Plan a pilot box
Menu

How to label AI-generated metadata in catalog records

Reader question: When AI suggests titles, subjects, names or descriptions for our records, how should we label that metadata, review it and explain it to users?

By Dastok Editorial TeamIssue: Published

Label AI-generated metadata field by field, not record by record. Every value a model suggested should carry a review status (machine-suggested, reviewed, edited or human-entered) with the date and the person who changed it, and that status should travel in every export: a status column in CSV, field 883 in MARC 21, a processing note in EAD. Show a label beside unreviewed values on the public record, publish a short statement explaining what the software did, and when a person changes a machine value, mark it edited rather than erasing where it came from.

The rest of this guide explains why labels matter, what the National Archives and American University already do, how fine a label should be, where it lives in each metadata standard, what happens to it when a cataloger edits, and how to explain it to the people who use your collection. A worked example with one fictional letter shows each step.

Why AI-generated metadata needs a label

AI-generated metadata needs a label because readers cannot reliably tell it apart from description a person wrote, and because its accuracy varies sharply from one field to the next. A label tells a researcher which values a person has checked, so they can decide how far to rely on each one.

Readers cannot see the difference

The clearest evidence comes from the ArchiveGPT study by Line Abele and colleagues, first posted on arXiv in July 2025 and published in Humanities and Social Sciences Communications in July 2026. A vision language model (InternVL2) drafted catalog descriptions for photographic prints on labeled cardboard mounts with archaeological content, and people curated the outputs. Then 139 participants, 29 of them archive and archaeology experts (most of them students) and 110 non-experts, judged 80 descriptions, half written by experts and half by the model, and said which were AI-generated.

Participants did better than chance, at what the authors call a moderate level. Average accuracy was about 62 percent, ranging from 54.9 percent for non-experts judging AI texts to 67.8 percent for experts judging expert texts, and experts did not significantly outperform non-experts. Most telling for labeling: the AI descriptions that participants rated as more accurate and more useful were the hardest to identify as AI-generated.

So the better the machine’s draft, the less likely a reader is to notice where it came from. A researcher who meets a fluent summary or a plausible subject heading in your catalog cannot know whether a person checked it unless the record says so. The authors also found that willingness to use AI tools and trust in them fell after participants had seen the model’s output, and they argue for an approach “where AI supports draft generation but remains subordinate to human verification.”

Accuracy differs by field

The Library of Congress’s Exploring Computational Description experiments (ECD1 in 2022 to 2023, ECD2 in 2023 to 2024) tested machine-generated catalog data, with catalogers reviewing the results. The Library reported that identifiers, author and title “can be extracted at around 85% accuracy, while others like subject headings, genre, and dates were only generated accurately at or under 50% of the time.” Its conclusion: “no current AI tools are currently good enough to run fully automatically for this task.”

That spread is the practical case for field-level labels. One record can hold a title that is almost certainly right and a date that is close to a coin toss, and a single flag on the record cannot say which is which.

Uncertainty is normal; hiding it is the problem

OCLC Research’s 2019 agenda Responsible Operations, by Thomas Padilla, put the principle plainly: “like human descriptions of collections, machine descriptions come with varying measures of certainty.” It named three challenges for library staff: setting certainty thresholds, bringing probabilistic data into existing systems without significant modification, and explaining algorithmic description in a way the communities they serve can understand. Labels are a practical answer to the second and third.

What the rules say

We found no professional rule, as of November 2026, that requires you to label machine-generated metadata. A 2026 document prepared for the Society of American Archivists’ Council recommends creating an AI task force whose deliverables would include best-practice guidance on “description and metadata,” but that is a proposal for future guidance, not a rule.

The content standard has moved, though. DACS (Describing Archives: A Content Standard), maintained by SAA, lists version 2022.0.4 in its changelog as “Updates to allow for documentation of AI-generated description.” Rule 8.1.5 in its Description Control element now reads: “Record the name(s) of the person(s) or generative artificial intelligence (AI) system(s) who created or revised the description, as well as the creation or revision date.” The element is marked “Added Value,” so it is optional, but it shows where descriptive practice is heading. The Library of Congress’s AI principles point the same way, committing the Library to “ensure AI implementations are transparent.”

What institutions already do

Three published examples show the main patterns: a public label on each machine contribution (the National Archives), dedicated fields backed by a public policy (American University), and review tags inside a vendor’s tool. None is a complete scheme, but together they cover most of what a small institution needs.

