right now.5
- When a public good can be separatedOpen the externality
- Release the useful asset and finance the capacity to use and maintain it.
- When an outcome can be specifiedCommit demand
- Give builders a credible buyer for something the public needs.
- When the market cannot read the returnUse first-loss capital
- Let one layer of funding absorb initial losses to reduce risk for others.
An editorial summary of the instruments Dorothy describes; each depends on the situation.
Betting on private rails to deliver public goods could read as technosolutionism, but it isn’t. Technosolutionists put their faith in the technology alone. I’m putting mine in the people around it: talented builders with the right incentives, working alongside the communities who know their problems best. This matters most in the domains the market underfunds, like climate, health, and education. Despite AlphaFold, biotech has gone through a funding winter. That only makes sense once you see that venture capital expects returns within 5-7 years or less, a timeline biology rarely keeps to. Breakthroughs will keep arriving on private rails, and leaving their survival to the market is not neutral but negligent. If we can’t engineer equitable delivery, the rest is academic. If we can, it’s the difference between a technology that widens who gets to benefit from progress and one that narrows it to whoever owns the cap table.
We don’t need another breakthrough to do this. We have the capital, the builders, and the need. What we lack is the will to point the capital at what we actually want, the ingenuity to build the pathways, and the imagination to believe progress can be shared, not hoarded. For now, the rails are private; someday that might change. But until then, let’s figure out how to get cutting-edge technology to benefit as many people as possible, starting with those who need it most.
Sources & further reading
- CASP: the protein-structure prediction community’s assessment programme ↗ ↩
- The Nobel Prize in Chemistry 2024 ↗ ↩
- AlphaFold and the Drugs for Neglected Diseases initiative ↗ ↩
- Spore.Bio: recipient of the Google.org AI for Science fund ↗ ↩
- An example of milestone-based funding: CEPI Programme Document (2019), §§4.3.2–4.3.3 ↗ ↩
These links provide context for the historical and scientific examples. Personal accounts and the argument are Dorothy’s own.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
I don’t particularly like this fact, and suspect many elder millennials feel similarly conflicted. We began university as post-9/11 aspiring policy wonks, believing public, international institutions like the World Trade Organisation, European Union and International Criminal Court were routes to an enlightened, globally stable ‘end of history,’ but then went corporate to cope with the 2008 crash. During that era of ‘hope and change,’ we convinced ourselves that morally coded tech companies were better paths to global development than slow government and nonprofit ranks. My friend Claire Stapleton’s book documents the mental and moral gymnastics many of us performed to justify the tradeoffs. Two decades later, I’m not sure we nailed the brief. Many of us have survived sexual harassment and endured the financialisation and bro-ification of our industry, yet we’ve done little to curb a new guard from promising the next generation that powerful technology will ‘make the world a better place.’
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
History shows economic and social structures change rapidly only in response to violent events like wars and famine, so without one this is a long game, AI or no AI.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
I came up in tech policy believing that if we got the laws, norms, and code right, everything else would follow. But I hadn’t properly read Lessig, who argued that four forces shape behaviour: law, social norms, the design of technology itself, and markets. We’ve spent the last two decades pulling hard on the first three and largely leaving the fourth to companies. Enough tech policy work has stalled, and enough well-meaning governance structures and social movements have run dry, for us to see that good intentions don’t scale; market incentives do. Regulation on its own has given us cookie pop-ups everyone clicks through, privacy dashboards no one uses, years of social media harms before governments could agree on age limits for access, and reporting requirements for frontier AI models that used more than 10^26 flops, which quickly became irrelevant despite industry cooperation. For now, the most durable alignment mechanism we have might be pointing markets in the right direction.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
What I am suggesting is that we start using private rails as the machinery (the speed, the capital, the talent, the market discipline) and put our energy into deliberately engineering the conditions that make the value they generate serve and respond to the public. This might include regulation but doesn’t rely on it alone. It could also mean redefining what’s possible in the private sector, for example, establishing entirely new markets, listening to communities and experts in ways tech has never done well, or orchestrating new asset classes to finance urgent work no one is doing at scale. These are new institutions we need to construct in the age of AI, because the traditional consumer and enterprise paths won’t get us to the public benefit we hope for. Create the runway the plane can actually land on, rather than spread-betting on launch pads and building in flight.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
The same pattern shows up closer to home. When UK founders recently pushed to curb notice periods and garden leave, the people with the most to lose, those in operating roles, far more exposed to layoffs than technical talent and more reliant on that notice as a cushion, weren’t the ones asking for it. Decisions about technology are no different: they need the people they affect most in the room. If we can’t consider the wellbeing of colleagues sitting right next to us, how can we be expected to do so for the rest of society?
