AI generated image: how to recognize it and verify its authenticity?

AI generated image

AI generated image: how to recognise it and verify its authenticity?

Is a photograph of water damage enough to substantiate an insurance claim? Does a picture of a damaged parcel actually show the product that was delivered? Does a photographed invoice reflect a genuine expense? An AI-generated image can depict an entirely fictional scene. AI can also alter an existing photograph by adding a crack, exaggerating damage or replacing information in a supporting document. For businesses relying on these images as evidence, appearance alone is not enough. They need to verify what the images show, look for signs of AI involvement and assess whether they support the claim being made. With AI creating image content and modifying existing photographs, what traces can help identify its use?

How to detect AI-generated images in 4 steps?

The use of AI to falsify supporting documents is already a reality. According to the Finovox France 2026 barometer, 10% of French respondents who admitted to having committed document fraud said they had used generative AI software.

When checking a photograph of an insured loss, a damaged product or a supporting document, these four steps help move from inspecting the image to verifying the facts.

1. Check the image’s provenance

Before examining the details, establish what needs to be verified. Are you trying to confirm that damage exists, identify a product or establish the circumstances of an incident?

For example, a photograph of a cracked screen should allow the damage to be linked to the specific product involved in the claim.

Initial checks include:

  • Requesting the original file, rather than a screenshot or an image embedded in a document.
  • Identifying the photographed product using a reference or serial number, where visible.
  • Establishing when and in what circumstances the photograph was taken, and how it relates to the reported incident.

The original file may retain information that is useful for analysis. However, missing information does not prove manipulation: some platforms remove metadata or compress images during transmission.

2. Examine the consistency of the scene and the damage

Visual inspection should answer specific questions relevant to what the image depicts.

For a broken screen, this could mean examining the continuity of the cracks, their interaction with reflections and their consistency across several photographs. For water damage, it could involve comparing visible marks on walls, objects and other views of the room.

Points to examine include:

  • Edges and transitions: does the damage blend into the object, or are there unusual transitions?
  • Shadows and reflections: are they consistent with the lighting and geometry of the scene?
  • Identifying details: do logos, text and reference numbers remain consistent?
  • Different views: do they show the same object and the same damage?

These observations may reveal inconsistencies. However, they cannot establish on their own that a picture was generated by AI.

The example below shows how AI can modify a photograph to simulate damage: on the left, the original scene; on the right, the same room with artificially added damage.

Original photo
Original photo of a room with an intact window.
AI-generated picture
AI-generated picture showing a destroyed window and artificially added damage.

3. Detect signs of editing and AI generation

Technical examination complements visual inspection. It aims to distinguish localised editing from the generation of an entirely new image.

Depending on the file and the tools available, the analysis may examine:

  • Metadata, to review retained information about how the image was created or processed.
  • Differences between areas of the image, particularly in digital noise, sharpness or compression.
  • Characteristic traces of generation, using specialised detection models.
  • Invisible watermarks embedded by certain generators to identify their outputs.

These checks do not all answer the same question.

Evidence of editing may indicate that an image was modified without identifying the tool used. A recognised watermark, on the other hand, may provide information about its generative origin.

4. Compare the image with the other documents in the file

The photograph should then be cross-checked against the invoice, incident report, repair estimate or expert report. The aim is to establish whether these documents describe the same item and the same damage.

  • Purchase invoice: do the brand, model and product reference in the photograph match the invoiced item?
  • Incident report: do the nature and location of the visible damage match what was reported?
  • Repair estimate: do the parts to be replaced or surfaces to be repaired correspond to the photographed damage?
  • Expert report: do the expert’s findings align with what the image shows?

For example, if a photograph shows a cracked screen but the repair estimate concerns a different model, this discrepancy should be explained before the image is accepted as supporting evidence. It does not prove AI generation, but it reveals an inconsistency between the documents.

Such comparisons help reveal inconsistencies within a file. To identify AI involvement in the image itself, invisible digital watermarks such as SynthID can provide additional evidence. How do they work, and what can they verify?

How can SynthID help identify an AI-generated picture?

Invisible watermarks are a distinctive form of technical evidence: they are deliberately embedded in generated content so that it can be recognised. SynthID is one example.

Developed by Google DeepMind, SynthID is an invisible watermark embedded in AI-generated content to enable its identification. By October 2025, the technology had already been used to watermark more than 10 billion still images and video frames across Google’s services.

A watermark embedded in the image’s pixels

Unlike a visible watermark placed over a photograph, SynthID embeds an imperceptible signal directly into the image’s pixels. This signal helps identify content created or modified by AI tools that use the technology.

The watermark is part of the image itself, rather than information stored in the file’s metadata. It is designed to withstand common changes, such as compression or the application of filters, although detection is not guaranteed after every transformation.

What does detecting SynthID establish?

The presence of a SynthID watermark indicates that an image was created or modified by an AI tool that uses this technology. It does not mean that the image is fraudulent: AI editing can have legitimate uses.

Conversely, the absence of a detectable watermark does not prove that an image is authentic. Some generators do not use SynthID, and certain transformations can make its watermark harder to detect.

So how can these invisible watermarks be detected, and how can attempts to remove them be identified? Both challenges are being explored by Finovox’s R&D team.

How is Finovox advancing AI generated image detection?

Through the work of its R&D teams, Finovox can now detect certain invisible watermarks, including SynthID, in AI-generated images. These marks help identify an artificial origin and, depending on the watermarking system, the generator used.

Some fraudsters attempt to remove these watermarks to conceal AI use. In some of these cases, Finovox can identify traces of their removal. Detection may therefore reveal an attempt at concealment, strengthening suspicions of fraud.

Turning research into practical checks for clients

To develop these capabilities, the team monitors scientific publications, studies model architectures and explores image characteristics and the internal layers of PDF files.

Each method is then tested under conditions that reflect client workflows: does it remain effective after compression or conversion? Does it flag files that have undergone legitimate processing? In which situations does it fail?

These experiments help the team retain sufficiently reliable approaches and discontinue those that do not deliver meaningful results.

“Taking a philosophical approach to R&D means being willing to question our assumptions. We need to explore, test and know when to stop pursuing an avenue that does not produce robust results. This judgement is what turns research into useful checks for our clients.”

Sami Touil, Lead R&D at Finovox

Advances that benefit multiple sectors

Unlike a business rule specific to a particular document or country, a method for detecting traces of AI generation may apply to several categories of images and documents.

A single advance can therefore enhance checks for multiple clients across different sectors and markets, within the method’s scope.

Do you receive photographs or supporting documents to assess requests or claims? Speak to Finovox about detecting content created or modified by AI.

Sommaire

Text Link