How to Tell If an Image Is AI-Generated

How to Tell If an Image Is AI-Generated: The Short Answer

Work from the outside in. Start with where the image came from, then check for provenance data, then look for earlier copies, and only then examine the pixels yourself or with a detector. Each step either answers the question or narrows it, and the cheap, reliable checks come first.

Be clear about what you are trying to establish. An image can be fully generated, partly edited with AI such as an object removed and filled in, or real but presented with a false caption. NIST AI 100-4 (November 2024) notes that partially synthetic images, where most content is real, are hard for every detection method, and whether an AI edit is substantial can be subjective.

Step 1: Check Where the Image Came From

Find the earliest post you can and look at the account behind it: how old it is, what else it shares, and whether it has a record of original photography. Then look for independent confirmation. A real photo of a news event is usually accompanied by other photos and videos from different angles and by reporting from outlets that were there. A single striking image with no corroboration, shared first by an anonymous account, deserves suspicion whatever it looks like.

Step 2: Look for Content Credentials and Watermarks

Upload the file to the Content Authenticity Initiative's Verify site at contentcredentials.org/verify. If the image carries Content Credentials, the C2PA standard's signed provenance records, you will see who signed it, which tools touched it, and whether a generative model was declared. A valid credential from a credible signer is strong evidence about origin. Google's Pixel 10, for example, attaches credentials to JPEG photos from Pixel Camera, and some AI image tools declare generation in theirs.

If you suspect a Google model, Google's SynthID Detector portal checks for its SynthID watermark; Google said in May 2025 that over 10 billion pieces of content had been watermarked. The portal detects only SynthID, so a negative result says nothing about images from other generators. Most images you meet online carry neither credentials nor a known watermark, because platforms strip metadata and many generators add none. Absence is not evidence either way.

Step 3: Run a Reverse Image Search

Search the image with two or more reverse image search engines. You may find the original photo the fake was built from, an earlier upload with a different caption, or the same face in a gallery of generated faces. FinCEN lists exactly that last case, an identity photo that matches an image in an online gallery of AI-generated faces, as a red flag for banks. Crop to the main subject and search again if the full image returns nothing.

Step 4: Inspect the Image for Visual Clues

Zoom in on the places generators historically struggle: hands and fingers, teeth, ears and jewellery that should be symmetrical, text on signs and clothing, reflections in glasses and windows, shadows that should fall the same way, and backgrounds where lines should be straight or patterns should repeat. Skin that looks airbrushed with no pores, hair that melts into the background and objects that merge into each other are further hints.

Treat these as hints only. Current generators fix many of them, and real photos processed by phone cameras, filters or compression can show similar artefacts. People are poor judges at scale: in iProov's February 2025 study of 2,000 UK and US consumers, only 0.1% correctly identified every real and fake item, and participants were 36% less likely to spot a synthetic video than a synthetic image. NIST AI 100-4 cites a 2024 study that found human performance near chance across visual and audiovisual content.

Step 5: Check Metadata and Run a Detector

Ordinary EXIF metadata, such as camera model and capture time, can support a story but is easy to edit and is usually stripped by social platforms, so it rarely settles anything. A deepfake detector analyses the pixels for traces generators leave. Free options such as the University at Buffalo's DeepFake-o-Meter let you compare several research models, but free detectors often disagree, and Deepfake-Eval-2024 found open-source models lost about 45% of their AUC on real 2024 images compared with older benchmarks.

Dedicated commercial detectors perform better on hard cases. On the public VendorBench-100 benchmark of 100 difficult still images, Neural Defend ranked first of 36 systems with MCC 0.876, ahead of Reality Defender at 0.646, while general chatbots scored lower, Gemini at 0.389 and OpenAI's GPT at 0.135. One VendorBench author also co-wrote Neural Defend's ARCAS 1B paper; the corpus and results are public, and the benchmark covers still images only.

How Businesses Check Whether Images Are AI-Generated

Manual checks do not scale to onboarding selfies, marketplace listings or insurance claim photos, and a fraudster's image will not come with a convenient caption. Businesses automate the content analysis and keep humans for context and edge cases. Neural Defend's NeuroVerify checks a face image, video or audio in one API call with six signals, including Face Manipulation Detection, AI Generative Model Analysis and AI Pixel-Level Manipulation, and returns one unified risk score with signal-level evidence in under 200 ms p95 on eligible image requests. NeuroVerify does not analyse ID documents.

For investigations, such as a disputed claim photo or a viral image a newsroom must verify, NeuroDetect stores the file in a case, records model and version context, shows explainability heatmaps over the regions that drove the score, and exports a forensic report with observations, uncertainty and provenance.