Skip to content
wexloAccessible by design
Customer portalFree scan
wexloFree scan
Customer portal

AI marking

How to see which part of a Wexlo report was written by AI.

Measurements, likely findings and AI indications each have their own label in your report.

For example, in the head of your report

<meta name="iptc.digitalsourcetype"
      content="http://cv.iptc.org/newscodes/digitalsourcetype/compositeWithTrainedAlgorithmicMedia">
<meta name="wexlo.ai-generated" content="true">
<meta name="wexlo.ai-generated-parts"
      content="summary finding-explanations">
<meta name="wexlo.deterministic-parts"
      content="findings severity counts criteria-matrix score verdict">

Recognise what an AI indication is.

AI indication: confirmation needed

Does the alt text describe this image?

The text seems generic. Check whether the description fits the function of this image.

Illustrative example. This is not an established error. AI indications only occur in paid scans. The AI only looks at the image itself if you turn on image analysis.

Reading it yourself

For the PDF, using ExifTool:

exiftool -XMP-iptcExt:DigitalSourceType report.pdf

For the HTML version, "view page source" in your browser is enough; search for digitalsourcetype.

In the PDF

In the document's XMP metadata, alongside the PDF/UA information that was already there.

In the HTML version

As <meta> lines in the document head.

Both carry the same value, so it does not matter which file you look at.

You will see one of these two values:

normal valuehttp://cv.iptc.org/newscodes/digitalsourcetype/compositeWithTrainedAlgorithmicMedia
The document mixes deterministic measurement results with text written by an AI model. This is the normal value for a Wexlo report.
no AI writing stephttp://cv.iptc.org/newscodes/digitalsourcetype/algorithmicMedia
No AI writing step ran on this scan. The text comes entirely from the rule engine.

Alongside that, there are fields naming which steps actually ran for that report and which parts stayed deterministic. A step that did not run is not listed.

What the model writes, and what it does not decide

Deterministic, from the rule engine

  • the findings
  • the severity of each finding
  • the counts
  • the criteria matrix
  • the overall verdict

The same site produces the same result.

Written by an AI model

  • the summary
  • the plain-language explanation, for each finding
  • the suggested fix
  • the developer task
  • the translation into your own language
  • on paid scans: observations about page fragments and about images

Those steps are identifiable as such in the report itself.

In short: the model puts into words what the measurement found. It does not decide what was found.

What this marking is not

There is no established standard today for marking AI-written text. Work is underway, but until it exists, every party picks its own approach.

  • not a seal of approval
  • not a signature
  • not proof that the document is unchanged

We use the IPTC vocabulary for the origin of digital media. It was written for images and video, and we apply it to a text document because it is the most widely supported vocabulary available today, and because the same terms sit underneath the C2PA standard. If a standard for text arrives, it may overturn this choice, and we will adjust.

So this marking is an honest description of how this document came about.

Want to be sure you have the original file?

To establish that you have exactly the file we delivered, use the verification in your portal: it shows the SHA-256 value of the file as we produced it.

Questions

Something not adding up, or want to know how a specific passage came about? Email us and we will look into it.