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Even democratic societies struggle with appropriate means of representation, so I concede there is no easy way to close that gap. But for the public to trust technology, they first need to believe someone in the room is looking out for their interests. Here are a few building blocks to consider, and they’re not exhaustive (as a colleague once said to me, therapy is also an option; I go weekly and it spares many people in my life):
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Survey market failures where we know there is a need but inadequate financing, and figure out why. Despite AlphaFold, biotech has gone through a funding winter. That only makes sense once you see that venture capital expects returns within 5-7 years or less, a timeline biology rarely keeps to. Moreover, AI’s introduction into biology has created a valley of death: tech investors finance platforms to simulate compounds, while bio investors only care about a target or molecule, and no one pays for the wet lab validation in between. Similarly, women’s health holds enormous opportunity, but most women rank their family’s wellbeing above their own, and mothers are less likely to get preventive care, which makes the market harder to read. These are structural and social dynamics worth investigating and unlocking.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Take a hard look at the communities suffering most from these market failures. Listen without judgement so the strategy doesn’t exclude people with lived experience. Maybe even hire a few of them. Be curious about the edge cases the market doesn’t prioritise or build for, since serving the margins can improve how you serve everyone. Accessibility technology is a great example: captions, voice control and text-to-speech were built for users with disabilities and are now used by almost everyone (the so-called curb-cut effect). Instead of assuming you’re doing good, stop to ask someone who might disagree.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Put time and money toward problem definition instead of starting by evaluating a founder or company. Figure out where a field needs to be in a decade and what conditions beyond good tech will enable safe uptake, from pilot to widespread launch. Pair it with proven techniques like gating funding on outcomes (releasing more as agreed milestones are met) and community-driven benchmarks rather than financial performance alone. It may also mean providing bridge resources between technological solutions and communities, like engineering capacity for implementation or new datasets and evaluations.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Decide what you’ll give away, and say why. Since the bulk of the work is happening in companies, create a plan for which goods will be kept private and which released publicly. Experiments with nonprofit structures, windfall clauses and Universal Basic Income are well-intended, and show foresight about the inequalities AI will likely produce. But how wealth gets redistributed is a democratic decision, not a corporate one. Why should the companies accruing the value also decide how it’s shared, or what UBI looks like? What they do control is what they build and release. What will you openly give away? What won’t you, and why? Shareholder value and fiduciary responsibility can be perfectly good reasons, as can safety, but these constraints need to be well articulated and publicly tested for people to believe you. A short release policy, published alongside each major launch and updated as things change, would go a long way.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Create ways for public goods from private companies to be sustainably financed. Sometimes companies remain good homes for public goods; sometimes they don’t. Should they be housed in public institutions with grants or government funding? Do we look to Jonathan Soros and Camille Canon’s proposals for capped returns, using the difference to finance public goods? We need more creative solutions like these to make public goods durable. Even ring-fencing funding to maintain the open version for a fixed number of years would help it outlast the launch.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
I’m still working out where this work should sit. It’s likely spread across governments, industry, funders, and think tanks or academia. But between bashing big tech, knee-jerk (or expiring) regulation, and letting the market do what it wants, there has to be a place that makes for a better, softer landing. And of course these mechanisms themselves may feel soft. But influence in its most effective form is soft power: shaping what people want to do, rather than forcing them.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
I was lucky (and remain grateful) that DeepMind’s leadership let my team pitch the Drugs for Neglected Diseases Initiative (DNDi) on validating protein structure predictions for two long-neglected diseases: Chagas disease and Leishmaniasis. The collaboration kept the company’s commercial options intact while giving the people who needed it most access in exchange for validation and feedback, and it was the precursor to openly releasing AlphaFold. This was the first time I saw how to, in my friend Halle Tecco’s words, align mission and margin to deliver public goods from private rails. I’d spent a decade trying to change technology through laws and norms. Here the change came from inside, through what a company chose to give away and how it justified the value exchange for it. But such interventions aren’t the norm, and they aren’t guaranteed. A few conditions enabled it and underpinned the later open-source release:
