How to Tell If a Photo Is AI-Generated or a Deepfake (2026 Guide)

AI-generated photos and face deepfakes have become good enough that a single odd finger or blurry background is no longer proof. The better question is not “Can I spot the fake instantly?” but “Can I gather enough signals to avoid being fooled?” Treat the image like evidence. Look at the visual details, then check whether the story around the image makes sense.

Start with the claim, not the pixels

Before zooming in, write down what the image asks you to believe. Is it claiming a person was in a place, that a product exists, that a disaster happened, or that someone sent a screenshot? Deepfake detection is easier when you know the claim. A generated portrait used as a dating profile has different warning signs from an edited invoice or a fabricated news image.

Ask: Who posted it first? Is there a named photographer, original publication, or primary source? Does the caption give a location and date? If the image appears only in reposts, cropped memes, or anonymous screenshots, increase your skepticism even if the picture itself looks polished.

Check lighting, shadows, and reflections

Real photos usually have a consistent light source. Look for shadows that point in different directions, highlights that do not match the scene, or reflections that omit important objects. In portraits, compare catchlights in both eyes. They do not need to be identical, but their position and shape should make physical sense.

Reflections are especially revealing in glasses, mirrors, glossy tables, car windows, water, and phone screens. AI images may produce a beautiful reflection that is compositionally wrong: a missing camera, a duplicated light, unreadable signage, or a reflection that shows a different room than the main scene.

Inspect hands, teeth, ears, jewelry, and hair

Older AI images often failed at hands. In 2026, hands are better, but they still deserve attention because they involve overlapping fingers, occlusion, and fine anatomy. Count visible fingers, trace where each finger connects, and check whether rings, sleeves, or shadows interrupt the hand in impossible ways.

Teeth and ears can also expose problems. Watch for teeth that blend together, gums that vanish, mismatched earrings, warped ear rims, or hair that melts into glasses and skin. One small oddity is not enough to declare an image fake; cameras, compression, and bad lighting create weird artifacts too. Multiple anatomy errors in the same region are a stronger signal.

Look for background warping and repeated textures

Backgrounds are where many synthetic images hide their mistakes. Zoom into door frames, railings, bookshelves, tiles, fences, keyboards, crowd faces, and text on signs. Straight lines may bend near the subject. Patterns may repeat too cleanly. Letters may look like language at a glance but collapse into nonsense when read.

Also check object relationships. A chair leg should meet the floor. A cup should cast a believable shadow. A necklace should continue behind hair. If the image is a screenshot, inspect UI spacing, icons, timestamps, and notification badges. Fake screenshots often mix fonts, padding, or interface details that a real app would keep consistent.

Run a reverse image search

Reverse image search is one of the highest-value checks because it shifts the question from “Does this look fake?” to “Where else has this appeared?” Search the full image, then crop to the face, landmark, logo, or unusual object and search again. Different crops can uncover an older original, a stock-photo source, or previous fact checks.

Pay attention to dates. If the same “breaking news” image appeared years earlier in a different country, the image may be real but miscaptioned. That still matters: many harmful fakes are not fully AI-generated; they are authentic photos reused with a false story.

Check metadata and EXIF, but do not over-trust it

EXIF metadata can show camera model, capture time, GPS location, editing software, and export history. If a file claims to be a fresh phone photo but the metadata says it was exported from an editing app, that is a clue. If the metadata is missing entirely, that is not proof of manipulation because social platforms often strip it automatically.

Treat metadata as supporting evidence, not a verdict. Metadata can be removed, changed, or forged. Use it to find inconsistencies: a timestamp after the claimed event, a time zone that does not fit, dimensions that suggest cropping, or software tags that match an editing workflow.

Look for provenance and Content Credentials

Provenance tells you how a file was created and edited. Some cameras, publishers, and creative tools attach Content Credentials or similar records that can show whether an image came from a camera, was edited, or used generative AI. When available, provenance is more useful than trying to guess from pixels alone.

The absence of credentials does not mean an image is fake. The presence of credentials does not mean the claim is true. A real image can be miscaptioned, and a credentialed edit can still be misleading. Use provenance to strengthen or weaken your confidence, then compare it with the source and visual evidence.

Use AI detection as a second opinion

AI detectors can help identify patterns that are hard for humans to see, especially in compression noise, texture consistency, face blending, and generated detail. They should not be treated as courtroom proof. Good models provide reasons and confidence, not just a dramatic red-or-green label.

That is the approach behind Veridox: upload an image, get a plain-language verdict, and review the specific reasons before you decide what to do next. A detector is most useful when it helps you ask better follow-up questions.

A practical 10-point checklist

1. Save the original file or highest-resolution version you can access.
2. Clarify the claim: person, place, date, event, or transaction.
3. Search the image and key cropped regions with reverse image search.
4. Compare shadows, reflections, and catchlights for physical consistency.
5. Inspect hands, teeth, ears, hairlines, jewelry, and glasses.
6. Zoom into text, logos, background faces, and repeated patterns.
7. Check EXIF metadata for camera, date, GPS, dimensions, and software tags.
8. Look for provenance records such as Content Credentials.
9. Scan with an AI-assisted detector and read the reasons, not just the label.
10. Do not share, pay, hire, accuse, or escalate until enough signals agree.

What to do if the image still feels uncertain

Uncertainty is normal. The safest response is to slow the decision down. Ask the sender for the original file, another angle, a live video call, a timestamped photo, or a link to the primary source. For high-stakes situations such as hiring, payments, legal claims, journalism, or family emergencies, require independent confirmation before acting.

The goal is not to become a professional forensic analyst overnight. The goal is to build a repeatable habit: source first, pixels second, metadata third, provenance fourth, and AI-assisted review when the stakes are high. That workflow catches far more suspicious photos than instinct alone.