AI vs manual tagging: when human metadata matters

Use AI-generated metadata for descriptions that can be inferred from an asset. Add human metadata for business facts that come from elsewhere: the campaign, product code, client or intended placement. Review generated information when an error affects retrieval or use. Manual approval of every tag should not be a default requirement.

The goal is finding and reusing the right asset

A useful metadata record helps someone find an asset and decide whether it fits the task. It does not need to describe every visible object. A photograph can be retrievable through a caption, filename, campaign term or product identifier even when its tag list is incomplete.

Do not review AI metadata simply because AI generated it. Add or correct information when it materially improves retrieval, business context or asset use.

This gives review a concrete purpose. A wrong product name that sends someone to the wrong image deserves attention. An omitted background object may not matter if nobody searches for it. Decide what needs work by testing the searches and decisions the library must support.

What AI can infer from the asset

Image analysis can often identify visible objects, colors, clothing and broad scenes. A generated caption can describe relationships that isolated tags miss, such as a person standing beside a vehicle or a jacket being worn outdoors. These descriptions provide starting points for visual searches.

Consider a photograph of a person wearing a red jacket against snow-covered mountains. Generated terms might include person, red jacket, snow, mountains and outdoor. A caption could combine them into a description of the scene.

These are inferences. Small details, partially hidden subjects and specialist products can be misidentified. Snow may suggest winter without establishing when the photograph was taken. A plausible description should not become a confirmed fact merely because it appears in a metadata field.

What the pixels cannot establish

The same photograph could belong to a hypothetical catalog record with season AW27, SKU AL-482, campaign Norway campaign and placement homepage hero. Those values describe how an organization uses the file. They are not established by recognizing a red jacket and mountains.

Generated metadata answers: what appears in this asset? Human or business metadata answers: what does this asset mean to our organization?

Campaign names, clients, project codes, product families, launch phases and intended channels need a source outside visual recognition. That source might be a person, a product catalog, a campaign brief or another application's records. Even a legible code in an image should be checked against the relevant business record before it becomes an operational identifier.

A campaign's target market also differs from the photograph's location. An image selected for a Norway campaign need not have been taken in Norway. Record the intended market from the brief; do not treat a landscape description as evidence of geography.

Approval and contractual context require their own evidence. Visual tagging cannot establish permission to reuse a photograph. The guide to digital asset rights management explains what to record alongside an asset and why a metadata summary does not enforce a restriction.

Keep generated and user-managed metadata distinct

Generated descriptions and confirmed business values have different sources. Keep that distinction visible wherever the asset system supports it. A generated caption describes the model's interpretation; a user-entered campaign identifier records information supplied by the team.

There is little reason to copy every generated tag into a manual field. Use human fields for the information automation cannot supply reliably, or for a correction that matters. Keep a short reference for recurring campaign names, product codes and project terms so people enter the same value consistently.

A controlled business term also needs an owner. If homepage hero describes the current placement, decide who changes it when that placement changes. Otherwise an accurate entry can become stale. Metadata should earn its maintenance cost.

When human metadata is worth adding

• Business identifiers: add the confirmed SKU, campaign, client or project code when people use it to find assets. Prefer the identifier from the source system over a guessed product name.
• Important factual errors: address a description that confuses two products or misidentifies the main subject when that mistake affects selection. Check the relevant searches after the correction.
• Frequently reused assets: add richer context to a homepage hero, campaign master or product image that people repeatedly need. Intended placement, version and product identity can matter more than additional generic visual tags.
• Internal terminology: record the team's actual search terms. A product family called Heritage Fit needs that confirmed business label; visual recognition of a jacket does not supply the internal classification.
• Operational dependencies: if an integration selects an image by a product code or placement value, use a confirmed value with agreed formatting. A probabilistic description should not determine an exact business identifier.

Match the correction to the problem. A missing campaign name needs business context. A false description needs a correction through the controls the asset system provides. Adding the correct term alone may leave the incorrect generated term present, so verify the result instead of assuming the edit resolved the conflict.

Public reuse deserves a separate check. An internal caption may help someone locate an image while omitting information needed on a particular page. Review public captions for their publishing context. Write alt text for the image's purpose in that placement: W3C's image guidance distinguishes informative, decorative and functional uses. A generic asset description is not automatically suitable alt text or finished SEO copy.

When manual review adds little value

Checking tree, road, car and daytime across thousands of photographs is hard to justify if people already find the required images. Review should answer a retrieval or usage problem, not make the tag inventory look complete.

The same applies to adding synonyms nobody uses, removing every harmless repetition or filling fields without a defined purpose. Each manual value creates work when an asset is reused, renamed or reassigned. A small set of consistent business terms can be more useful than an elaborate taxonomy that nobody maintains.

If an error recurs across a relevant group of assets, inspect that group and address the affected searches. That is a reason for targeted review. It does not make manual approval of all generated metadata a necessary step before the library can be used.

A lightweight hybrid tagging workflow

1. Upload the asset and retain any useful source filename or identifier. For a system with automated enrichment, let it generate its descriptive context.
2. Use the library for normal tasks. Try the visual descriptions and business terms people already use instead of beginning with a tag-by-tag inspection.
3. Add missing business context from a reliable source. Record a campaign, product code or project term when it answers a real retrieval need.
4. Correct consequential errors as they appear. Prioritize mistakes that lead to the wrong asset or affect how someone uses it.
5. Reuse consistent human terms. Keep exact identifiers consistent with the source record; use a brief shared naming reference for recurring business labels.
6. Check assets used in publication or an integration for the values that particular use requires. Keep this check separate from approving every descriptive tag.
7. Repeat representative searches after changes. Keep the metadata work that improves the result and reconsider fields that nobody uses.

A process can start with this workflow and become more specific when an actual dependency requires it. For example, a product feed may need a confirmed SKU on every image it consumes. That requirement belongs to the feed's workflow; unrelated archive photographs do not automatically need the same review.

How Vyso uses generated and user-managed context

Vyso's asset enrichment and search features include background generation of image metadata. Generated titles, captions and tags are stored as aiTitle, aiCaption and aiTags. This supplies descriptive context without requiring a manual tag list for each uploaded image.

The metadata API also supports user-managed fields, including userTitle, userDescription, userCaption and userTags. These are separate from the generated fields. Use them for confirmed descriptions and business terms where needed.

Vyso's indexed search draws on filenames, generated titles and captions, user-entered descriptions and both sets of tags. Its search service combines full-text and semantic results. Descriptive context from the image and terms supplied by the team can therefore contribute to retrieval without depending on one exhaustive tag list.

User and generated tags can coexist. Adding a user tag does not delete a generated tag. Check searches after adding a correction, especially where a wrong generated term could still influence the results.

Check whether people can find the assets

Evaluate the workflow through real retrieval tasks. For the illustrative jacket photograph, try a visual request such as red jacket in snow, an internal campaign term and the confirmed product code. These test different sources of context. No single query needs to depend on every individual tag being correct.

Keep a small set of common requests and ask a colleague to find the intended assets. Record what they searched for, whether they found a useful result and what information was missing. Repeat the exercise after a meaningful metadata change.

• Do common campaign and product terms return the assets people expect?
• Can someone find a useful image without knowing its filename?
• Are people repeatedly browsing collections or folders because a needed term is absent?
• Do the same factual mistakes keep causing corrections or unsuitable selections?
• Does the time spent finding an asset improve after a targeted metadata change?

If a search fails, check whether the missing information is visual description, business context or an indexing problem before assigning manual tagging work. Correct the relevant cause, then rerun the task. A longer tag list is useful only when it helps someone find or use the right asset.