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Clear the fiduciary bar. Giving something away can’t breach the company’s duties, which puts this with the board and investors.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Find a partner who proves the public value. It should be on a real, unglamorous problem, as DNDi’s work on neglected diseases was for AlphaFold.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Design and finance the open release. It needs to be sustained rather than abandoned, often with a catalytic funder.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Open-sourcing the externality is the method I know best, but it is only one instrument. We need a movement with many shots on goal. This series will describe some of the others. Many already work well in a single sector, like climate or global health, but none is sufficient alone. Together they are the beginnings of a discipline: ways to route the value on private rails toward the public, chosen to fit the situation. Open-source the externality when it’s separable, as Spore did with its AMR data. Commit demand when the outcome can be specified, as Frontier did by pre-buying carbon removal. De-risk with first-loss capital when the market can’t read the return, as the Gates Foundation and Sida did by guaranteeing investors in the Global Health Investment Fund against part of their losses. There will be more than I can name, and inventing them is some of the most important work available right now.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Betting on private rails to deliver public goods could read as technosolutionism, but it isn’t. Technosolutionists put their faith in the technology alone. I’m putting mine in the people around it, from the communities who know their problems best to the builders, and those who set their incentives. That matters most in the domains the market underfunds, like climate, health, and education. Breakthroughs will keep arriving on private rails there, and leaving their survival to the market is not neutral but negligent. If we can’t engineer equitable delivery, the rest is academic. If we can, it’s the difference between a technology that widens who gets to benefit from progress and one that narrows it to whoever owns the cap table.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
That serves no one. Confusing, hyperbolic headlines erode our belief that our choices matter, and that erodes our willingness to make crucial decisions for our lives. What I hope to do in the pieces that follow is to provide a few ideas for a path between dystopia and utopia to help decision-makers, from public officials to private individuals, consider some practical responses to the latest developments in AI. All of it requires reclaiming agency, both personal and collective, and taking responsibility for how AI diffuses across society.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Science and engineering can only take us so far. They are the practice of finding the best solution within a given set of constraints. Democracy is something else: the practice of “forming a more perfect union,” or, put more simply, figuring out how we live together.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
None of this is about whether people use AI. Billions already do, even as many resist the infrastructure it runs on. And in much of the non-Western world, people are optimistic about what it can do for them. What matters is whether we can collectively shape how it's absorbed so that everyone benefits. It’s true that we will be able to solve problems that we couldn’t without AI, but it’s equally true that there are plenty of problems AI won’t be able to solve without us. Take the NHS. AI could tackle the acute challenge of queueing for care, but turning the system toward preventative care rather than treating the sick isn't something it can operationalise alone. Building the trust and incentives for transformational change is work only the people within and around the system can do.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Like it or not, many of the most important innovations of the next few decades will come from the private sector. Whether they make for a better society is far from settled, and that is where all of us, from the policymakers who represent us to the CEOs who are citizens too, have the most leverage. The five pieces that follow explore who gets to share in the value created, how proven technology gets stuck before it reaches those who need it most in public services like the NHS, what ‘sovereign AI’ should actually mean, why governments should focus on shaping demand rather than building supply, and where we need to build what the market won’t deliver on its own. Underneath it all is a bet that when people experience science and technology in ways that benefit them directly, they are far more likely to welcome it into daily life and to trust the institutions that brought it to them.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
For many of us who have worked on this technology, and believe in its ability to benefit society, this is a crucial wager. Like Arthur Galston, whose work on soybean flowering eventually led to Agent Orange, I have come to believe that “the only recourse for a scientist concerned about social consequences is to remain involved with it to the end.” Whatever happens to valuations, the question is not whether AI will create value. It is who that value is for, and what we are willing to do about how it is developed and deployed, given the knowledge and the agency that are ours.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
After spending nearly 20 years in frontier technology, I find the current conversation on AI exhausting. Rather than earning the public mandate to build world-changing science and technology, competing platitudes about apocalypse and abundance dominate the headlines, making it hard for anyone to know what to believe, let alone what to work toward. Instead of being focused on casting a vision, bringing the public along, and creating the operational capacity to deliver it safely, we’ve regressed to the attention economy’s default of oversimplifying the message at the risk of inciting collective panic.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
