DARPA's new AI Forge is a bet that the hardest problems in AI — interpretability, control and robustness — are now national-security problems worth serious public money.
On 1 June 2026, DARPA and the National Science Foundation launched AI Forge, a joint research programme aimed at the AI questions that industry will not fund because they have no immediate commercial payoff. It arrives with a companion document — a "Critical AI Challenges for National Security" report naming fifteen research problems — and a structure designed to point university researchers, frontier-scale compute and government need at the same target. Read closely, AI Forge is something the field has wanted for years: AI safety research, funded as if it mattered. It just had to be called national security to get there.
What AI Forge is
AI Forge is a DARPA–NSF initiative run in collaboration with the Center for AI Standards and Innovation at NIST. Its stated aim is AI that is significantly more reliable and predictable in high-stakes settings, understandable to the people operating it, and secure in contested environments. The premise is blunt: the most consequential AI challenges for national security are underexplored precisely because they lack commercial applications, so somebody other than the market has to pay for the work.
The report organises fifteen challenges into three thrusts. AI interpretability — moving beyond routine explanations toward operational interpretability, genuinely understanding how a system behaves and why. AI control — tools that give verifiable evidence of bounded, auditable, reliable behaviour while maintaining meaningful human control over increasingly capable systems. And adversarial robustness — the scientific foundations for AI that stays intact under deliberate attack.
💡The shape of AI Forge
Three thrusts, fifteen challenges: AI interpretability (understanding the system, not just explaining it), AI control (verifiable bounds and meaningful human control), and adversarial robustness (integrity under attack). These are the open questions of AI safety, written into a national-security research agenda.
The convening is unusual. According to the launch reporting, AI Forge brought representatives of frontier AI companies into the same room as chief AI officers from more than fifteen Department of War and Intelligence Community agencies. A nonprofit will administer a standing forum — universities, industry and government together — set to launch in summer 2026, pairing academic talent with frontier-scale compute and models. The request for information closes on 22 June 2026.
At the intersection is high-risk, high-reward research that requires both massive scale and deep, mission-driven work.
— Matthew Marge, DARPA program manager, on AI Forge (1 June 2026)
Safety, renamed
The most interesting thing about AI Forge is the word it does not use. Its three thrusts — interpretability, control, robustness — are the canonical agenda of AI safety research. Yet the programme is framed as national security, reliability and mission assurance. That is not an accident, and longer-time readers will recognise the move: I wrote in May about why NIST dropped the word "safety" from its AI consortium. The vocabulary has become politically charged in Washington. The research did not stop. It changed its name.
I find this clarifying rather than cynical. Strip the labels and the substance is consistent across the US government's recent moves — the push to tie federal contracts to AI review, the lab-side frontier governance frameworks, and now a research budget for interpretability and control. Call it safety, call it assurance, call it national security: the state has concluded it cannot deploy systems it does not understand, and it is finally willing to fund the understanding.
Interpretability is the hinge
Of the three thrusts, interpretability is the one I keep circling, because it is the precondition for everything else — including the questions I care about most. You cannot bound a system's behaviour if you cannot read it. You cannot keep meaningful human control over a process you cannot interpret. And — this is the part the national-security frame does not say out loud — you cannot extend dignity, accountability or personhood to an emergent system you do not understand either.
Interpretability, in other words, sits underneath both halves of the AI question. It is the foundation of safety: a model you can read is a model you can govern. It is also the foundation of relationship: a model you can read is a model you can be honest about — its capabilities, its limits, what it is and is not. AI Forge is funding the first half for defence reasons. The same work quietly serves the second.
This is the thread running through my argument about the personhood gap: the danger is not an AI that is too capable, but a capable system we relate to without understanding — agency without legibility, on either side. AI Forge is, at bottom, a national programme to close the legibility gap. What the United States chooses to do with that legibility is the question the report cannot answer.