The National Archives: a label on each contribution

The National Archives catalog defines an element called Attribution Type: Artificial Intelligence (AI) / Machine Generated Label in its Lifecycle Data Requirements Guide. The definition: “AI / machine-generated labels attribute contributions created through the use of artificial intelligence and machine learning tools by NARA or a NARA partner.” The stated purpose is “to provide additional metadata to make content in the Catalog more discoverable and provide transparency for AI / machine generated content.”

The element’s details are instructive:

  • It is a public element, repeatable, not mandatory, and read only.
  • It is available on contributions such as tags, OCR transcriptions, translations, location names and occupations, rather than on the description as a whole.
  • It depends on a contribution and can combine with other attribution, so a tag can be both machine-generated and donated by a named partner.
  • The guide’s example is a Revolutionary War pension file whose catalog entry carries AI / machine-generated tags such as “Continental Army” and “Divorce.”

The lesson is that NARA labels the piece the machine made, not the record, and shows that label to the public. The page was last reviewed on September 24, 2024.

American University: two fields and a policy

American University Library published a policy on AI-generated descriptive metadata, effective August 2025, when it began piloting AI-generated descriptions for text and image resources; it already used AI-generated audiovisual transcripts. The library generates the metadata with JSTOR Seeklight, through JSTOR’s Digital Stewardship Services. “All records containing AI-generated metadata are clearly labeled,” the policy says, in two fields: an “AI-Generated Metadata Control field” and an “AI-Generated Metadata Statement field.”

The policy is specific about review. Its quality control measures include “100% human review of core fields such as title, description, and subject,” “targeted review of other fields based on AI confidence scores and representative sampling,” refinement of AI content where needed, and manual creation of many additional fields by catalogers.

It also lists the thirteen fields that may contain AI-generated information, from title, creator and date to name and topical subjects and audiovisual transcripts. It names the factors used to select collections, including privacy, donor intent and deed of gift restrictions, the “impact of incorrect information vs access,” and the amount of existing description. And it gives an email address for corrections, asking for the item’s URL, local identifier and title.

Vendor tools: review tags inside the interface

JSTOR’s Seeklight documentation describes “tags for AI-generated data,” confidence scores to prioritize review, and a way to mark records “Reviewed.” It also says that “unreviewed records can still be published,” with a prompt reminding users to review them first.

Labels inside a tool help reviewers, but they are not the same as labels in your catalog. Before you adopt any tool, ask whether the review status leaves the tool with the record, in which field, and whether it survives export to your collection system and a later migration.

What the examples leave out

All three label the unit the machine produced, explain the label publicly or in documentation, and review by risk. None publishes a rule for what the label becomes after a person edits a machine value, or a mapping to Dublin Core, MARC and EAD. The next sections fill those gaps.

Record, field or value: how fine should the label be?

Label at the field level as a minimum, and at the value level for repeating fields such as subjects, names and places. A record-level flag is better than nothing, but it cannot tell a researcher which part of the record to doubt.

Level What the label says Good for Weakness
Record This record contains AI-generated metadata A first step; a filter for staff Does not say which fields
Field The subjects were machine-suggested; the date was reviewed Most catalogs and exports A field with mixed values needs a rule
Value “Credit” reviewed; “Labor disputes” machine-suggested Subjects, names, places, tags Needs paired lists or a separate table

American University’s policy describes labeling the records that contain AI-generated metadata, with a statement field that explains it. That is a sensible minimum for a pilot. As soon as review is partial, though, field-level status becomes necessary, because “reviewed in part” is exactly the condition a record-level flag cannot express.

A rule for mixed fields

If a field holds several values with different statuses and you cannot label each value, label the field with its least-reviewed status. “Subjects: machine-suggested” when one of three terms is unreviewed is cautious and honest; “Subjects: reviewed” would be false for one term.

Record flags as a summary

A record-level flag is still useful for staff, as long as it is calculated from the field labels rather than set by hand. “Contains unreviewed machine values: yes” lets you filter a review queue, count progress and find every record touched by a given batch.

A review-status vocabulary that stays honest

Use four statuses: machine-suggested, reviewed, edited and human-entered, plus a log entry for suggestions a person rejected. Together they answer the two questions a researcher has about any value: did a person look at it, and did they change it?

Status Meaning Suggested public wording
Machine-suggested Produced by a model; no person has checked it “Suggested by software, not yet reviewed”
Reviewed Produced by a model; a person checked it against the item and accepted it unchanged No label, or “Reviewed”
Edited Produced by a model; a person changed it No label, or “Reviewed”
Human-entered Written by a person with no machine suggestion No label
Rejected (log only) A person removed a suggestion Not shown

Legacy records need one more value. If no one recorded how older description was made, call it “unknown” rather than assuming a person wrote it. EAD3 makes the same choice: its <agenttype> element takes the values “human,” “machine” or “unknown.”