That serves no one. Confusing, hyperbolic headlines erode our belief that our choices matter, and that erodes our willingness to make crucial decisions for our lives.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Companies aren't governments and CEOs aren't elected, but when a company claims to be building world-altering technology, democratic legitimacy still matters. When people's will is overridden by decision-makers they can't or won't trust, the social and economic fabric starts to fray. The strain is already visible: in data centres becoming a flashpoint during the US midterms, in people asking who decided this technology should be built at all, and in widespread fear about job loss and surveillance. Governments, which are meant to represent the public, have been reactive at best. They could be setting out a collective vision for the outcomes they want from AI, like faster, more accurate diagnosis for patients or helping students who’ve fallen behind to read, and using their buying power to steer the public and private sectors toward them. Instead, the focus has stayed on geopolitical advantage. Now, as politics swing from boosterism to backlash, responsibility for where this technology leads still falls between companies and governments, with neither fully owning it.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
None of this is about whether people use AI. Billions already do, even as many resist the infrastructure it runs on. And in much of the non-Western world, people are optimistic about what it can do for them.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Like it or not, many of the most important innovations of the next few decades will come from the private sector. Whether they make for a better society is far from settled, and that is where all of us, from the policymakers who represent us to the CEOs who are citizens too, have the most leverage.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
For those of us who have worked on this technology, and believe in its ability to benefit society, that responsibility is personal. Like Arthur Galston, whose work on soybean flowering eventually led to Agent Orange, I have come to believe that “the only recourse for a scientist concerned about social consequences is to remain involved with it to the end.” Whatever happens to valuations, the question is not whether AI will create value. It is who that value is for, and what we are willing to do about how it is developed and deployed, given the knowledge and the agency that are ours.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
After spending nearly 20 years in frontier technology, I find the current conversation on AI exhausting. Competing platitudes about apocalypse and abundance make it hard to know what to believe, let alone what to work toward. They erode our belief that our choices matter. We should be earning a public mandate and creating the capacity to deliver safely.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Companies aren't governments and CEOs aren't elected, but democratic legitimacy matters when a company claims to be building world-altering technology. Governments could set out a collective vision for AI and use their buying power to steer toward it. Instead, as politics swing from boosterism to backlash, responsibility still falls between companies and governments, with neither fully owning it.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
None of this is about whether people use AI. Billions already do, and many in the non-Western world are optimistic about it. Many of the most important innovations will come from the private sector. Whether they make for a better society is far from settled.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
For those of us who have worked on this technology, that responsibility is personal. Like Arthur Galston, whose work on soybean flowering eventually led to Agent Orange, I have come to believe that “the only recourse for a scientist concerned about social consequences is to remain involved with it to the end.” The question is not whether AI will create value. It is who that value is for, and what we are willing to do about how it is developed and deployed.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
For the rest of my working lifespan, I expect the defining breakthroughs in the world to come from the private sector. Capital, and the talent that follows it, are concentrated there, so public institutions (in the West) are unlikely to fund or build frontier science and technology at the scale and speed our global crises demand. The pursuit of artificial general intelligence (AGI) will deepen that pull, drawing even more of the world’s value formation inside companies, where a few people decide what’s worth building and market forces decide who gets it. Unless we build deliberate paths out, that value won’t reach the people who need it most, on terms that work for them.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
I don’t particularly like this fact, and suspect many elder millennials feel similarly conflicted. We began university as post-9/11 aspiring policy wonks, believing public, international institutions like the World Trade Organisation, European Union and International Criminal Court were routes to an enlightened, globally stable ‘end of history,’ and then went corporate in an era of ‘hope and change,’ convinced that tech companies could be paths to global development, too. That may have been naive, but it wasn’t unfounded. At Google, I wrote the testimony for, and briefed Congress on, why we pulled our search engine out of China in response to a security hack that targeted human rights activists. Two decades later, having survived sexual harassment and the financialisation and bro-ification of our industry, many of us see far more clearly what technology companies can and can’t achieve for society. Yet we’ve done little to stop a new guard from blindly promising that AI will ‘make the world a better place.’
