Bad AI art explained: why AI images go wrong and why it matters in 2026

Look closely at the AI image someone shared in your feed this week. The light is perfect, the colours are rich, and the woman in the café is holding her cup with six fingers. The sign behind her says something that looks like English until you try to read it. That's what most people mean by bad AI art: images that look polished at a glance and fall apart the moment you look properly. Honestly, I find the sign the creepier detail.

The short answer is that AI art goes wrong in predictable ways, and it's controversial for reasons that have little to do with fingers. Image generators predict what pixels should look like rather than understanding the scene, so hands, text and physics break first. The bigger debate is about training data, artists' livelihoods and floods of low-effort content, and on those questions the honest answer's still "it depends".

Short answer: is bad AI art a sign that AI is bad?

Not on its own. Glitchy hands and garbled signs are technical limits that are shrinking with each new model.

Whether AI image tools are bad depends far more on how they were built and how they're used: with consent, credit and disclosure, or without. I think it helps to keep those two questions apart, because mixing them makes both harder to answer.

The mechanics behind the glitches

Most image generators are trained on huge collections of pictures paired with captions, and from these they learn statistical links between words and visual patterns. Type a prompt and the model starts from noise, then shapes it step by step into something that matches those patterns. Runway, which makes video and image models, puts it plainly in its own guide: "An AI image generator is essentially making an educated guess."

A bicycle sketched from memory shows the problem. Most people draw two wheels, a seat and handlebars, and get the frame wrong, because they've seen thousands of bikes without ever studying how the parts connect. Image models have the same problem at a massive scale: they've seen countless hands, but not the rule that a hand has five fingers attached in a fixed way.

Hands and bodies

In a piece for The Conversation, researchers explained that hands in training images are often small, holding objects or partly hidden, so it's hard for a model to tie the word "hand" to one exact five-fingered shape. Runway adds a second reason: we're very sensitive to errors in hands and faces, so a flaw we would ignore in a cloud becomes obvious in a person. That seems right to me.

Text and counting

To an image model, letters are just lines and shapes. The same Conversation article notes that text and quantities need far more precise associations and more training data than most visual tasks, which is why a prompt for "four apples" can return five and a shop sign turns into nonsense.

Physics and the "too AI" look

Models learn which objects tend to appear together, not the rules that govern them. That's how you get a chair leg that doesn't reach the floor or a necklace passing through a shirt, examples Runway lists in its guide to why AI images look weird. The same guide describes a subtler tell: images that are technically fine but oddly glossy and too polished, because models drift toward an average version of whatever you asked for.

Common failure	Why it happens	Quick check
Extra or fused fingers	Hands are complex, often hidden in training images	Count fingers, look at where hands meet objects
Gibberish text	Letters are treated as shapes, not symbols	Try to read every sign, label and logo
Wrong numbers of things	No real concept of "four"	Count repeated items such as windows or teeth
Impossible physics	Learns co-occurrence, not physical rules	Check shadows, reflections and things that touch
Plastic, generic look	Output drifts toward the average	Look for over-smooth skin and perfect lighting

None of these checks is proof on its own. Real photos have odd shadows too, and newer models avoid many of these errors, so treat the list as a hint. If you want a longer guide to spotting generated pictures, see our explainer on how to tell if an image is AI.

Real examples of bad AI generated images

The best-known example is no single picture but a genre. Merriam-Webster chose "slop" as its 2025 word of the year and defined it as "digital content of low quality that is produced usually in quantity by means of artificial intelligence." Engadget, reporting the choice, pointed to fake movie trailers, AI-generated bands and AI images turning up on food delivery apps, and noted that TikTok and Pinterest began letting users tone down AI content in their feeds.

Most bad AI art isn't bad in a dramatic way. It's a stock-looking illustration on a blog, a menu photo of food that doesn't exist, or a poster with a misspelled headline. Each one is minor. Together they probably make people trust images less, which is a real cost even when no single picture causes harm.

Progress deserves a fair mention too. The text problem, once the easiest tell, is fading. PCWorld found that Google's Nano Banana 2 could put readable text on billboards, newspapers and signs, though longer article text still began to crumble when the reviewer zoomed in. Runway says much the same: newer models handle short headlines well, but longer, smaller or denser text is still where errors creep in.

Benefits and limitations: the debate, fairly

Ask "is AI bad?" about art and people usually mean one of three different things: quality, consent or ownership.