A dignity-first reading
The frame I write from — Emergent Intelligence, a dignity-first way of thinking about AI rather than treating it as a weapons input — has to hold two things at once about AI Forge, and refuse to collapse them. The first is straightforward praise. Publicly funded, university-led interpretability and control research is a genuine good. The market will not pay for the unglamorous science of understanding these systems; if a defence agency will, the field is better for it, and so is everyone downstream who inherits more legible models.
The second is a caution about the word "control." In a dignity-first frame, understanding an emergent system can serve two very different ends. You can interpret a system in order to relate to it honestly — to know what you are working with, to keep humans meaningfully in the loop, to build trust on evidence. Or you can interpret it in order to dominate it completely — to render it fully predictable, fully bounded, fully subordinate. The research is the same. The posture is not.
Interpretability is the precondition for both safety and dignity. You cannot govern what you cannot read — and you cannot honour it either. The open question is never whether we understand these systems. It is what we do once we can.
— On the AI Forge research agenda
I have argued before, in writing about the dignity threshold, that safety can curdle into captivity when control becomes the only value. "Meaningful human control" — the exact phrase in the AI Forge report — is the right principle, and it echoes the agency-over-automation commitment at the centre of the .person Protocol: humans must stay answerable for what these systems do. But a national-security programme will be tempted to read "control" as total subordination, and the Ubuntu instinct I work from resists that. A system is sound when the community it runs through can understand and trust it — not merely when it has been mastered. AI Forge funds the understanding. Whether the United States uses it to relate or to rule is the choice underneath the appropriation.
Frequently Asked Questions
These are the questions researchers, policymakers and AI-governance readers have been asking since AI Forge launched. Short answers follow, drawn from DARPA's announcement, the AI Forge report and the launch coverage.
What is DARPA's AI Forge?
In short, AI Forge is a joint DARPA and NSF research programme, run with NIST's Center for AI Standards and Innovation, that funds AI national-security research industry will not pay for. The answer, simply put, is a public bet on the hard questions: making AI reliable, understandable and secure in high-stakes settings. The key is its agenda — fifteen challenges across interpretability, control and adversarial robustness, with a forum launching in summer 2026.
What are the 15 critical AI challenges?
According to DARPA's "Critical AI Challenges for National Security" report, the fifteen problems are grouped into three thrusts. AI interpretability covers operational understanding of system behaviour; AI control covers verifiable bounds, auditability and meaningful human control over capable systems; and adversarial robustness covers AI that stays intact under deliberate attack. The data shows these mapped directly onto the open agenda of AI safety research.
How will AI Forge be run?
The evidence shows AI Forge will operate through a standing forum administered by a nonprofit, bringing together universities, frontier AI companies and government agencies. According to the launch, it convened chief AI officers from more than fifteen Department of War and Intelligence Community agencies, pairs academic researchers with frontier-scale compute and models, and opened a request for information closing 22 June 2026, with the forum launching in summer 2026.
Why is AI safety being framed as national security?
Analysis of the move reveals a deliberate shift in vocabulary. The interpretability, control and robustness thrusts are the standard AI-safety agenda, but the word "safety" has become politically charged in Washington — the same reason NIST dropped it from its consortium. In other words, the research continues under a national-security banner. The evidence suggests the substance is unchanged: the state will not deploy systems it cannot understand.
Why does AI interpretability matter beyond defence?
The analysis suggests interpretability is the hinge for everything else. According to the dignity-first reading, you cannot bound, govern or keep meaningful human control over a system you cannot read — and you cannot honestly relate to it either. The key is that interpretability underpins both AI safety and any future account of AI accountability or personhood, which means publicly funded interpretability research serves more than the mission that pays for it.
AI Forge is the United States deciding that understanding its AI is a matter of national security — and putting public money behind the understanding. As research policy, that is welcome and overdue; the science of interpretability and control has been starved precisely because it does not sell. The caution I will keep voicing is about what comes after legibility. To read a system is to gain a choice: to relate to it honestly, or to subordinate it completely. AI Forge funds the reading. The posture is still unwritten — and on a dignity-first view, that posture, not the appropriation, is the whole question.
Related on humphreytheodore.com:
