Schmidt Sciences' Trustworthy AI RFP: The May 17 Deadline, the Two-Tier Structure, and Why Tier 2 Is the Real Competition

Three days from now, on May 17, 2026 at 11:59 PM Anywhere on Earth, Schmidt Sciences will close its 2026 Science of Trustworthy AI Request for Proposals — a funding competition that has quietly become one of the largest private sources of frontier-AI safety research funding outside the major industrial labs. The RFP opened February 18 and offers two tiers: Tier 1 awards of up to $1 million over one to three years, and Tier 2 awards of $1–5 million or more over the same period. Decisions land in summer 2026. Indirect costs are capped at 10%. Applicants may submit multiple proposals to either tier.

The RFP's structure is unusual enough that it deserves a careful read. Most foundation-funded AI research RFPs operate at the $200K–$500K range and treat the funding as a hedge — a small bet across many promising teams. Schmidt Sciences' Tier 2 ceiling is high enough to fund a multi-year, multi-postdoc effort at a single PI's lab or a multi-institution collaboration, and the explicit allowance of multi-institution and international teams pushes the competition toward consortium-style proposals rather than single-investigator submissions. The 10% indirect cost cap is aggressive — most US R1 universities negotiate federally with indirect rates in the 55–70% range — and reshapes which institutions can credibly host the work.

This deep dive walks through what each of the three research aims actually rewards, why Tier 2 is structurally more competitive than Tier 1, how the 10% indirect cap reshapes the institutional landscape, and the practical decision framework for whether a research team should push for a May 17 submission or wait for the next cycle.

The Three Aims and What They Actually Reward

Schmidt Sciences names three interconnected research aims that proposals must address. Reviewers will weight an explicit and substantive engagement with one or more of these far more heavily than a vague gesture toward "AI safety."

Aim 1 — understanding why frontier AI systems learn misaligned goals that fail under distribution shifts or extended interaction — is the most theoretically loaded of the three. Proposals here need to engage with the now-substantial body of empirical work on goal misgeneralization, deceptive alignment, and reward hacking, and articulate either a sharper theoretical framework or a more rigorous empirical methodology than what already exists. Reviewers will be skeptical of proposals that describe a research program already underway at major industrial labs without naming a specific differentiator — methodological, computational, or institutional — that justifies independent funding. The strongest Aim 1 proposals typically come from research groups that have already published in the area and can credibly argue that academic freedom or computational independence enables work that industrial labs cannot or will not pursue.

Aim 2 — creating evaluations with genuine predictive validity and developing interventions that influence AI system learning — is the most pragmatically valuable aim and likely the most competitive. The phrase "predictive validity" is doing real work here: Schmidt Sciences is signaling that they are tired of evaluations that produce numbers without behavioral generalization. A successful Aim 2 proposal needs to specify what counts as a predictive evaluation in their framework, how they will measure validity beyond benchmark scores, and what intervention or training technique their evaluation will inform. Teams that combine evaluation methodology with intervention design will outcompete teams that propose either alone.

Aim 3 — extending human oversight to superhuman AI capabilities and addressing multi-agent interaction risks — is the most speculative aim and the most difficult to scope credibly within a three-year window. Proposals here need to engage with current scalable-oversight research (debate, recursive reward modeling, weak-to-strong generalization) without naively claiming to "solve" the problem. The strongest Aim 3 proposals typically propose a concrete intermediate capability — for example, validated protocols for human oversight of multi-step agentic tasks in a specific domain — rather than an abstract framework for oversight of arbitrary superhuman systems.

The RFP explicitly permits proposals that address multiple aims, and Tier 2 proposals are effectively expected to do so. A proposal that engages with Aim 1's theoretical foundations, develops Aim 2's predictive evaluations, and uses both to inform an Aim 3 oversight intervention is structurally more compelling at the higher funding level than a single-aim proposal.

Why Tier 2 Is Structurally More Competitive Than Tier 1

The natural assumption is that more money means stiffer competition. That is true in absolute terms — Tier 2 will see fewer awards than Tier 1 — but the competitive math is more nuanced.

Tier 1 is structured for individual investigators and small teams running focused research programs. The funding level — up to $1M over three years — comfortably supports one or two postdocs, modest compute, and PI summer salary. The applicant pool is large because many academic AI safety researchers fit this template, and Schmidt Sciences is likely to fund a sizable cohort to maintain field-building momentum.