New model run adds suggestions only Machine-suggested model output, unchecked accepted as is changed deleted Reviewed person, date recorded Edited original kept in history later change Rejected not published, logged Human-entered no machine suggestion Lime: checked or written by a person. Blue: machine output.
How a value's review status changes. Editorial scheme based on the practices described in this article; the four statuses match the labels in Dastok's AI policy.

What to record alongside the status

The status alone tells a reader whether to trust a value. A few more data points let staff audit the work and fix problems later:

  1. Who changed the status: a name, initials or staff number.
  2. When, as an ISO 8601 date such as 2026-11-18.
  3. Which process made the suggestion: the tool or model and its version, and the prompt version if you wrote your own.
  4. Confidence, if the tool reports one. Use it to order the review queue, not as a public label.
  5. Evidence, where the value came from in the item: the dateline, a postmark, a studio imprint on the mount.
  6. Batch, so that if one run turns out to have a systematic error you can find every value it touched.

Keep “reviewed” meaning reviewed

Reviewed should mean that a person compared the value with the item, its image or its text. It should not mean that a batch passed a sample check. If you accept a batch because a random sample met your error threshold, the values nobody looked at stay machine-suggested, and the batch’s measured error rate goes in your public statement instead. This keeps the most important word in the scheme trustworthy.

What happens to the label when a person edits a machine value

When a person changes a machine value, its status becomes edited, the person and date are recorded, and the original suggestion stays in the record’s history. The label describes how the current value came to be; the history keeps the evidence of what the machine proposed.

Edit rules

  1. Accepted unchanged: the status becomes reviewed. The history still shows the value came from a model.
  2. Changed but built on the suggestion: a corrected spelling, the authorized form of a name, a narrower subject term or a tightened summary. The status becomes edited, because part of the machine’s choice survives.
  3. Discarded and replaced: if the reviewer rejects the suggestion and supplies a value that shares nothing with it, log the suggestion as rejected and mark the new value human-entered. When in doubt between this rule and rule 2, choose edited; it errs toward disclosure.
  4. Deleted with no replacement: log the suggestion as rejected. It is not published.
  5. Later change to a reviewed value: the status becomes edited, with the new person and date.
  6. Bulk changes: a global update, such as changing an authority form across a thousand records, is logged as an event but does not promote machine-suggested values. Nobody compared those values with the items.
  7. Model reruns: a new run may add suggestions to the review queue, but it never overwrites a reviewed, edited or human-entered value. If a better model proposes a different date for a reviewed letter, a person decides.
  8. Corrections from users: when a reader reports an error and staff fix it, the status becomes edited, with the source of the correction noted in the history.

Why keep the original suggestion

Keeping what the machine proposed costs little and pays back three ways. You can count how often each field’s suggestions are accepted, edited or rejected, which tells you where the tool saves time and where a cataloger may as well start from scratch. You can find every value a faulty run produced if a systematic error turns up later, such as a model reading day-month dates as month-day. And you can show funders and boards how much of the description people checked.

How the statuses map to MARC

MARC 21 field 883 has a first indicator that distinguishes fully machine-generated (0), partially machine-generated (1) and not machine-generated (2) data. Machine-suggested and reviewed values map to 0, because the content is still entirely the machine’s. Edited values map to 1. Human-entered values map to 2, or simply carry no 883. The indicator records origin, not review, so review status needs a home of its own, as the next section explains.

Where the label goes in Dublin Core, MARC, EAD and your CMS

Each standard handles provenance differently. Dublin Core has no element for it, MARC 21 has a dedicated field, EAD records machine activity for the finding aid as a whole, and most collection systems need a custom field. Decide where the status lives in each format you export before the first batch is cataloged.

FormatWhere the status goesWhat it capturesLimit
CSV for importA status column beside each labeled field, plus reviewer and dateField or value statusLost if the import template has no matching field
Dublin CoreNo element exists; use the CSV columns or a local element, never dcterms:provenanceWhatever you defineSimple harvests drop local additions
MARC 21Field 883 linked by $8 with link type p; review in a local 9XX fieldOrigin per field, process, date, confidence883 does not record review
EAD3 and EAD4Maintenance history (agent type machine or mechanism) and a processing information noteWho changed the finding aid; a readable note per componentNot value level
DACS8.1.5 Archivist and Date names persons or generative AI systemsHuman-readable historyOptional element
Collection systemDedicated or user-defined fields; a note with a fixed prefix as a fallbackDepends on the systemTest that exports keep them
Sources: Library of Congress MARC 21 field 883 and MARC introduction; EAD3 and EAD4 tag libraries; DCMI Metadata Terms; DACS 8.1, all checked November 2026.

