The $200K Corrigibility Research Fund: A Small, Fast AI-Safety Grant With an August 23 Deadline and No Application Form

Most of the AI-safety funding Granted covers arrives in large, structured packages: Schmidt Sciences' multi-agent safety program with its $1M ceilings, ARIA's £100M trust-infrastructure push, Coefficient Giving's technical-safety RFPs. Those are the battleships. The Corrigibility Research Fund is something different — a fast, small, deliberately low-friction grant program that will distribute at least $200,000 across 2026, in checks that mostly land between $5,000 and $35,000, and where you apply by sending a single email. For an independent researcher, a graduate student, or a small team with a sharp idea and no institutional grants office, this is exactly the kind of instrument that is easy to miss and disproportionately worth catching. Round 1 closes August 23, 2026.

This is the deep dive on what the fund is, what it is trying to buy, and how to approach it — because the mechanics here are unusual enough that treating it like a normal grant will cause you to over-prepare the wrong things.

What "corrigibility" actually means

Start with the concept, because the word does a lot of work and the fund's definition is precise. The Corrigibility Research Fund defines corrigibility as "the property of an AI agent that keeps its human principal informed and in control — one that helps us notice and fix the flaws in its thoughts, structure, and actions." In plainer terms: a corrigible AI is one that keeps humans in the driver's seat, that does not resist being corrected, and that actively helps its operators catch and repair its own mistakes.

The fund is managed by Max Harms, author of "CAST: Corrigibility as Singular Target," and administered by Lightcone Infrastructure Inc., a 501(c)(3) nonprofit that retains final approval on grants and prizes. The intellectual bet is specific: Harms argues that corrigibility offers a kind of robustness that direct value-alignment does not. "A purely-corrigible agent can be expected to avoid scheming," as he frames it. Where value-alignment tries to get the AI's goals exactly right the first time — a target that is unforgiving if you miss — corrigibility is designed to be an iterative property that accommodates human error and leaves room to fix things as you go. That distinction is the entire thesis, and it is why a dedicated fund exists rather than folding this work into a general alignment pool.

The unusual structure: grants and prizes, two clocks

The fund runs on two parallel tracks, and understanding the split is the key to using it well.

Grants (~$100K of the pool) are traditional, apply-in-advance funding for prospective work. This is money for research you intend to do. The application is deliberately lightweight — a short email describing the project and the funding need. Typical grants run $5,000 to $35,000. There are two rounds:

• Round 1 deadline: August 23, 2026
• Round 2 deadline: October 31, 2026

Prizes (~$100K of the pool) are the mirror image: retroactive awards recognizing excellent corrigibility research produced during 2026. There is no application for prizes. If you did strong, relevant work this year, you are eligible to be recognized whether or not you ever contacted the fund. Prize announcements come in two waves — roughly $40,000 at the end of September and $60,000+ in mid-December 2026.

That two-track design changes the strategic calculus. If you have work already underway or completed this year, you do not need to wait for a deadline or fill out anything — the prize track will consider it. If you have a plan and need capital to execute, the grant track is your lane, and Round 1's August 23 date is the near-term action item. Most researchers should think about both: apply for a grant to fund the next phase, and let the prize track reward the phase you already shipped.

What they will and won't fund

The scope statement is refreshingly concrete about both directions. The fund supports work that "predictably advances humanity's understanding" of corrigibility, spanning the full range from formal theory to empirical training experiments. It wants research that is "legible and relevant to people making decisions about real systems" — meaning results a lab or a policymaker could actually act on, not purely abstract exercises.

There is one hard constraint worth internalizing: the work must avoid accelerating AI capabilities. This is a safety fund, and research whose main effect is to make frontier systems more powerful is out of scope by design. If your proposal's honest summary is "this makes models better at X," reframe or reconsider — the fund is buying control and legibility, not capability.

Within those bounds, the range is wide. Formal work on what corrigibility means and whether it is coherent; training experiments that test whether you can instill it; interpretability work that lets operators see when an agent is drifting toward incorrigibility; evaluations and benchmarks that make the property measurable. The unifying test is Harms's phrase — does it help us "get everything right the first time" by helping us notice and correct our mistakes?

Why a $200K fund deserves a deep dive

It is fair to ask why a fund this size warrants the same attention as a nine-figure program. Three reasons.

First, the check-to-effort ratio is extraordinary. A $5,000–$35,000 grant that you apply for with one email is one of the highest expected-value uses of an afternoon in the entire safety-funding landscape. There is no budget-office overhead, no 15-page narrative, no institutional cost-share. For an independent researcher, that low friction is the whole point — it removes the administrative tax that keeps small, good ideas from ever getting funded.

Second, small dedicated funds are how niche research areas get born. Corrigibility is not yet a large, well-funded subfield with its own conference track and standing NSF program. Funds like this one are the seed capital that turns a research thesis into a community. Being an early grantee or prize-winner in a nascent area is a positioning move — it establishes you in a space before it is crowded, which is exactly the kind of first-mover advantage that pays off when the larger funders eventually build programs here.

Third, the prize track is free optionality. Because prizes require no application and consider all 2026 work, any researcher already active in corrigibility has a live, no-cost shot at recognition and money simply by doing good work and making it public. The rational move is to make sure your 2026 output is legible and discoverable — a clear writeup, posted where the fund's judges will see it — so it is in the consideration set when the September and December awards are decided.

How to actually approach it

If you want a grant, act before August 23. Email grants@corrigibilityresearch.org with a short, concrete description: what you will do, why it advances corrigibility understanding, why it is legible to real-world decision-makers, and how much you need. Given the $5K–$35K band, scope the ask realistically — this funds a focused piece of work or a phase, not a multi-year lab. Missing Round 1 is not fatal; Round 2 closes October 31. But August 23 is the first and best shot.

If you have 2026 work, make it visible now. The prize track rewards what already exists. Publish your writeup, make it findable, and frame it explicitly in corrigibility terms so the connection is obvious to a judge scanning the field. The September wave (~$40K) is decided soon; the December wave ($60K+) rewards the full year.

Frame everything as control, not capability. The one disqualifier is capability acceleration. Lead with how your work helps humans stay informed and in control, and how a lab or policymaker could use the result. That framing is both the fund's stated priority and the honest description of what safety research is for.

The Corrigibility Research Fund will not move the macro numbers on AI-safety spending — $200,000 is a rounding error next to the battleships. But it is a clean, fast, low-friction instrument aimed at a specific and underfunded idea, run by people who have thought hard about that idea, with a live deadline this month. For the right researcher, one email before August 23 is among the best-leveraged grant applications available anywhere right now.