Hire a Research Data Curator for AI Before You Train on Your Archives

What Does a Research Data Curator for AI Actually Fix?

Your fine-tuning run finished last Thursday and the model confidently reports a binding affinity your own team retracted in 2023. The retraction lived in a lab notebook; the archive kept the paper. A Research Data Curator for AI is the person who catches that before training: annotation protocols written with the scientists who ran the assays, provenance traced per record, and hard gates on synthetic data.

The job splits into three pieces that labs usually hand to three people who never speak. The first is the protocol: what counts as a positive result, how a borderline reading gets labeled, what an annotator does with a run whose instrument log is missing. In research and development that document cannot be written by a vendor, because the ambiguities are scientific and so is the arbitration.

The second is provenance and licensing. Consortium data arrives with terms attached. Some instrument contracts claim rights over raw output. A curator who cannot tell you which of your sixty datasets may legally sit in a training corpus has left your counsel a problem to find later, and the answer will be needed under time pressure.

The third is synthetic data. When an assay produced forty usable runs, augmentation is tempting and the failure mode is quiet: the model learns the generator's assumptions and reports them back as evidence. Deloitte's 2026 technology outlook names data quality specialists for synthetic data among the roles AI-native organizations are adding, describing the work as making the training data trustworthy and representative ². In a lab that means holdouts drawn only from real runs, and a written rule for the ratio.

Which Backgrounds Produce a Good Scientific Data Curator?

The reliable feeders are people who already spent years being accountable for someone else's data: core facility managers, staff scientists on long instrument-heavy programs, biocurators from model-organism databases, clinical data managers, and research software engineers who have written an extraction pipeline for a lab. What they share is having been blamed for a bad record, which teaches a caution no course does.

The unexpected backgrounds deserve a longer interview than the obvious ones. Library and information science graduates who worked in institutional repositories arrive fluent in metadata standards and licensing, and they often cost less than a bench PhD because nobody has told them this job exists. Systematic reviewers from evidence synthesis teams have done little for years except define inclusion criteria and adjudicate disagreement between two readers, which is annotation under a different name. Field ecologists record collection conditions by habit, a discipline most wet-lab scientists never acquire.

What matters more than the degree is how the candidate got good with AI on their own work. The strong ones tell a story with a date in it: they ran an extraction over 900 papers, found the model inventing units on one journal's table format, and built a check for that specific failure instead of abandoning the pipeline. They can say where they stopped trusting output and what test they wrote at that boundary.

Ask what they threw away. A curator who has never deleted a dataset they built has never priced their own errors, and the archive you are handing them will ask for exactly that.

Where Do You Source a Research Data Curator, and How Do You Test One?

Start inside the building. The strongest candidate is usually a staff scientist two years into being annoyed by your archive, already trusted by the principal investigators, already cleared for the restricted data, and already aware of which instrument drifts. Outside, look at data-curation networks and repository staff rather than at annotation marketplaces, where the work is priced as piecework.

Venues that genuinely hold these people: Research Data Alliance and CODATA meetings, the Force11 community around scholarly data, and the curation staff behind Dryad, Zenodo and the Dataverse installations that large universities run. Domain databases are the deepest pool of all. Whoever curates entries at UniProt, the Protein Data Bank or FlyBase has spent a career turning other people's messy submissions into records other people rely on. Adjacent titles worth searching: data steward, biocurator, research data manager, clinical data manager, ontology curator.

The test that separates real from performed takes forty minutes and needs no take-home. Hand the candidate two hundred rows out of your own archive, ugly parts included: a column whose units changed in 2019, three duplicate sample identifiers, a free-text notes field. Ask for a first draft of an annotation guideline and for the questions they would put to you before labeling anything.

Performed expertise returns a tidy schema and no questions. Real expertise returns four questions, two of which you cannot answer, and a guideline with a section on what to do when the answer stays unknown. The same instinct shows up in the support knowledge curator role, where the failure is identical: content that reads clean and teaches a model something false.

What Does a Research Data Curator for AI Cost, and Can the Work Be Remote?

The person who would have caught that 2023 retraction is a scientist, and scientists have a price. No published salary series carries this title yet, so anyone quoting a precise band for it is guessing. Price it against the two roles it borrows from. The levels.fyi data scientist page reported a United States median total compensation near $180,000 as of mid-2026, with the middle half falling roughly between $134,000 and $250,000 ⁴. Lab scientist bands sit lower.

Demand and supply pull in opposite directions here. Demand pushes up: LinkedIn's Jobs on the Rise 2026 put data annotator fourth among the fastest-growing United States roles, describing people who label and review data under detailed guidelines and quality checks so that models train on accurate datasets ¹. A second ranking gets quoted in the same breath, the World Economic Forum's 2025 jobs report on big data specialists, but its page did not resolve for this piece, so no claim here rests on it ³. Supply pushes down at the bottom of the market, because generic annotation is genuinely commoditized.

Your candidate is not in the generic market. Someone who can arbitrate a scientific disagreement is a scientist and will not take a pay cut to stop being one. In practice, budget between a senior research associate and a senior data scientist for your region, and expect the top of that when the domain is regulated. Ask two or three candidates what they earn now and what a competing offer looked like; with no survey to cite, your own pipeline is the best series available.

On location the honest answer is split. Protocol writing, provenance work, licensing review and synthetic data audits are all remote-friendly. Anything touching instruments, physical samples, restricted patient records or an air-gapped archive is not. Most teams settle on two or three days on site through the first quarter, while the curator learns which records lie, then mostly remote. Decide which half your archive is before you post the role, because a fully remote posting for on-premise data wastes six weeks.

Close a Curator by Giving Them Authority Over the Protocol

Curators leave jobs where they are asked to clean data and forbidden from changing how it gets collected. So the closing argument is authority: the curator signs off on the annotation guideline, can refuse a dataset for training, and sits in the conversation upstream where instruments and intake forms are chosen. Offer that in writing and it competes with more money elsewhere.

Offers die most often on a reporting line into central IT with no scientific peer, on a title carrying the word technician, on no budget for annotator time from the principal investigators whose people they need, and on a mandate that stops at the archive door. That last one tells a curator they will be blamed for input errors they had no standing to prevent.

What they care about instead is durable and cheap to give. Publication credit on dataset papers. A named line in the data management plan of the grants they support. Travel to the curation meetings where their peers are, because this is a small field and reputation inside it is the career. Ownership of one metric they choose and report, such as the share of the archive with complete provenance.

Where the role sits matters too. Under a head of AI for research and development the curator has an ally with budget; under a lab manager they usually have neither budget nor standing. Write the first ninety days into the offer: a protocol for one dataset, a provenance audit of the archive, and one written recommendation you have agreed in advance to act on. If what comes back is that every paper from 2023 has to be reconciled against the notebooks before anything trains on it, you hired the right person and the archive was worse than you thought.

See a sample report >