Spreadsheets and Dublin Core

Dublin Core’s fifteen elements describe the resource, not the metadata about it, so none of them holds a review status. One term looks tempting and should be avoided: DCMI defines dcterms:provenance as “a statement of any changes in ownership and custody of the resource since its creation that are significant for its authenticity, integrity, and interpretation.” That is the item’s custody history. Putting AI labels there would mislead anyone who maps it to a provenance note.

The practical pattern is paired columns in the spreadsheet you deliver or import. For each field a machine can fill, add a status column, and keep repeating values and their statuses in the same order:

identifier,subject,subject_status,subject_reviewer,subject_date
ACHS-3-019,Labor disputes; Credit,machine-suggested; reviewed,JR,2026-11-18

Paired lists are fragile if someone re-sorts one column and not the other. For large projects, a separate assertions file with one row per value (identifier, field, value, status, reviewer, date, process) is safer and easy to join back. Our metadata mapping worksheet is the place to record which fields carry a status column, and our comparison of Dublin Core, MARC and EAD explains which standard suits which system.

Aggregators that harvest only simple Dublin Core will not pick up your status columns. For records that still hold unreviewed values, add one sentence to the description, such as “Some subject terms in this record were suggested by software and have not been reviewed.” It is blunt, but it travels.

MARC 21: field 883

MARC 21 has a field built for this job. Field 883, Metadata Provenance, was added in 2012 as “Machine-generated Metadata Provenance” and renamed in 2020 “to allow the recording of non-machine-generated content.” It is “used to provide information about the provenance of metadata in data fields in the record,” and it attaches to the field it describes through subfield $8, using the field link type “p” for metadata provenance defined in MARC’s control subfields appendix.

The parts that matter for labeling:

  • First indicator: 0 fully machine-generated, 1 partially machine-generated, 2 not machine-generated.
  • $a creation process, a process name or other description.
  • $c confidence value, a number between 0 and 1.
  • $d creation date, written yyyymmdd.
  • $q the MARC organization code of the agency that ran the process.
  • $u a URI describing the process, and $x a validity end date.

A machine-suggested subject heading, linked to its provenance, looks like this, with your own values in the brackets:

650 #0$81\p$aLabor disputes.
883 0#$81\p$a[tool and version]$d20261110$q[your organization code]$c0.7

Field 883 says how a value was produced, not whether a person later checked it. MARC reserves the 9XX fields “for local implementation,” so a local field linked with the same $8 number can carry the status, the reviewer and the review date. Before relying on either field, export a few records from your catalog and confirm that 883 and your local field come back out intact.

EAD and DACS

EAD records machine involvement for the finding aid as a whole. In EAD3, the <agenttype> element inside <maintenanceevent> takes “human,” “machine” or “unknown” for the agent responsible for “the creation, modification, or deletion of an EAD instance.” EAD4, the current version, whose Tag Library 4.0.0 is dated July 2026, moves this to an @agentType attribute on <agent>, with a standard list that includes “human,” “mechanism” and “unknown.”

Neither tells a reader which element a model drafted. For that, use the processing information element (<processinfo> in EAD3, <processInfo> in EAD4), which holds “basic information about accessioning, arranging, describing, preserving” and other preparation of the materials, and can sit at the collection level or on any component. A note such as “Scope and content notes for series 2 were drafted by [tool] in November 2026 and reviewed by [name]; subject terms for folders 4 to 9 are machine suggestions not yet reviewed” is readable in any finding aid display.

DACS 8.1.5 supplies the content rule for that note. Its own example reads: “Finding aid written by Jan Jones, 2021. In 2025, Sam Johnson used ChatGPT to revise the biographical note and create an inventory for the 2025 accession.”

Collection management systems

Most collection systems were not built with field-level provenance in mind. In rough order of preference:

  1. A dedicated status field for each labeled field, if your system supports user-defined fields that also export.
  2. A pair of record-level fields like American University’s control and statement fields.
  3. A note field with a fixed, searchable prefix such as “Machine suggestion, unreviewed:” as a fallback.

