By Kyle Demes and Rohan Mantena
Every funder we work with wants the same thing: a list of the papers its money produced. They need it to report to governments and donors, to evaluate their programs, and to decide what to fund next. And sooner or later, nearly all of them ask us the same question: where does the funding information on a paper actually come from?
The answer surprises people. For most papers, the funders in the DOI record weren’t entered by the authors. Software in the publisher’s production system inferred them from the acknowledgements section, and the result was deposited as fact.
This post is about how that happened, why asking for it made sense at the time, and why we think it’s the wrong approach now. Our proposal is simpler: publishers share the acknowledgements section itself, as open metadata, with every DOI, and the links get made from the text.
Where funding metadata comes from today
Outside of our own work, funding metadata reaches open infrastructure mostly through the DOI registration agencies, and the route differs by the kind of output. (OpenAlex now reads funders and grant numbers from the full text of papers, and by volume that supplies most of our funding links; we described that project here. It isn’t the subject of this post.)
Datasets. When researchers deposit a dataset in a repository like Zenodo, they can enter its funders and grant numbers themselves, and the repository sends them to DataCite. We wrote about this in July (Who funded this dataset? Let’s ask DataCite.): 716,297 datasets carry a funder link, and 97% of those carry a grant number. Asking the people who did the work, at the moment they submit it, is a good way to get accurate structured metadata. For datasets it’s also the only way, because a dataset has no acknowledgements section to read.
Articles. Years ago, many of us in the open-metadata community, ourselves included, started asking publishers to deposit structured funding metadata with their Crossref records: an identifier for each funder and a number for each grant. Funders and Crossref asked too, most recently through the Barcelona Declaration’s funding-metadata Call to Action. It made sense at the time. We wanted funding data that machines could use, and we assumed that structured would also mean clean and accurate. Publishers responded: about a quarter of the journal articles registered with Crossref in 2024 to 2026 now carry at least one funder (25.9% of 19.0 million).
But how they responded matters, and it isn’t quite what we expected. Some submission systems ask authors to pick their funders from a list as they submit (putting the onus on authors, upstream of publishing), the same good path datasets take. For most papers, though, the structured funding metadata is created later, in production. Third-party software reads the funding section of the paper, tries to match the names it finds to a registry of funders, extracts grant numbers, and sends those funders and grant numbers to Crossref as structured metadata.
Nothing in the record says it was inferred, so it reads like a source of truth. But the text it was inferred from isn’t deposited in Crossref at all, so nobody downstream can verify it. We just have to trust that those assertions are true. We asked for clean, structured data. The way publishers delivered it was software that infers it, and everyone has had to treat those inferences as the source of truth.
Now that OpenAlex extracts funder names and grant numbers from the full text of open access papers ourselves, we can compare them with the Crossref records for the same papers, and see what these otherwise invisible systems get wrong.
What goes wrong
Reading funders out of free text is hard, and it fails in predictable ways. Every example below was checked in Crossref this week:
- Same name, different funder. A paper thanks India’s Anusandhan National Research Foundation (ANRF) for grant ANRF/IRG/2024/000620/LS. Its Crossref record credits the US Arthritis National Research Foundation, with the same grant number (10.1016/j.matt.2025.102257).
- A paper funded by Taiwan’s National Science and Technology Council (NSTC 114-2221-E-006-050-MY3) is credited to Zambia’s NSTC (10.1021/acsomega.5c10463).
- A US National Institutes of Health grant (K08 HD116931) is filed under “NIH”, the Norwegian School of Sport Sciences (10.1210/jcemcr/luag301).
- Four US National Science Foundation grants are credited to the National Science Foundation of Sri Lanka (10.1128/mbio.03793-25).
- A grant from the National Research Foundation of Korea is credited to South Africa’s National Research Foundation (10.1080/00036811.2025.2491643).
- The wrong section. A paper that says it “received no external funding” lists five companies as funders, taken from the authors’ competing-interest disclosures (10.1093/ehjopen/oeag004).
- The wrong funder entirely. A paper acknowledging grants from France’s ANR, Germany’s DFG and the US NSF has all three filed under a university in Indonesia (10.1145/3725245).
These examples come from seven different publishers, and the problem isn’t any one publisher’s. We see it across all the major publishers: every one of the 15 that deposit the most funder metadata has errors our checks catch, though at very different rates.
None of this is unique to publishers. On another paper thanking India’s ANRF, the publisher deposited only the name “ANRF” and Crossref’s own matcher picked the American foundation (10.1007/s11139-025-01125-3). These mistakes are easy for anyone to make. What matters is being able to catch them, and that takes the raw acknowledgements. We can measure our own errors because we have the text, and we rebuild our matcher every time we learn from them; right now we’re rebuilding it to read only the funding and acknowledgement sections. That only works when the text is open.
These aren’t rare quirks. They run to millions of papers:
- Crossref holds funder metadata for about 16 million works. In two samples where we checked every funder entry against the paper’s own text, about 9 to 13% of papers carried at least one funder that was clearly wrong. If that holds across Crossref, roughly 1.5 to 2 million papers carry at least one wrong funder entry.
- The gaps are bigger. In one of those samples, at least a quarter of papers (444 of 1,792) were also missing funders their own funding statements named. Others have found the same from the other side: about 30% of funding acknowledgements in full text never reach Crossref at all (de Jonge and Rieck 2026).
