Schmidt Sciences Funded 19 of 501 Humanities AI Proposals. The 3.8 Percent Hit Rate Is the Least Interesting Number in the Announcement.
October 17, 2026 · 6 min read
Granted Research Team · Editorial policy
On October 14, 2026, Schmidt Sciences announced $10 million to 19 projects under its Humanities and AI Virtual Institute (HAVI). An expert advisory board selected them from 501 proposals spanning 42 countries. Combined with the prior round — $11 million announced in late 2025 — HAVI's two-year commitment now exceeds $21 million.
The funded work is genuinely strange in the way good humanities computing is strange. One team is using AI to reconstruct the Forma Urbis Romae, the marble map of Rome that shattered in 226 AD. A University of Michigan-led international team received more than $340,000 to develop methods for unlocking the guqin canon — over 30,000 surviving compositions in a tablature notation that encodes finger position rather than pitch. Another project, Humanitext, is building a verified knowledge graph of quotations, allusions, paraphrases, and criticisms running between Greek and Latin texts over three years.
The 3.8 percent selection rate is the number that will get quoted. It is also the number least useful to anyone deciding whether to apply next time.
Why the hit rate misleads
A 3.8 percent acceptance rate sounds like NIH R01 territory in a bad year, and it invites the conclusion that HAVI is effectively a lottery. That reading misses two structural facts.
First, 501 proposals from 42 countries is an open global call with no institutional eligibility filter. Compare that to a typical NEH Collaborative Research competition, which is restricted to U.S. nonprofit institutions and cultural organizations and draws a few hundred applications from a professionalized applicant pool that already knows the rules. A global open call with no eligibility gate collects a large volume of proposals from people who have never written a competitive research proposal. The effective rate for a serious, well-scoped submission from an experienced team is meaningfully higher than 3.8 percent — nobody publishes that number, but every program officer knows it exists.
Second, HAVI is not a single-tier program. Earlier in 2026, HAVI separately funded 11 Development Awards — explicitly short-term projects to "demonstrate how new technical approaches can be successfully implemented at scale." Those included Georgetown's work tracing enslaved people from archival records to the U.S. Census, Hebrew University's dating of Akkadian literary texts from small data, Cardiff University's adaptation of handwritten text recognition for colonial Paraguayan archives, Ohio State and the Royal Central School of Speech and Drama on motion modeling with dance expertise, and Carnegie Mellon's ArchAIa foundation models for artifactual understanding.
A two-tier structure where a smaller, shorter, lower-risk award exists alongside the flagship is a feeder, whether or not the funder describes it that way. The development tier is where a program builds confidence in a team's technical competence and in whether a method scales. Applicants who treat the development awards as a consolation prize are reading the structure backward.
The open-source mandate is the real eligibility test
Buried in the 2026 announcement is the sentence that should reorganize your proposal: teams will openly share tools, code, model weights, and benchmark datasets.
Model weights. Not just code. Not just a data availability statement. Weights.
This is a stronger requirement than almost anything in federal humanities funding, and it has consequences that cascade through a project budget and a team roster:
| Requirement | What it forces into the proposal |
|---|---|
| Open code | A software engineer or RSE line, not a grad student who "also codes" |
| Open model weights | Clean provenance on every training corpus — you cannot release weights trained on material you only have reading rights to |
| Open benchmark datasets | A benchmark design task treated as a deliverable, not a byproduct |
| Open tools | Documentation, packaging, and a maintenance commitment past the award period |
The provenance point is the one that kills otherwise strong humanities AI proposals. Humanities research runs on licensed corpora: subscription text databases, museum image rights, archival reproductions with restrictive terms, publisher-controlled critical editions. A team can do excellent work on licensed material and be structurally unable to release weights derived from it.
If you are planning a HAVI submission, the rights audit is not a compliance afterthought. It is a design constraint that should determine which corpus you choose. The funded projects skew noticeably toward material that is public domain, institutionally owned by a team member, or being newly created by the project — ancient texts, archaeological artifacts, government archival records, pre-1929 periodicals. That is not a coincidence.
Philanthropy is now the load-bearing funder here
The context nobody in the announcement says out loud: U.S. federal humanities funding has been in sustained disruption. The National Endowment for the Humanities has been through terminations litigation, a rescission of Title VI disparate-impact regulations affecting arts and humanities grantmaking, and competitions reshaped around a narrow set of mandated themes. NEH's Collaborative Research line — the closest federal analogue to a HAVI award — ran a cycle with a roughly 11 percent funding rate and a restriction toward American history and Western civilization topics.
Against that, a private funder putting $21 million over two years into globally distributed, methodologically experimental humanities work is not a supplement. For a lot of the 19 funded teams, it is the only mechanism that would have funded that specific project.
This creates a dependency worth naming. Philanthropic research funding is faster, less bureaucratic, and more willing to fund weirdness than federal funding — and it is also discretionary, unappropriated, and subject to a single donor family's continued interest. HAVI has run two rounds. It has not announced a third, and the announcement contains no dates or commitments for future calls. Building a lab's five-year plan on a program with no published cadence is a risk that should be priced into how you structure the work: fund the method through HAVI, and position the method so that it becomes fundable through NEH, NSF, ERC, or a national research council once it exists.
What to do if you want to be in the next cohort
Pick your corpus for releasability first, research question second. This feels backward to a humanist and it is the single highest-leverage change you can make. A brilliant question applied to a corpus you cannot redistribute produces a proposal that fails the program's core commitment.
Build the international team early. Forty-two countries submitted; the funded list spans Carnegie Mellon, Chicago, Northwestern, UC Berkeley, Oxford, Vanderbilt, Tec de Monterrey, Lancaster, Cardiff, Hebrew University, and Reading. Cross-national collaboration is visibly rewarded, and unlike federal programs there is no domestic-lead requirement forcing you to contort the partnership.
Propose a benchmark, not just a result. The projects that recur across both HAVI rounds share a shape: they produce an artifact other researchers can use — a knowledge graph, a recognition model for a specific script, a benchmark dataset for a task that did not previously have one. "We will answer question X about corpus Y" is a research paper. "We will build the first benchmark for task Z, answer X with it, and release everything" is a HAVI proposal.
Consider the development tier deliberately. If your method is unproven at scale, a short development project that demonstrates it is a stronger path to a flagship award than a flagship application that asks reviewers to take the scaling on faith.
Watch the smapply portal rather than waiting for a press release. HAVI's Development Awards call has run through Schmidt Sciences' SM Apply instance. Programs that announce awards in October frequently open the next call within a few months, and open calls from private funders routinely carry shorter windows than federal solicitations — six to ten weeks is common. A team that starts assembling after the call opens is usually too late.
The quieter implication
Nineteen funded projects across art, dance, history, music, literature, archaeology, and ancient games is a portfolio bet that AI methods transfer across humanistic domains. If that bet pays off, the artifacts these teams release — the weights, the benchmarks, the recognition models for specific scripts and notations — become shared infrastructure, and the next round of humanities AI work gets cheaper for everyone.
That is the actual argument for the open-release mandate, and it is why a funder running a 3.8 percent selection rate can still be increasing the total amount of work that gets done. The 482 teams that did not get funded this round will, in eighteen months, have tools they did not have before. Some of them will be competitive precisely because of what this cohort released.