Whichever you choose, test a round trip: export ten records, import them into a test copy, and confirm the labels survive. Our guide to importing digitized records into collection software covers that kind of test.

Deciding what machines draft and how much people review

Decide by risk. Fields that control access or carry legal weight are reviewed every time or written by people, descriptive fields can be sampled, and low-confidence suggestions go to a person first. Labels make sampling honest, because values nobody checked stay visibly unchecked.

Field Machine draft? Review level Why
Devised title Yes Every record First thing users see
Dates Yes, from datelines and postmarks Every record Dates were at or under 50% accurate in LC’s tests; they drive sorting and filtering
Names of people Yes Every record, matched to your authority list Wrong identification can harm living people and families
Places Yes Every record, or authority-matched Historic and current place names differ
Subjects Yes Sample, or every record in small collections Subject headings were at or under 50% accurate in LC’s tests
Summary or description Yes Every record, or a large sample Fluent errors are the hardest to spot
Rights statement and access restrictions No People decide Legal and donor obligations
Culturally sensitive description Only if the community agrees The community decides Community protocols take precedence

American University’s policy follows the same logic: full review of title, description and subject, and targeted review elsewhere based on confidence scores and sampling. Where you set the line depends on your users. A genealogy-heavy collection may need every name reviewed; a large run of municipal minutes searched mostly by full text may tolerate sampled subjects.

Sampling with labels

For example, suppose a hypothetical batch of 400 letters has machine-suggested subjects. You review a random sample sized with the QC sample size calculator and count how many terms needed changing. If the error rate exceeds your threshold, you review the whole batch or stop using the model for subjects. If it passes, the sampled values become reviewed or edited, the rest stay machine-suggested, and the measured error rate goes into your public statement.

Watch the acceptance rate

After each batch, count accepted, edited and rejected suggestions per field. A field where most suggestions are rejected is costing reviewers time rather than saving it; turn the machine off for that field or change the prompt, model or vocabulary. Our AI cataloging service counts the same three outcomes per field for the same reason.

Showing labels to users

Show a plain-language label beside every unreviewed value on the public record, link it to your statement, and offer a way to report errors. Reviewed and human-entered values usually need no label on screen, as long as the statement explains the scheme and the exported data carries it.

Wording and placement

  • Put the label next to the value, not in a footer or a separate tab. A researcher copying a date should see its status in the same glance.
  • Use plain words. “Suggested by software, not yet reviewed” says what happened. Confidence scores mean little to most readers and suggest a precision the numbers do not have.
  • Do not rely on color alone. Pair any color or icon with text, so the label works for people who cannot distinguish colors and for screen readers.
  • Link the label to your public statement, so a reader can learn what the software did and how review works.

Search, browse and citations

Machine-suggested values can help people find things, so most institutions will want them in keyword search. Browse lists and facets are different: they present values as curated access points, so consider holding machine-suggested subjects and names out of facets until a person has reviewed them. If you offer meaning-based search, our article on semantic search for archives covers how it changes what users find and what they trust.

Citation tools and metadata downloads should carry the labels too. A researcher who downloads records as CSV should receive the status columns, and a generated citation for an item whose date is unreviewed should not present that date as settled.

Corrections from users

Give readers a simple route to report errors, as American University does with an email address and a request for the item’s URL, identifier and title. Route reports into the same review queue as machine suggestions. A value corrected after a report becomes edited, with the report noted as the source.

A public statement you can adapt

A public statement explains, in a few short paragraphs, what the software did, which fields it touched, how people reviewed them, what the labels mean and how to report errors. American University’s policy is a good model, and the template below follows the same structure.

A useful statement covers:

  1. When machine description started, and in which collections.
  2. Which tools or models were used, by name and version.
  3. Which fields they suggested, and which they never touch.
  4. How review works for each field: every record, a sample or none yet.
  5. What each label means on the page.
  6. Measured error rates from sampling, with the date measured.
  7. Collections excluded and why, such as donor restrictions, privacy or community protocols.
  8. Whether images and text were used to train anyone’s model.
  9. How to report an error, and the date the statement was last updated.

Here is a template. Replace everything in brackets and delete what does not apply.

How we use software to describe our collections. Since [month and year], some records in [collection names] include descriptive information suggested by [tool or model and version], software that reads [scanned images, recognized text]. It suggested [titles, dates, names, places, subjects, summaries]. It did not decide rights statements, access restrictions or [other fields].