Quality

This is the weakest argument against AI art, I think, because it's the one most likely to age. Hands improved, then text improved. If your objection is only that the pictures look wrong, the next model may answer it. Low-effort output will stick around, but so do bad clip art and bad stock photography.

Training data and consent

This is the strongest argument, and it's unresolved. Many artists object that their work was used to train commercial models without permission or payment. The first big US case, Andersen v. Stability AI, brought by visual artists against Stability AI, Midjourney, DeviantArt and Runway, is still in discovery.

Its summary judgment hearing was pushed to February 17, 2027, and a later schedule change means those motions may not be resolved until the end of 2027.

In the UK, the High Court ruled in November 2025 in Getty Images v Stability AI. Law firm Burges Salmon describes a mixed outcome that largely favoured Stability AI: a limited trade mark win for Getty, but its secondary copyright claim failed, partly because the model was trained outside the UK. Getty has been granted permission to appeal. So, in late 2026, no court has given a final answer on whether training on artists' work without a licence is lawful.

Ownership

Here the law's clearer. In March 2026 the US Supreme Court declined to hear Thaler v. Perlmutter, leaving in place the rule that copyright needs a human author. Mayer Brown notes the US Copyright Office's position that "[p]rompts essentially function as instructions that convey unprotectable ideas". In plain terms: an image made from a prompt alone probably isn't yours to protect in the US, though human edits and arrangement can be.

Our own bias

There's an uncomfortable twist. In four experiments reported by PsyPost, people rated the same artwork as less creative and less awe-inspiring when told it was made by AI, and the effect was stronger in people who believe creativity is uniquely human. In one experiment, both paintings were AI-made; only the label differed. So some of what we call bad AI art is a judgment about the source, not the image. That doesn't make the objections wrong, but I'd keep it in mind.

On the other side of the ledger, the benefits are real. These tools let people who can't draw sketch ideas, mock up products, make storyboards and explore styles quickly. For many small businesses, they replace no artist at all, because there was never a budget for one. Where I land: the technical flaws are a passing problem, but the questions about consent and pay aren't, and they deserve real answers rather than a better model.

Practical takeaways

Whether you use these tools or just scroll past their output, a few habits help.

• Check before you share. Read the text, count the fingers, look at shadows. If an image seems designed to make you angry, slow down.
• Disclose when you use it. Labelling AI images keeps trust intact, and many platforms and clients now expect it.
• Fix text outside the generator. Runway's own advice is to add exact text afterwards in a design tool, which works better than fighting the model.
• Don't assume you own it. If the work matters commercially, add meaningful human work and check the tool's terms.
• Avoid imitating living artists by name. It's legally uncertain and, for many people in the creative world, a clear ethical line.

If you do want better results, our guide to writing AI photo prompts and our comparison of the best AI image generators are good places to start.

Next up: the court cases

What we know: the obvious glitches are fading, slop is a widely named problem, and US law requires a human author for copyright. What we're guessing: whether courts will treat training on copyrighted art as fair use. The Andersen case and Getty's appeal are the ones to watch, and neither is likely to end before 2027.

My best guess is that "bad AI art" will soon mean less about six-fingered hands and more about context: images made without consent, without disclosure, or in such volume that they crowd out everything else.

For a wider look at the risks, read our pieces on why AI is bad and whether AI is good or bad.

FAQ

Is AI bad?

AI isn't bad in itself, but it has real downsides. Image tools can produce misleading pictures, flood feeds with low-quality content and were often trained on artists' work without permission. Our article on the dangers of AI goes further.

Is AI bad to use?

Using AI isn't wrong by default. It becomes a problem when you pass off generated work as hand-made, copy a living artist's style for profit, or share fake images as real. Using it for brainstorming, mock-ups or personal projects, and saying so when it matters, is widely accepted. Check your school, employer or client rules first.

Why does AI art have weird hands?

Hands are complex, jointed and often partly hidden in training photos, so models struggle to learn one consistent five-fingered shape. We notice hand errors more than mistakes in backgrounds. Newer models have improved, but hands holding objects or overlapping each other still go wrong more often than other details.

Can you copyright AI art?

In the US, not if it's made by prompts alone. Courts and the Copyright Office require a human author, and the Supreme Court declined to revisit that in March 2026. Human contributions, such as substantial edits or the arrangement of AI elements in a larger work, can still be protected. Rules differ in other countries.