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
As the first to experience social media before our prefrontal cortexes were fully developed, the first to take jobs in a hot, internet-fuelled tech industry, and now the ones ushering in AI from leadership roles, we’ve seen first-hand that good intentions and high-performing technology aren’t enough. Now in our 40s, we have much to answer for. As the newly feted East of Eden rendition posits, we must take responsibility for how we show up in the world. Our (technological) futures are not predetermined. We may choose, and choosing a better path to shared progress is what we really owe the future.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
History suggests economic and social structures rarely change quickly without a systemic shock on the scale of a world war or a great depression. Barring one, this is a long game, AI or no AI.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
I came up in tech policy believing that if we got the laws, norms, and code right, everything else would follow. But I hadn’t properly read Lessig, who argued that four forces shape behaviour: law, social norms, the design of technology itself, and markets. We’ve spent the last two decades pulling hard on the first three and largely leaving the fourth, the one that decides what gets built and who it reaches, to companies. Laws gave us cookie pop-ups everyone clicks through, years of social media harms before governments could agree on age limits for access, and reporting requirements for frontier AI models that used more than 10^26 flops, which quickly became irrelevant despite industry cooperation. Better design gave us privacy dashboards no one uses. Enough tech policy work has stalled, and enough well-meaning governance structures and social movements have run dry, for us to see that good intentions don’t scale; market incentives do. For now, the most durable alignment mechanism we have might be pointing markets in the right direction.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
That’s possible because ‘the market’ isn’t some immutable force. It’s a human-designed system. The question isn’t whether competition or ethics wins the day, but what kind of competition society chooses to incentivise. And that might include regulation, but doesn’t rely on it alone. It could also mean redefining what’s possible in the private sector, for example, establishing entirely new markets, listening to communities and experts in ways tech has never done well, or orchestrating new asset classes to finance urgent work no one is doing at scale. These are new institutions we need to construct in the age of AI, because the traditional consumer and enterprise paths won’t get us to the public benefit we hope for. Create the runway the plane can actually land on, rather than spread-betting on launch pads and building in flight.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Many of the loudest voices in AI claim to be the only ones capable of delivering beneficial tech. To me, this betrays a lack of self-awareness and humility: who asked them? Benefit is not unilaterally bestowed. As my friends and AI scholars Shakir Mohamed, William Isaac, and Marie-Therese Png have written, it requires the receiving party to independently agree that the shift is indeed beneficial. Asked for a single word to describe AI, Britons most often say “scary” and “worried”, yet technologists tend to ignore that or put it down to a lack of technical understanding. The result is widespread gaslighting: what you feel is inaccurate because you’re too uninformed to understand it.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Even democratic societies struggle with appropriate means of representation, so I concede there is no easy way to close that gap. But for the public to trust technology, they first need to believe someone in the room is looking out for their interests. That starts with going back to first principles about who has the political legitimacy to decide what. Companies decide what is built and released, but elected representatives are the ones with the legitimacy to decide how value should be shared, with input from the public, companies included. Corporate experiments with nonprofit structures, windfall clauses and proposals for Universal Basic Income are well-intended, and show foresight about the inequalities AI will likely produce, but they can’t stand in for democratic choices. That’s part of the problem with over-extending the role of the private sector: companies cannot attempt to both play a role traditionally held by government and also escape the responsibility of being accountable to the public. Representation, in its most basic sense, needs to be real, and that means working through elected officials rather than ceding the government’s role to corporations.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
