How CESOERA detection works

How CESOERA works

Each piece of content is analysed by CESOERA’s detection models, running on CESOERA’s own infrastructure. The analysis produces an AI-likelihood score from 0 to 100% and one of four indication levels. Where the content is analysed in parts - video frames, text passages, document sections - the result also shows which parts carried the stronger indications.

Image analysis

CESOERA analyses visual signals in the submitted image and estimates how strongly those signals align with patterns associated with AI-generated content. JPEG, PNG and WebP images are supported, uploaded directly or fetched from a link.

Editing, screenshots, resizing and heavy compression can alter some of the visual signals available for analysis, so an original file gives the clearest reading.

Some tools sign a declaration into the file - Content Credentials (C2PA) - saying it was made or edited with AI. CESOERA reads that declaration and reports it as a separate signal next to the visual score. When a file carries one, the verdict rests on the declaration and the report says so; the visual model’s own score is still shown beside it. Credentials are removed by screenshots, re-saving and most platform uploads, so a copy without them is judged on its pixels alone. A missing declaration is not evidence that content was made by a person.

Video analysis

CESOERA samples frames across a video and analyses those samples for visual signals associated with AI-generated content. Frames are taken about every five seconds - at least 8 and at most 48 per clip - and the overall result is the median of the sampled frames, so a single unusual frame does not decide it.

The result includes a timeline of the sampled frames. Each point corresponds to a sampled timestamp, not a continuous reading of every frame. Longer clips get more sampled frames and cost more credits; the cost is shown before a video is analysed.

Re-encoding, compression, resizing and the processing platforms apply on upload change a video’s pixels, and with them the visual signals available. The same clip can therefore read differently once it has been shared.

Conventional 3D animation and computer-generated imagery - animated films, rendered product videos, visual effects - can score as AI-generated even when no generative AI was involved. In our own testing on openly licensed animated films, about a third of sampled clips received some or strong indication, far more than ordinary camera footage. Treat a high score on rendered content with particular caution.

Text analysis

CESOERA evaluates passages for statistical patterns associated with AI-generated writing and combines those observations into the overall result. The text is read as overlapping passages, and the result weighs the most strongly indicated fifth of them - so an AI-written section inside otherwise human writing is not averaged away.

Every accepted text is analysed in full, up to 20,000 characters. One analysis covers up to 32 passages, which ordinary prose never reaches; very dense input - long tables of codes or numbers, or text written mostly in characters such as Chinese, Japanese or Korean - can. Such text is declined before anything is charged, with a request to send it in parts. CESOERA does not present partial analysis as complete analysis.

Document analysis

Through the Business API, CESOERA extracts the text sections of PDF, DOCX, XLSX and TXT files and analyses them while preserving useful locations - pages, sections, sheets or cells - so a result can say where in the file the stronger indications are. Scanned pages, images inside documents and handwriting are not read.

Documents are processed and discarded. Business detection history keeps the result and its locations, not the document or its text.

Understanding results

Each model’s score is placed on a common 0-100% scale whose bands correspond to the indication levels, so the percentage and the level always agree:

The bands are fitted separately for images, video and text against CESOERA’s own evaluation data, so “strong indication” carries comparable weight whichever kind of content it describes.

Content Credentials and provenance

When Content Credentials (C2PA) or compatible provenance information is present, CESOERA surfaces it separately from model-based analysis. The two answer different questions: provenance records what a tool declared about a file; the models assess the content itself.

Provenance metadata is often removed by exporting, screenshots, re-encoding and platform processing, so its absence says nothing about how content was created.

What can affect analysis

Detection signals can change when content is transformed after creation:

• compression and repeated re-saving, including by messaging apps and social platforms;
• resizing, cropping and screenshots, which replace the original pixels;
• re-encoding of video by editors and platforms;
• heavy editing, filters and overlays;
• very short text, which carries little statistical signal - text under 200 characters is not analysed.

Getting the strongest analysis

• Use the highest-quality original file available, rather than a screenshot or a re-shared copy.
• Avoid re-compressing or resizing content before submitting it.
• Submit complete text rather than short excerpts.
• For video, upload the original file where you have it; a platform link returns the platform’s re-encoded copy.
• Check Content Credentials where they exist, as a separate signal alongside the result.

Responsible interpretation

CESOERA results are designed to support content review and investigation. For high-impact decisions, combine detection results with the surrounding context and other available evidence. Academic-integrity decisions should include human review and relevant supporting evidence.

Formal terms of use are in the Terms.