Can You Spot AI-Generated Text and Images? What Actually Works

You see a photo of a flooded street shared by a neighbor. Your child hands in an essay that sounds a bit too polished. A product review reads oddly smooth. You wonder: was this made by AI?

It’s a fair question, and the honest answer is that you usually can’t be sure. You can still get much better at weighing the evidence. This post covers why that’s hard, which clues are worth noticing, and where labels and watermarks fit in.

Why this is harder than it sounds

Picture a good forger copying handwriting. The more they practice, the fewer slips they make. AI tools work in a similar way. They’re built to produce text and images that look like what people make. Each improvement removes some of the old giveaways.

A few years ago, AI images often had obvious mistakes, like six fingers or melted-looking faces. Many of those flaws have faded. AI text used to repeat itself or wander off topic, and that has improved too. A trick that worked last year may not work now.

Text has an extra problem. A sentence typed by a person and the same sentence produced by AI are identical. Unless the tool deliberately watermarks its output, nothing in the words themselves reveals where they came from. Even watermarks only show up across a longer passage, and they get weaker when the text is reworded.

Why AI detectors are unreliable

Many websites offer to scan text and tell you whether AI wrote it. They’re tempting to use, but they have serious limits.

• They guess from style, not proof. Most detectors look for writing that seems predictable or very even. Plenty of people write that way naturally, especially in formal or simple language.
• They wrongly flag real people. Detectors have been known to flag human writing as AI-made. People writing in their second language and people who write in a plain, careful style seem especially likely to be caught. A wrong accusation can do real harm to a student or an employee.
• They’re easy to fool. A little rewording, by a person or by another tool, can change the result.
• Their confidence scores can mislead. “87% likely AI” sounds precise. It isn’t a measurement like a thermometer reading. It’s the tool’s best guess.

Image detectors share some of these problems. They may work on images from tools they were trained on and fail on newer ones. Cropping, compressing or screenshotting an image can throw them off.

One more thing to keep in mind: pasting someone else’s writing or private photos into an online checker shares that content with an outside service. If you’re a teacher checking a student’s work, your school’s privacy rules may also apply.

What’s well established: current detectors make mistakes in both directions, and you shouldn’t use one as the only evidence for an important decision.

What’s still debated: how much detectors will improve, and whether they can keep up as the tools that generate content get better. Researchers disagree on this.

Clues worth noticing (as hints, not proof)

None of these prove anything alone. Treat them like clues in a mystery. One clue means little. Several pointing the same way are worth a closer look.

In text

• Confident but vague. Lots of smooth sentences that don’t say anything specific. For example, a review of a bakery that praises “delightful flavors and a welcoming atmosphere” but never names a single pastry.
• Made-up details. AI can invent quotes, sources, book titles or facts that sound real. If something is cited, try to find it.
• A sudden change in voice. If a student who usually writes casually hands in something that reads like a corporate report, that’s worth a conversation. It doesn’t prove anything.
• Leftover instructions. Sometimes people paste AI output without reading it, so phrases like “As an AI…” or “Here’s a revised version” are left in.

In images

• Odd details on a closer look. Text on signs that turns into gibberish, jewelry that merges into skin, mismatched earrings, hands holding objects strangely.
• Physics that doesn’t add up. Shadows falling in different directions, reflections that don’t match, or patterns on tiles and fences that drift.
• Too perfect. Skin that looks airbrushed, lighting that’s oddly flawless, a scene that feels staged.

Keep in mind that real photos can have strange shadows and blurry text too, and editing apps can make anything look smooth.

Better questions than “is this AI?”

Often the more useful question is: can I trust this? That’s answerable even when the AI question isn’t.

• Where did it come from? Who first posted it? Is that a source you know?
• Can you find it elsewhere? For a news photo, check whether trusted outlets are reporting the same event. Many search tools let you search using an image to find where else it appears.
• Does the context fit? A dramatic photo with no date, place or photographer named deserves extra doubt.
• Can you just ask? With a student’s essay, ask them to talk you through their argument. Someone who did the thinking can usually explain it. That’s fairer than a detector score.

Labels, watermarks and “content credentials”

Some help is coming from the other direction. Instead of guessing after the fact, some companies attach information to content when it’s created.

There are two main approaches:

• Watermarks. Some AI tools mark the images, audio or text they produce. A few marks are visible, like a logo or badge in a corner. Others are hidden in the content itself and usually can only be checked with the maker’s own tool. There’s no single checker that works for all of them.
• Provenance labels. “Provenance” means a record of where something came from. Some cameras, editing apps and AI tools can attach a record that says how a file was made and changed. Some websites and social platforms show a small label when that information is present.

These are useful, but they have limits:

• Not everyone uses them. Plenty of AI tools don’t add any marks.
• They can be removed. Provenance records attached to a file are easy to lose: a screenshot or re-saving the file often strips them. Hidden watermarks are built into the content and are designed to survive screenshots, cropping and compression, but heavy editing can still weaken or remove them.
• A missing label proves nothing. Most real photos don’t carry these records either.

So if you see a label saying something was made with AI, that’s good information. If you don’t see a label, that doesn’t tell you the content is real.

The bottom line

You can’t reliably spot AI-generated content by eye, and detectors aren’t reliable enough to settle the question. What you can do:

1. Treat detector scores and visual clues as hints, never as proof.
2. Focus on whether the content is trustworthy: its source, its context, and whether others confirm it.
3. Pay attention to labels when they appear, and don’t read anything into their absence.
4. Before accusing someone, talk with them. Don’t rely on a tool.

Healthy doubt, a quick check of the source and a willingness to ask questions will protect you better than any detector.