I’m still working out where the work of building that trust should sit. It’s likely spread across governments, industry, funders, and think tanks or academia. But between bashing big tech, knee-jerk (or expiring) regulation, and letting the market do what it wants, there has to be a place that makes for a better, softer landing. And of course this may all sound soft. But influence in its most effective form is soft power: shaping what people want to do, rather than forcing them.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Clear the fiduciary bar. Giving something away can’t breach the company’s duties, which puts this with the board and investors. The choice is not between values and no values, but between implicit, unexaminable ones and explicit, contestable ones. Shareholder value and fiduciary responsibility can be perfectly good reasons for keeping something private, as can safety, but these constraints need to be well articulated and publicly tested for people to believe you. A short release policy, published alongside each major launch and updated as things change, would go a long way.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Betting on private rails to deliver public goods could read as technosolutionism, but it isn’t. Technosolutionists put their faith in the technology alone. I’m putting mine in the people around it: talented builders with the right incentives, working alongside the communities who know their problems best. That matters most in the domains the market underfunds, like climate, health, and education. Despite AlphaFold, biotech has gone through a funding winter. That only makes sense once you see that venture capital expects returns within 5-7 years or less, a timeline biology rarely keeps to. Breakthroughs will keep arriving on private rails, and leaving their survival to the market is not neutral but negligent. If we can’t engineer equitable delivery, the rest is academic. If we can, it’s the difference between a technology that widens who gets to benefit from progress and one that narrows it to whoever owns the cap table.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Science and engineering can only take us so far. They are the practice of finding the best solution within a given set of constraints. Democracy is something else: the practice of “forming a more perfect union,” or, put more simply, figuring out how we live together. Companies aren’t governments and CEOs aren’t elected, but when a company claims to be building world-altering technology, democratic legitimacy still matters. When people’s will is overridden by decision-makers they can’t or won’t trust, the social and economic fabric starts to fray.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Who gets to define what it means for technology to deliver benefits to society? Benefit is not unilaterally bestowed. As my friends and AI scholars Shakir Mohamed, William Isaac, and Marie-Therese Png have written, it requires the receiving party to independently agree that the shift is indeed beneficial. Asked for a single word to describe AI, Britons most often say “scary” and “worried”, yet technologists tend to ignore that or put it down to a lack of technical understanding.
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.
Even democratic societies struggle with appropriate means of representation, so I concede there is no easy way to close the gap between the public and those building the technology. But for the public to trust technology, they first need to believe someone in the room is looking out for their interests. That starts with going back to first principles about who has the political legitimacy to decide what. Companies decide what is built and released, but elected representatives are the ones with the legitimacy to decide how value should be shared, with input from the public, companies included. Corporate experiments with nonprofit structures, windfall clauses and proposals for Universal Basic Income are well-intended, and show foresight about the inequalities AI will likely produce, but they can’t stand in for democratic choices. That’s part of the problem with over-extending the role of the private sector: companies cannot attempt to both play a role traditionally held by government and also escape the responsibility of being accountable to the public. Representation, in its most basic sense, needs to be real, and that means working through elected officials rather than ceding the government’s role to corporations.
Thinking together
Join the conversation.
For thoughts about the foreword and essay as a whole: what rings true, what are we missing, and what should we read next? To respond to a particular idea, highlight a few words in the passage above.
Commenting uses ChatGPT sign-in.
Loading the conversation…
From an earlier draft
This passage has since changed or been removed. Its original wording and conversation remain here.