Skip to content
Interface language: English
Audit / Signal

AI images in company materials: what file provenance actually establishes

Distinguish a file provenance record from the truth of the scene and review an image through publication.

Article language: English

A person holds a red comb against a long dark braid beside a bright window.
In this article

An image shows a spacious studio, and the file has a readable editing history. Someone says, 'We have confirmation that the photograph is authentic.' Yet the studio may have been generated, and the history may document that honestly. A provenance credential can be valid for a scene that never existed. It still would not be a good idea to put that studio's address on a meeting invitation.

Separate three questions when using AI images in company materials: where the file came from, what was recorded about its processing and what the image is meant to represent to the audience. The answers can support one another, but they are not interchangeable. Visible history does not automatically confirm product conformity, an actual completed project or appropriate use in a particular offer.

An operational review should follow the material from source to published version. This helps establish what information survived, who approved the use and what actually reaches the website. A screenshot of a green indicator in a tool does not cover the file's whole journey.

Start with what the image says on the page

Imagine a company designing studio equipment. It uses three kinds of material: a photograph of an existing shelving unit, a visualisation of a proposed variant and a generated scene conveying the brand's mood. Each can be useful. Each also needs the right placement and description so readers understand its role.

A photograph of an existing product helps evaluate a particular object. A visualisation can show a proposed design direction. A brand scene creates an aesthetic context. If the third appears under 'Our completed projects' without explanation, the problem is not missing technical metadata. It is the message created by publication in that location.

Record the image's role before checking a tool. Does it show a product, a service result, a concept or only atmosphere? Do the caption and neighbouring text match that status? These questions separate checking the file from editorial approval of what its use means.

The guide to using brand photography on a website (in Polish) explains how to give each image a role in the page sequence. For AI material, add information about how it was made and the limits of what it represents. The person publishing should not have to guess its status from an impressive appearance or 'final' in the filename.

Prepare a short record: source, approved role description, version, responsible person and places of use. It need not contain every experiment. It should explain why this particular version appeared beside this particular offer.

Do not turn this card into a claim that technical credentials answer every question about the material. Conditions of use, consistency with the offer and publication decisions have their own bases and owners. This review checks the flow and the scope of available evidence rather than assigning one tool the role of universal seal of approval.

What provenance credentials contribute

The C2PA explainer distinguishes an asset's history and integrity from the truth of the depicted content. Provenance information may help identify a source and changes, but does not itself say whether a scene corresponds to reality. Validity of the signature and its binding to the file does not by itself establish the truth of individual assertions recorded in the history. This is a basic limit on interpreting verification.

For the fictional studio, a successfully checked history of a generated image remains a history of a generated image. It does not become documentation of a built interior. The team may use it as a concept or brand material if that is how it is presented to the audience.

When reading a credential, check which assertions are actually available. Is information visible about a tool, editing stage or issuer? Do not infer the whole process from an empty field. The report should name the scope read from that file instead of replacing it with a general 'authenticity verified'.

Adobe's description of Content Credentials describes adding further credentials during editing. This describes recording history in supported tools. For your own workflow, check the particular file after export and the surface where the audience is expected to read the information.

Do not assume every asset has a complete history. A file may have passed through a tool or stage that provided no suitable record. The C2PA explainer describes possible gaps and removal of metadata. Absence of a credential alone cannot determine whether an image came from a camera or a model. Other process materials and an honest indication of the gap are needed.

This differs from image recognition by a model. A detector or visual similarity can provide another signal, but does not replace the history of the particular file. The guide to finding a product from a photo (in Polish) also concerns similarity, not confirmation of who prepared the material or how.

Follow the actual route from source to publication

For a trial choose one asset with a recognisable status and several transformations actually used, such as cropping, resizing and export to the format used on the site. There is no need to cover the whole archive at once. You do need to pass through the tools the team really uses.

Keep the source and output versions. At each stage record what was checked and the result.

Attach the result to the file actually opened: its identifier, size and place of retrieval. Record the tool and time of the check separately. Another person can then repeat it. The name 'shelf-final' may be too common, particularly in a folder that also contains final, ultimate and really-ultimate versions.

Aeneia conceptual photography

If information can no longer be read, identify when and the scope of the observation. Do not automatically call it evidence of intentional removal. Establish the cause from the behaviour of the particular workflow.

Check the version retrieved from the place of publication. The approver may inspect a file in a folder while the website serves a transformed variant. Acceptance should concern the version actually reaching the audience. Otherwise the report confirms something different from what was published.

