What other fields suggest about curation

Agents arrived in other kinds of expert work before they arrived in biocuration. Radiologists, systematic reviewers, and software developers have been through some of what curators are starting now, and there is evidence about what happened to them. This page collects that evidence and asks what it might mean for biocuration.

It does not settle the question. Biocuration has only its first results, and the other fields differ from it in ways that matter. The companion page, How agents changed software development, covers software in more detail.

Radiology: the replacement that did not happen

In 2016 Geoffrey Hinton said people should stop training radiologists, because AI would read images better within five years. The models did get very good at reading images. More than 700 radiology AI models have FDA clearance.

Radiologists did not go away. US radiology residencies offered a record 1,208 positions in 2025, vacancy rates are at all-time highs, and average pay is about 48% higher than in 2015 (Works in Progress, 2025). The UK is short of 32% of the radiology consultants it needs (Royal College of Radiologists, 2025).

Three reasons, each with a parallel in curation:

Reading images is about a third of the job. Radiologists spend about 36% of their time on direct interpretation. The rest is consulting with clinicians, procedures, protocols, and teaching. Assigning a GO term is likewise a fraction of what curators do: modelling hard cases, settling disputes, setting standards, training others, and deciding what the knowledge base should say at all.
Each model answers one narrow question. Real work joins many such questions together, and the joins are where judgment lives.
Cheaper reading created more reading. When a task gets cheaper, demand for it rises. Curation has a backlog measured in decades, so there is no shortage of demand.

The lesson is not "AI will not change the job". It changed radiology a great deal. The lesson is that tasks get automated and jobs get rebuilt around what is left. This was the argument put to curators at Biocuration 2026, setting radiology ("AI performs first pass interpretation" / "Demand for radiologists is higher than ever") beside software engineering, and concluding that the curator's role "will involve doing higher level tasks, and more managing and organization" (Mungall, Staying in the Loop, February 2026).

Second eyes: the mammography result

In European breast cancer screening every mammogram is read independently by two radiologists. This "double reading" is the standard of care because a second reader catches cancers the first one misses. It is also expensive, which is why most countries cannot afford it. Most curation is in the same place: an annotation is made by one curator and is seen again only if someone happens to look.

The Swedish MASAI trial randomised 105,934 women between standard double reading and AI-supported screening. In the AI arm, the AI triaged low-risk exams to a single reader and flagged suspicious areas for the readers. The results, published in The Lancet in January 2026:

	Two radiologists	AI-supported
Sensitivity	73.8%	80.5%
Specificity	98.5%	98.5%
Interval cancers per 1,000	1.76	1.55 (non-inferior)
Screen readings	109,692	61,248 (44% fewer)

(Gommers et al., Lancet, 2026; Hernström et al., Lancet Digital Health, 2025)

In a randomised trial, then, AI support meant less expert time and better quality at once. Germany's nationwide rollout found 17.6% more cancers detected with no increase in recalls (Eisemann et al., Nature Medicine, 2025).

Two caveats. Pooled across the three large European studies, the gain in detection is real but modest, about one more cancer per thousand women (Ferre et al., 2026). Screening was already highly optimised, with two expert readers on every exam. Most curation has one, which raises the question of whether the room for improvement is larger.

Second eyes: evidence synthesis

Systematic reviews have their own version of double reading: two people screen every paper independently. LLM screening now reaches a pooled sensitivity of 0.92 and specificity of 0.94, with workload reductions of 50% to 99% (Xie et al., J Evid Based Med, 2026). Cochrane and its partner organisations permit it, with a condition: "AI and automation in evidence synthesis should be used with human oversight," and the authors remain responsible (joint position statement, 2025). Screening papers for relevance is close to the literature triage curators do.

The costs of keeping a human in the loop

"Human in the loop" is often said as if it settles the matter. It does not. The human is a component with known failure modes.

People defer to confident machines, including experts. When an AI gave deliberately wrong answers on mammograms, the most experienced radiologists' accuracy fell from 82% to 46% (Dratsch et al., Radiology, 2023). Physicians given planted errors in LLM advice dropped from 90.5% to 76.1% diagnostic accuracy (a 2025 preprint). The human factors literature has found for decades that this kind of complacency "cannot be prevented by training or instructions" (Parasuraman and Manzey, 2010).

Skills fade when they are not used. After AI-assisted colonoscopy was introduced in four Polish centres, endoscopists' detection rate on procedures done without AI fell from 28.4% to 22.4% (Budzyń et al., Lancet Gastroenterology & Hepatology, 2025). The study is observational and has been debated, but the direction matches what software studies find for junior developers.

