Spotting AI Images and Deepfakes

A photo used to be decent evidence that something happened. Not anymore. AI image generators produce photorealistic pictures of people, places, and events that never existed, and deepfake video and audio can put words in a real person's mouth. These show up in scams, political manipulation, fake news, and harassment. This lesson gives you a layered approach: quick visual checks first, then the more reliable move of tracing where an image actually came from.

The key mental shift, echoing earlier lessons: visual glitches are getting rarer as the technology improves, so context and provenance beat appearance. Where an image came from tells you more than how clean it looks.

What You'll Learn

• Visual red flags in AI images and deepfake videos, and why they are weakening
• How to use reverse image search to trace an image's origin
• What Content Credentials (C2PA) can and cannot tell you
• A practical order of operations for checking a suspicious image

Visual Red Flags (Weakening, but Still Useful)

Older AI images were full of giveaways. Newer ones have fixed many of them, so treat these as quick first-pass signals, not proof:

• Hands and fingers. Historically the biggest tell: extra fingers, missing or fused digits, hands in impossible positions. Newer models have improved at hands, so a normal-looking hand no longer clears an image.
• Text in the image. Signs, labels, and logos often come out as garbled, dream-like lettering that looks like text but spells nothing.
• Lighting and shadows that disagree. A face lit like a dark studio while the background is a sunny park. AI often fails at the physics of light bouncing between objects.
• Eyes and blinking in video. In some deepfakes a person blinks rarely or in a mechanical rhythm, and gaze can drift unnaturally. Top-tier fakes now mimic natural blinking, so this is a fading clue.
• Backgrounds that fall apart. Warped patterns, melting edges, repeated textures, jewelry or teeth that do not quite hold together, accessories that merge into skin.
• Too perfect. Flawless symmetry, plastic skin, and an image that looks more like an illustration of a moment than a captured one.

For video and audio deepfakes specifically, also watch for lips that do not sync to the words, a voice with no breaths or odd flat emotion, and edges around the face that shimmer or blur when the head turns.

The More Reliable Move: Reverse Image Search

Because visual tells are unreliable, the stronger technique is to find out where an image actually came from. Reverse image search lets you take a picture and ask the web, "where else does this appear, and when did it first show up?"

• Google Lens is broad and convenient. Upload or right-click an image and search to find visually similar images, the pages that use it, and often the original context. Great for a fast first pass.
• TinEye specializes in tracing an exact image across edits and crops. Its standout feature is sorting results by oldest first, which helps you find the earliest appearance, often revealing that a "breaking news" photo is actually years old or from an unrelated event.

What you are looking for:

• A mismatch in context. An image presented as a recent event that reverse search shows was published years ago, or somewhere else entirely, is being recycled to mislead. This is one of the most common forms of misinformation, and it usually involves real photos used falsely, not AI fakes at all.
• An original source. Tracing back to a credible outlet, a real photographer, or an official account raises trust. Finding only anonymous reposts lowers it.
• No trace at all. A dramatic, supposedly newsworthy image that appears nowhere reputable is a reason for suspicion.

Content Credentials (C2PA): Helpful but Limited

You may start seeing Content Credentials, an open standard called C2PA that attaches a signed record to a file describing who made it, when, and with what tools. Some cameras, editing software, and AI generators now add these, and some platforms display them. When present and valid, they are useful provenance.

But understand the limits clearly, because they are easy to misread:

• Missing credentials are not proof of anything. Most images online carry no credential. Absence does not mean an image is fake, and it does not mean it is real.
• A valid credential does not prove truth. It can confirm a file was made by a certain tool and not altered since signing. It does not prove the image is fair, in the right context, or honestly captioned.
• Credentials can be stripped. Screenshotting or re-saving an image often removes the data entirely.

So treat Content Credentials as a bonus signal when you find one, not a detector you can rely on. The same caution applies to "AI image detector" websites: they give probabilities, not proof, and they make mistakes in both directions.

Practical Order of Operations

When an image matters, work from quick-and-weak to slow-and-strong:

1. Glance for glitches. Hands, text, lighting, warped backgrounds. Found something obvious? Strong reason for doubt. Found nothing? You have learned almost nothing yet.
2. Reverse image search. Use Google Lens for context and TinEye for the oldest appearance. This is your most reliable single step.
3. Check the context. Does the date, location, and event match how the image is being presented? Recycled real photos are more common than AI fakes.
4. Look for provenance. A valid Content Credential or a trace to a credible original source adds confidence. Their absence is not a verdict.
5. Corroborate. If the image supports an important claim, confirm the claim through independent reporting, which the next lessons cover.

Key Takeaways

• Visual red flags (bad hands, garbled text, mismatched lighting, warped backgrounds, unnatural blinking) are useful first-pass clues but are weakening as models improve.
• Reverse image search with Google Lens and TinEye is the more reliable move; sorting by oldest can expose recycled or out-of-context photos.
• A lot of image-based misinformation is real photos used in the wrong context, not AI fakes, so always check date and place.
• Content Credentials (C2PA) are helpful provenance when present and valid, but missing credentials prove nothing and a valid one does not prove the image is truthful.