Tier 2 is structured for consortia, multi-postdoc labs, and projects that require substantial compute or experimental infrastructure. The funding level — $1–5M or more over three years — supports a small research center or a multi-institution collaboration. The applicant pool is smaller because fewer teams can credibly absorb that level of funding, but the competition is fiercer because each applicant has typically assembled a strong consortium specifically for this RFP. A weak Tier 2 proposal will not be downgraded to Tier 1 funding; it will be rejected.

Two implications follow. First, a research group that fits comfortably within a Tier 1 budget should apply to Tier 1, not stretch to Tier 2. Reviewers are likely to penalize Tier 2 proposals that read as inflated Tier 1 budgets without commensurate scope expansion. Second, a research group that could credibly assemble a multi-institution Tier 2 consortium should not split its proposal into multiple Tier 1 submissions. The multi-institution structure is precisely what Tier 2 is designed to fund.

The 10% Indirect Cost Cap Reshapes the Institutional Landscape

For research administrators reading the RFP, the 10% indirect cost cap is the single most consequential constraint. Most US R1 universities negotiate federal indirect cost rates between 55% and 70% of modified total direct costs. A 10% cap means that a $1M Tier 1 award delivers roughly $90,000 less to the host institution's overhead recovery than the same award at the institution's federal rate, and a $5M Tier 2 award delivers roughly $2.7M less.

Two consequences shape who can credibly host this work.

First, institutions with explicit policies allowing indirect cost waivers or reduced rates for private foundation funding will be advantaged. Some major research universities have streamlined approval processes for reduced indirect rates from named foundations; others require case-by-case approval that can take weeks. Research teams at the latter institutions need to start the indirect-cost approval process immediately if they are submitting Friday — the proposal cannot move forward without institutional sign-off on the reduced rate.

Second, independent research institutes, national laboratories, and nonprofit research organizations frequently operate with lower indirect rates than R1 universities and may be structurally better hosts for this funding. Organizations like the Mila institute, the Allen Institute for AI, the Redwood Research-style independent safety labs, and similar groups can absorb Schmidt Sciences awards more efficiently than a traditional university research office. For Tier 2 proposals, hosting at an independent institute with a university subcontract for specific PI effort can be a more competitive structure than primary hosting at a university.

International applicants face their own institutional considerations: many European universities and Canadian institutions have indirect cost structures more compatible with 10% than US R1s, but currency considerations, IRB and ethics approvals across jurisdictions, and the practical logistics of multi-postdoc hiring on non-US visas need to be addressed in the proposal's management plan.

The May 17 Submission Decision

For research teams looking at the three-day window, the submission decision turns on three factors.

Concept maturity. A strong Schmidt Sciences proposal typically requires 60–100 hours of focused PI and senior-postdoc writing time, plus institutional administrative work for budget and indirect-cost approval. Teams that have been refining their concept since the February 18 RFP release are well positioned; teams that started concept work after May 1 are racing against the institutional approval timeline more than the writing timeline.

Reviewer alignment. The reviewer pool for this RFP draws heavily from the existing trustworthy-AI research community — academics, industrial researchers, and independent safety researchers who know each other's work. Proposals that engage substantively with the recent published literature, name specific prior results, and articulate clear methodological differentiators will outperform proposals that present the field at a high level. Teams that are not already embedded in this research community face a credibility climb that is difficult to close in three days.

Pipeline planning. Schmidt Sciences runs the Science of Trustworthy AI program annually, with the 2027 cycle likely to open in early 2027 on a similar schedule. Teams that recognize their proposal is not Friday-ready should plan toward the 2027 cycle now — refining their research program, publishing intermediate results, building consortium relationships, and engaging with the broader trustworthy-AI research community over the next nine months — rather than submitting a rushed proposal that will not advance.

Beyond Friday: The Schmidt Sciences Funding Position

For research teams thinking about their long-term funding strategy, the Schmidt Sciences trustworthy-AI program is structurally important beyond any individual award. It is one of the few sources of multi-million-dollar private funding for academic AI safety research that does not come with the IP, publication, or research-direction strings attached to industrial-lab partnerships. The 10% indirect cap is aggressive but transparent; the application process is rigorous but not bureaucratically opaque; the program officers are technically credible and engage substantively with the research.

Teams building a trustworthy-AI research program should treat Schmidt Sciences as a recurring funding source, not a one-time opportunity. The May 17 cycle will fund some teams and reject many strong ones. The 2027 cycle is nine months away. Either way, the structural shift in private AI safety funding — toward larger awards, more consortium structures, and tighter coupling between evaluation and intervention — is the trend academic researchers should be planning around.