Every suggested value is labeled. “Suggested by software, not yet reviewed” means no staff member or volunteer has checked the value against the item. Values without that label were written or checked by our staff. We review [all titles, dates and names] in every record and [a random sample of subject terms] in each batch. In [month and year], [number] percent of sampled subject terms needed correction.

We did not use software on [collections excluded] because [reason]. [Your images and text were not used to train third-party models.]

If you find an error, write to [address] with the item’s web address, identifier and title. We will review the value and record the correction. Last updated [date].

Date the statement and update it whenever you change tools, models or review levels. Keep earlier versions, so a researcher who cited a record in a given year can see the rules that applied then.

A worked example: labeling one fictional letter

The example below follows one letter through a machine pass, review, a later sample check and a model rerun, showing the label at each step.

Example (fictional). The Aster County Historical Society runs a machine pass on Box 3 of the Okerlund family papers. For letter ACHS-3-019, Ida Okerlund to her sister Marit, the recognized text reads: “The smelter men walked out on Monday and Walter says the store will carry them on credit until the matter is settled.”

The model suggests these values, all labeled machine-suggested:

Field Suggested value Evidence
Title Letter, Ida Okerlund to Marit Okerlund Salutation and signature
Date 1894-02-03 Dateline
Names Okerlund, Ida; Okerlund, Walter Signature; text
Places Wren Gulch (Colo.) Dateline
Subjects Labor disputes; Credit Text
Summary Ida reports that smelter workers went on strike and that Walter’s store will give them credit. Text

A volunteer cataloger, initials JR, reviews the record on November 18, 2026, with the page image beside the text. Title, date, names and place match the letter and the society’s name list, so they become reviewed. The summary contains a quiet error: the letter says “the store,” not Walter’s store, so JR rewrites it as “Ida reports that smelter workers walked out and that Walter says the store will extend them credit.” Its status becomes edited, and the model’s version stays in the history. Subjects are left for the batch sample and stay machine-suggested.

On the public record, only the subjects carry a label: “Suggested by software, not yet reviewed.” The CSV export carries a status column for every field, and the MARC record links its 650 fields to field 883 with first indicator 0.

A week later, the sample check reaches this letter. JR replaces “Labor disputes” with the society’s local heading “Strikes and lockouts” (edited, 883 first indicator 1) and accepts “Credit” (reviewed). The record no longer has an unreviewed value, so the public label disappears and the record-level flag turns off by itself.

A year later, the society tests a newer model, which proposes the subject “Smelters” for the same letter. The term enters the review queue as a new machine suggestion. Nothing already reviewed or edited changes until a person accepts it.

The pattern generalizes. Every value has one status at a time, every change of status has a person and a date, and nothing a person has checked is overwritten by a machine. Our example collection shows records like this one with machine and reviewed fields side by side.

What to do next

Start small: pick one collection that already has, or will soon have, machine suggestions, and put the scheme in place there before the backlog grows.

Checklist

  1. List every entry point for machine output: recognized text, suggested fields, summaries, tags and anything a vendor delivers.
  2. Adopt the four statuses and write the edit rules on a single page your reviewers keep at hand.
  3. Choose the home for the status in every format you export, and record it in your mapping document.
  4. Add the fields to your collection system and test a round trip with ten records.
  5. Set review levels by field, with a sampling plan and an error threshold for each sampled field.
  6. Write and date the public statement, including a route for corrections.
  7. Train reviewers that reviewed means checked against the item, not accepted in bulk.
  8. Report the numbers after each batch: accepted, edited and rejected suggestions per field.

Mistakes to avoid

  • Labels that exist only inside a vendor’s interface and vanish on export.
  • Promoting a whole batch to reviewed because a sample passed.
  • Letting a model rerun overwrite values a person has already checked.
  • Using dcterms:provenance for metadata origin.
  • Deleting the original suggestion when a person edits it.
  • A public label that relies on color alone.

If you work with an outside cataloger or vendor, put the statuses, the edit rules and the export mapping in the agreement, so labeled records come back in a form your system keeps. Dastok, a Colorado Springs firm that digitizes and catalogs historical collections, labels every machine-suggested field until a person reviews it, in the records it delivers; our AI policy sets out what we label, what people review and what you can refuse. For the decision that comes before labeling, how deeply to describe each series in the first place, see our guide to item-level and collection-level description.

Working through this with a real collection?

If this article describes a decision you are facing, tell us about the collection. A pilot box answers most questions with your own material.

Plan a pilot box

Studio and contacts

24 South Weber Street, Suite 225
Colorado Springs, CO 80903
(719) 244-8450sales@dastok.com