Only the most expert reviewers catch the subtle errors. In the DRAGON-AI evaluation of AI-written ontology definitions, the GO Consortium noted that the AI "would give plausible yet subtly incorrect term definitions for some terms, and … this was only noticed by more expert curators" (GO Consortium, Nucleic Acids Research, 2026). Expertise is not made redundant by review. Review is where it is needed most.

Human plus AI is not automatically better than either. A meta-analysis of 106 experiments found human–AI teams did worse than the better of the two alone on decision tasks, and better on creation tasks (Vaccaro et al., Nature Human Behaviour, 2024). Curation has both kinds of task, and the combination has to be designed for each.

Supervision is tiring. See Supervision is tiring, and it has a limit.

So asking curators to be "hypervigilant" does not scale. Vigilance decays. The practical response is to spend vigilance where it pays: let deterministic checks handle what can be checked mechanically, and direct human attention to the items most likely to be wrong.

AI reviewing AI

AI review is useful as a filter. It is not an independent second opinion, for three reasons:

Models make the same mistakes. Across more than 350 models, when two models are both wrong they agree on the wrong answer 60% of the time, and more capable models are more correlated (Kim et al., 2025).
Models favour their own output (Panickssery et al., 2024). AI reviewers of ICLR 2026 papers showed a "hivemind effect of excessive agreement" (Baumann et al., 2026).
Error-finding is still weak on hard problems. On a benchmark of real errors in published papers, no model exceeded 21% recall (Son et al., 2025).

Software developers reached the same place: they accept about a third of an AI reviewer's comments, and human reviewers raise things agents do not (see software).

That suggests using AI review as a separate reviewer with an explicit checklist, run on every change including the editing agent's own, followed by a human. GO's ontology repository works this way. See GO ontology and One review checklist.

How to tell whether AI review works

We do not yet measure this (see Evidence). These benchmarks would tell us, and none needs new infrastructure:

Planted errors. Seed a sample of pull requests with known errors (a wrong evidence code, a taxon violation, a fabricated PMID) and count how many the AI reviewer catches. This is how the automation-bias studies above were done.
Held-out curation. Take annotations curators have already made, have the agent redo them blind, and compare. GOFlowLLM and the FlyBase FlyAOC benchmark do this.
Curator overrides. Track how often a curator rejects or changes an agent's work, and why. A falling rate is progress; a rate near zero may mean rubber-stamping.
Disagreement between model families. When two different models disagree on an item, send it to a curator. A 2026 trial (a preprint) used agreement across three model families as a confidence signal for physicians, and it improved their decisions.

What might this mean for biocuration?

Two questions follow from the other fields. Could every piece of curation get a second pair of eyes? And could curation become much more efficient, perhaps by an order of magnitude, without losing quality? The "10x curator" was put forward in October 2025 as an opportunity with a "reimagined role", alongside covering more biology in more depth (Mungall, Global Core Biodata Resource Forum, 2025). It is a hypothesis, not a result. These are the first signs, and the reasons for caution.

Second eyes in curation

AI reviewing human work. In the GO ontology, an AI reviewer now checks nearly every pull request: whether the right references are cited, whether the statements in a definition are correct, whether cross-references are right, and whether the change addresses the issue and nothing else. Over two months it reviewed 172 pull requests and caught real errors in biology, axioms, and provenance that passed the automated checks, in pull requests written by curators as well as by the editing agent (GO ontology team report, GO Consortium meeting, autumn 2026). A model checking annotations in the Gemma gene expression database found manual curation errors in about 2% of experiments, and because it quoted its evidence a curator could check each one quickly (Rogic et al., Database, 2026).

Humans reviewing AI work. At PomBase, an agent curates a whole paper for the author and a curator to review. In a pilot of nine papers it came close to curator recall and accuracy, with "no gross errors, only minor disagreements about specificity/interpretation" (PomBase, GO Consortium meeting, autumn 2026). GOFlowLLM drafted 2,538 candidate microRNA annotations, and curators agreed with 87% of the terms (Green et al., Bioinformatics, 2026).

These are small samples and early pilots. They suggest that a second reader on everything is no longer out of reach, which was not true a few years ago.

Could it be much more efficient?

Some observations that point that way:

Much of the time goes on work an agent can draft. Curating a paper means reading it, finding terms and identifiers, and entering the annotations: about 45 per paper in PomBase's community curation. GOFlowLLM produced its 2,538 candidates in 58 hours; about 1,400 microRNA papers had been curated by hand in the previous decade.
Review can be fast when the draft carries its evidence. In lipid transport curation, a batch of 16 GO-CAMs was quality-checked in one sitting, and the review led to a new GO term, a new Rhea reaction, and UniProtKB updates (UniProt, GO Consortium meeting, autumn 2026).
Some valuable tasks were never affordable. Re-reviewing hundreds of annotations affected by an obsoletion, or comparing GO-CAMs of the same pathway across species, are what have been called Tasks 2.0: higher value than the entity recognition and term-filling that earlier AI aimed at (Mungall, AIBIO-UK, February 2026).