- Some funder identifiers are mostly errors. Of 12,759 deposits crediting Zambia’s NSTC, 12,609 belong to someone else. Of 1,501 papers crediting the Arthritis National Research Foundation, 1,151 don’t.
- Even our narrowest check, which only catches errors that leave a trace (a name two funders share, or a grant number that belongs to another country’s funder), finds 62,154 wrong funder entries on journal articles from 2024 to 2026 alone. 98% of them were deposited by publishers, not matched by Crossref.
Why this matters
A Crossref record looks like a source of truth. It’s structured, it carries persistent identifiers, and it comes from the publisher. So everyone downstream treats it that way: OpenAlex, commercial indexes, funder dashboards, research offices, and whoever is writing next year’s impact report. A wrong guess made once spreads everywhere. A funder can find its name on a paper it never funded, sometimes one with integrity problems, with no way to see where the claim came from or to correct it.
The gaps cost as much as the errors. A funder counting its outputs comes up short, and can’t tell which papers are missing.
We can’t fix this from our side. For open access papers we read the acknowledgements ourselves: OpenAlex has acknowledgement or funding sections for about 15.7 million works, and we use them to add funders the DOI record missed, and we’re starting to remove ones the text doesn’t support. That only works for papers we can read. For closed-access papers, the DOI record is the only funding information in the open, so its errors and its gaps stay. Funders ask us all the time what we do about closed-access papers, and we’ve never had a good answer beyond “trust Crossref”, which has funders for only about a quarter of recent articles.
What’s changed
When the community first asked for structured funding metadata, turning a sentence of free text into funder identifiers was hard, and doing it once, at the source, looked better than everyone downstream doing it badly. That’s no longer true. Software can now read an acknowledgements section and pick out who paid for what with high accuracy: Crossref’s new matcher reports precision of 0.99 and recall of 0.81 (and approaches built on large language models can do better still), and the methods keep improving. With the raw text, anyone can make cleaner links than the structured metadata we’re handed today, check them, and redo them when the methods get better. Without the text, we’re stuck with whatever was guessed at deposit.
So we think the right approach has flipped. Asking every publisher to structure its funding metadata made sense then. Now it’s easier for everyone, more accurate, and better for the open community if publishers share the raw acknowledgements and the links are made from them. It’s better for publishers too: they no longer have to build, buy and maintain software to infer funders.
We could keep the current approach and make the assertions accountable instead: a shared standard for how funders are extracted and matched, a label on each assertion saying how it was made, regular checks on what vendors deposit, and a commitment to correct records when errors are reported. The Barcelona Declaration’s Call to Action already asks publishers to monitor third-party performance. That can work, but it’s slow and expensive, every publisher has to repeat it, and the result is still a set of inferences nobody else can check.
Sharing the text is simpler and better:
- Publishers already have the text. JATS XML has elements for both (
<funding-statement>and<ack>). Depositing them needs no matching, no registry lookups and no vendor audits. They can keep depositing structured funders too, but those become what they really are: inferences from that text that anyone can recompute and check. - Anyone can do the matching. Crossref, OpenAlex, funders, national research information systems and researchers can all run their own funder matching against the same text. With the text open, Crossref’s matcher could run on every paper, not just the ones that came with a funder name.
- Open benchmarks become possible. Today nobody can measure how well any funder-matching system works across the literature, because the ground truth is closed. Open text lets the community build shared test sets and compare methods in public, ours included.
- Search gets far more useful. Acknowledgements name much more than funders: core facilities, field stations, research infrastructure, programs, collaborators and communities. A research office asked us recently whether they could search acknowledgements for papers that thanked one of their funding programs by name. For most papers the answer today is no. Open acknowledgements would also let anyone search for the people a paper thanks by name.
- Errors become fixable. With the text and the inference side by side, anyone can spot a mismatch and report it.
- Closed access stops being a blind spot. Funders could finally see every paper that acknowledges them, whoever can read the full text.
We’ve seen a smaller version of this already. Among papers that authors also deposited themselves in Zenodo, entering funders by hand, 2% of the author-entered records had a clearly wrong funder, against 13% of the same papers’ publisher records. When the people who know the answer write it down, the data is better. The acknowledgements section is where authors already write it down.
None of this is a new idea, and the pieces are already moving. The Barcelona Declaration’s funding-metadata working group recommends letting publishers “deposit raw funding information from the acknowledgment section alongside structured funding metadata”. Crossref has announced that its next major schema will accept free-text statements, funding acknowledgements among them. What’s missing is the commitment that makes it the default: publishers filling that field for every paper, and funders expecting them to.
What we’re asking for
- Publishers: deposit the funding statement, and ideally the whole acknowledgements section, with every DOI. You don’t need to build or buy software to infer structured funders as well; with the text open, the community can make those links.
- Crossref and DataCite: Crossref’s planned statements are the right home for this text, and we’ll help test them. DataCite’s funding fields have no place for free text yet; a matching field would bring datasets and repositories along.
- Funders: ask for it. You asked publishers for funder IDs and they delivered. You can ask for the text the same way. The Call to Action already asks funders to consider making standard acknowledgement wording a grant condition; asking for that wording to be deposited openly goes naturally with it.
- The Barcelona Declaration community: take this up as a concrete next step for open funding metadata.
For our part, OpenAlex will use deposited acknowledgement text in our funder matching as soon as it exists, and we’re writing up the full analysis behind these numbers as a preprint. If you think we’ve got something wrong, tell us.
Who funded this paper? The authors already told us. Let’s share what they said.