In your CMS content publication process, include this check alongside the caption, alternative text and placement. The guide to publishing content in a CMS (in Polish) can help structure that process. Source data and a public explanation have different roles: one can support verification, while the other helps readers understand the image's status without knowing the tools.

Include transfer to a partner. Send the same package normally prepared for publication: image, role description and agreed caption version. Ask to see the final use. If the partner received only the file, find out where the description came from. This can reveal a gap invisible in the internal archive.

When accepting external work, compare the image itself too. Replacing it with a similar asset from the same session may keep the visual style while changing what is represented. Responsibility for the version should not depend on the memory of the person who chose thumbnails a month earlier. If a model makes changes, separately record your own permitted limits for editing a product photograph. The guide to setting boundaries for AI edits to product photos (in Polish) can help frame that decision. Retained history does not establish that product features remained unchanged.

Test image replacement within an existing component. When an editor replaces an old photograph with a new visualisation, the caption should not automatically remain 'Completed studio'. This is a content-to-asset association error even if the new file retains a complete history. Image and description versions must be accepted together.

Add a bulk-export case. Does each version have its own identifier and a retained connection to the approved source? Were names swapped during publication? Correct metadata on one file does not prevent the correct file being used in the wrong card.

In a catalogue, pay special attention to visually similar variants. The same shelving unit may come in several dimensions, while a generated scene shows only a general design direction. Readers should not treat scene elements as part of the set if the offer does not include them. Check the image against the product description.

In AI catalogue recommendations, the image may later appear outside its original page. The guide to assessing catalogue recommendations (in Polish) explains how to review the cards. A shortened card still needs an identifiable status. If a concept becomes a thumbnail of a finished product, check the card without opening its details too. A reader may decide from that abbreviated view; an explanation found only later has not yet accompanied the choice.

If the process transfers materials between services, record the materials and tools involved. The AI data-flow mapping guide (in Polish) explains how to trace that route. The record should reflect the real flow, without listing unused platforms just in case. This helps establish where a resulting copy should be checked.

You can discuss your needs concerning AI in this workflow with Aeneia Audit and agree on a possible scope of cooperation. Prepare the source, one published version and a description of the stages connecting them for the conversation.

The recorded result should end in a specific decision

Separate data read, the technical verification result and assessment of use in the offer in the report. Each needs its own basis. 'The history includes generation' differs from 'The material is suitable to present as a concept'. The latter requires review of the publication context.

Distinguish 'checked and not read', 'not checked' and 'result needs explanation'. The first describes a test performed, the second an outstanding task and the third an open question. A shared 'missing' label hides these differences and makes the next action harder to choose. Assign a responsible person and a return date to each status.

Describe an open question so the next person knows what to seek: the source file, confirmation of an editing stage or permission for a particular use. Missing information should become neither automatic acceptance nor a confident accusation of improper origin. The team needs a route to complete the evidence or choose another image, depending on the importance of its role.

For doubtful material, choose an action suited to its role. Unclear provenance of a decorative background may lead to replacing it. Unclear representation of an actual project needs clarification before it is used as evidence of completed work. The report should show the difference rather than give every thumbnail in the archive the same priority.

For a correction, retain the previous version and reason. Changing a caption, replacing a file and adding a visible explanation are different actions. Acceptance should confirm what actually changed. Merely reopening the image in a tool does not prove the public card was updated.

Decide who reviews a new use of the same asset. An image accepted for a concept page may later appear in advertising or a catalogue. Earlier approval does not necessarily cover a different message. When placement changes, check status, caption and relationship to the offer rather than treating the file as unconditionally ready for anything.

Two or three examples can help the team: material with a readable history, material with a gap and material technically described correctly but placed misleadingly on a page. The comparison shows why one green indicator does not replace an editorial decision, without requiring a lecture on cryptography.

Well-organised provenance enables more reliable statements about the file and its journey. Accurate context enables reliable statements about what the image represents to a customer. Both are needed. A generated studio can then remain an interesting concept, without the company later having to explain why nobody can travel there for the scheduled meeting.

Aeneia / Aeneia editorial

Editorial article. Source links are included in the text.

About Aeneia

Brands we have worked with

  • L’Oréal
  • Garnier
  • Netflix
  • Carrefour
  • Warner Bros. Discovery
  • Biedronka
  • HBO Max
  • Wolt
  • Żabka
  • COSRX
  • Embassy of China in Poland
  • Mixa
Aeneia / Initium

From knowledge to a decision

Tell us what you want to put in order. We will agree on where to start.

Contact Aeneia

We partner with and use solutions from:

  • Google
  • OpenAI
  • Microsoft
  • Stripe
  • Anthropic
  • Linear