And some that counsel caution:

Review capacity caps the gain. PomBase's community curation is limited by how fast curators can check submissions, not by how fast they arrive. More drafts with the same review process give a backlog. Software has seen the same thing (Writing code got cheap; shipping it did not).
The frontier is jagged. Agents in these reports were good at some steps and weak at neighbouring ones: drifting to generic binding terms, missing a sibling GO term a curator spotted, hedging on uncertain papers. Where the boundary lies in biology has not been mapped. See The jagged frontier.
Radiology's gains were real but modest. The best-evidenced field showed better quality with 44% fewer reads, not an order of magnitude.
Nobody has measured it end to end. The pilots measure steps, not curated knowledge reaching users.

If large gains come, they are more likely to come from changing what curators spend time on than from doing the same steps faster: making each review cheaper (the draft carries its evidence, validators catch the mechanical errors, an AI reviewer flags what to look at), and building quality in rather than checking it afterwards. See Make identifiers hard to fake and Fast and slow validation.

Where to look: output most likely to be wrong

From the studies above, from the GO AI Hub workshop, and from the repositories in our case studies, these are the places errors cluster.

Look harder at	Why
Anything not looked up: identifiers, PMIDs, quotes	Models fabricate plausible identifiers from memory. Lookups and validators catch these mechanically.
A step where a tool failed	Agents recover by guessing. In the workshop, an agent filled in missing abstracts from memory after PubMed failed, and did not say so.
Claims without a verbatim supporting quote	A quote makes a claim checkable in seconds; its absence is a signal.
Specific, deep terms and the long tail	Models know common genes and general terms best. Rare organisms, specific leaf terms, and recent findings are where they guess.
Output that is plausible and matches expectations	The DRAGON-AI errors were "plausible yet subtly incorrect". An answer that confirms what you expected is the one you check least.
Qualifiers, negation, taxon, and direction of effect	Small words that flip meaning: NOT, regulates versus directly acts, positive versus negative, the right species.
Long sessions	Context compaction makes agents forget earlier instructions. Most workshop participants hit it.

The first three rows can be checked by software. The rest need a curator. Put the curator's time there.

How curator roles might change

Software suggests what to expect. Developers still exist, but they spend more time specifying, reviewing, designing boundaries, and managing agents, and less writing code. For curators, the equivalents might be:

Reviewing drafts rather than writing every annotation.
Writing skills: turning a standard operating procedure into something an agent follows. A skill is a curation guideline that runs. See Write your own skills.
Regulating the loop: when the same mistake appears twice, fixing the process that produced it. See Regulate the loop. In this picture, people govern and manage (assign work to agents, read their reports, make the core decisions, write the rubrics and skills) on top of defined processes and deterministic checks (Mungall, AIBIO-UK, February 2026).
Modelling the hard cases that agents get wrong, and adjudicating disagreements.
Managing several sessions. Two or three at once is a realistic limit for most people. GO's ontology editors report the same problems developers do: getting lost in the agent's verbiage, and staying focused while switching between sessions. See What we do in practice.

New curators need to build judgment somewhere. Junior developers who learned with AI writing the code understood it less well, but those who asked the AI conceptual questions did fine (Shen and Tamkin, Anthropic, 2026). Have new curators do some work by hand, and use the agent to explain rather than only to do.

Practical starting points

Match the check to the task.

Task	Agent's role	Check
Literature triage	Screen and rank papers	Curator samples the rejects as well as the accepts
Identifier and term lookup	Do it, always with a tool	Validator, no human needed
Drafting annotations or GO-CAM models	Draft with quoted evidence	Curator reviews every item; validators run first
Reviewing existing annotations	Flag candidates for change	Curator decides; track how often flags are right
New ontology terms and definitions	Draft from a design pattern	Expert editor reviews; this is where subtle errors hide
Reports, dashboards, one-off scripts	Write them	Check the output looks right; throw away when done

Share what works. Many curators on the GO AI Hub converged on the same problems without knowing it: the agent's pull toward generic binding terms, the need for a clear rule on which molecular functions are worth annotating, papers it could not get the full text of. A regular call where curators show each other their workflows, and a habit of turning what works into a shared skill, spreads lessons faster than documentation does. See Writing and sharing skills on the hub.

Measure from the start. Pick one task, record the time and the error rate before and after, and publish the result. Whether curation can be much more efficient will be settled by numbers like these, and curation does not have them yet.