Guides

The problems you actually run into when checking whether something is AI-generated, where it came from and whether it was altered. What can be established is written down next to what cannot.

Basics

• How to check whether an image is AI-generated What to look at first when deciding whether a photograph was generated, why looking with your eyes no longer works, and how far a check can actually take you.
• How to check a video for a deepfake Video carries more to check than an image does. How the container, the frames and the audio are each examined, and what compression takes away from you.
• Verifying C2PA Content Credentials What a C2PA signature tells you and what it does not. Valid and trusted are different words, and the difference is where forgery is caught.
• The limits of AI text detection Detecting AI writing is far weaker than analysing an image. Why that is, which texts it fails on most dangerously, and what we therefore refuse to do.
• Does metadata prove a photo is real? EXIF present means real, EXIF absent means fake — both halves are wrong. How far metadata can be trusted, with the measurements behind the answer.
• What an evidence ledger is Not one verdict but every signal that produced it. What each line records, and why evidence pointing the other way is kept as well.
• Checking a photo before you publish it What to check, in what order, before a submitted photograph runs — and how to record the check so you can explain it afterwards.
• Supported file formats Which image, document, text, audio and video formats can be read today, and what kind of evidence each of them tends to yield.
• Why a screenshot destroys the evidence Which layers of a file disappear the moment you capture the screen, and what you can reasonably expect from a capture afterwards.
• What inconclusive means Inconclusive is a real verdict, not a failure. The conditions that produce it, and what you can do with it once you have it.
• How AI labels and watermarks are read AI output is marked in three layers: metadata, invisible watermarks in the pixels, and visible text. Which of them can be read locally, and how much weight each one carries.
• Digital forensics: forgery, tampering and evidence analysis for files How to check whether a photo, video, voice recording or document was forged or altered — using provenance signatures, metadata, compression traces and hashes — and how to hand it on as evidence.

Digital forensics

• Image forgery detection: splicing, copy-move and region edits How splices, copy-move clones and region edits show up in JPEG grids, error level analysis (ELA) and noise — and what a tamper heatmap can and cannot tell you.
• Video tampering and deepfake forensics: frame and encoder traces How to tell whether a video was cut, spliced or face-swapped from its frame structure, encoder data and re-encoding traces — and what we can and cannot conclude.
• Audio forensics: edited recordings and AI voice-clone evidence How to check whether a recording was edited or a voice was cloned with AI using metadata, codec traces and our audio detector — and the limits of call recordings.
• Document forgery forensics: reading PDF, DOCX and HWP edit history How PDF incremental saves, DOCX revision IDs (rsid) and tracked changes, and HWP metadata and fonts reveal whether a contract or certificate was changed later.
• Metadata and EXIF forensics: capture time, location and editing software How to read capture time, device, GPS location and editing software from EXIF and XMP, which platforms strip what, and why metadata can be faked as well.
• Chain of custody and hashes: how to hand over digital evidence How to show a file was not changed with a SHA-256 hash, and hand over evidence that others can reproduce: original copies, notes, engine versions and signed PDF reports.

By generator

• How to tell if an image was made with Midjourney What a Midjourney image keeps and loses, what the detectors look at when there is no provenance signature, and how to read the result honestly.
• How to tell if an image was made with ChatGPT When the C2PA signature on ChatGPT and gpt-image images survives and when it disappears, and what decides the result for copies that lost it.
• How to tell if an image was made with Gemini (Nano Banana) Gemini and Nano Banana originals are caught well through provenance signatures and marks, while copies sent through messengers are still weak. Why, and what to do.
• How to tell if an image was made with FLUX or Stable Diffusion For open models like FLUX and Stable Diffusion, the prompt and workflow inside the PNG are decisive clues. What those traces are and how we judge once they are gone.
• How to check Seedream, ERNIE and Z-Image images How well our gate catches generators released since 2025 such as Seedream, ERNIE and Z-Image, generator by generator, and which marks they leave, stated plainly.
• How to tell if a Sora, Veo or Kling video is AI-generated Signatures and watermarks on AI videos from Sora, Veo, Kling and Hailuo, what re-uploads strip, how frames and sound are judged separately, and our gate result.
• How to tell if a voice is AI-generated or cloned Traces left by speech generation and voice cloning such as ElevenLabs, the limits on phone recordings and re-encoded files, and our gate rate and false-alarm budget.
• How to tell if ChatGPT, Claude or Gemini wrote a text How far detection of chatbot writing can be trusted, why translated, short and quote-heavy texts are easy to get wrong, with our gate result.

File formats

• What iPhone HEIC photos keep, and checking them for AI The capture information iPhone stores in HEIC, what JPEG conversion and messenger sending remove, and what that means for checking a photo for AI.
• Reading the AI prompt record left in a PNG What the prompt, seed and workflow records AI tools leave inside PNGs are, how to read them, when they get erased and how the judgement works afterwards.
• Checking PDF, DOCX and HWP documents written with AI When checking PDF, DOCX and HWP files for AI writing, what body text detection and document structure traces each say, and where they go wrong.

By use case

• Handling possibly AI-generated evidence in legal disputes When checking whether photos, audio or documents submitted in a dispute were generated or altered with AI, what to secure, how to record it and how to hand it over.
• Checking portfolios and assignments for AI in hiring How to read results when checking a candidate's portfolio images, written assignments or cover letter for AI, and what to observe so no one is rejected without grounds.
• Spotting AI-made product photos and listing images How to check whether photos on secondhand markets, shops or rental listings were really taken, or generated or retouched with AI, and how to judge the result.
• Checking insurance claim photos for AI edits When checking whether accident or damage photos were generated with AI or had damage exaggerated, what to look at and record, from securing originals to marking edited regions and the limits.
• What to watch for when checking student work for AI What detection results do and do not say when checking student assignments and writing for AI use, and how to use them without unfairly suspecting students.

Explainers

• Why messenger and social media copies lose AI evidence Why sending through messengers and social networks removes AI signatures and metadata and weakens detection, and how to get the original instead.
• How real photos edited with AI come out How real photos that went through AI edits such as generative fill or object removal appear on the AI generation and alteration axes, and what is caught or missed.

This is general guidance, not legal advice. A scan result is a probabilistic assessment, not proof.