Now inviting all designers, artists, photographers, marketers to our Artist + Designer plan (4TB)

AI Art Fails That Made Us Wash Our Eyes

Here’s what happens when you type the prompt “art” into AI art generator Dall-E 2:

Pretty impressive, right? It might even make some graphic designers quake in their boots at the power and potential of AI text-to-image generators.

AI art tools such as Stable Diffusion and Midjourney have made significant advancements since their inception way back in 2014. For instance, in 2018, an artist collective named Obvious sold an AI-generated painting at Christie’s for $432,500. Fast forward to 2023, and AI art exhibits are suddenly everywhere — including the world’s largest AI art show in New York, featuring over 100 AI artists.

However, creating AI art that’s even halfway decent (and not the stuff of nightmares) requires learning another skill: the art of prompting. While humans have an intuitive understanding of the world — its shapes, sizes, and what looks natural versus unnatural — AI, well…doesn’t.

Generative AI tools don’t always quite get what’s required from prompts that would be obvious to us — often with hilarious (and sometimes terrifying) results.

Even with great prompts, it’s never guaranteed that AI won’t produce something utterly baffling — like the ten examples listed below. Read on to see some digital art gone horrifically wrong — plus some prompting tips to help you get better results from your AI artwork.

1. “Salmon in the River”

There’s definitely “salmon” fishy about this piece of AI art that gives a whole new meaning to the term “salmon run.”

2. “A Human Hand That Actually Looks Like a Hand With 5 Fingers”

Nothing to see here, just some totally normal pictures of totally normal human hands…

3. “A Cocky Expression”

Next time someone gets a little too cocky, you can take them down a peg or two by turning them into an actual rooster.

4. “Christopher Reeve”

Is it a bird? Is it a plane? No, it’s just another piece of terrifying AI art!

5. “Fishermen”

Ahh…nothing beats a nice, relaxing day of fishing. Unless you happen to bump into these guys on the riverbank.

6. “AI Garden Rakes”

These rakes “leaf” a lot to be desired.

7. “Practicing Some Yoga”

We’ve heard yoga is transformational, but this seems like a step too far…

8. “Man Sitting at Rocky Beach”

Now you know what to do when your butt starts hurting from sitting on all those rocks.

9. “Reverse Benjamin Button”

This AI art generator took TikTok’s “old age” filter to the next level by turning this man’s baby into his own personal Mr Miyagi.

10. “Camping

We now have high expectations for our next camping trip.

Why do AI image generators fail so spectacularly?

This is just a tiny slice of the many examples of artificial intelligence fails floating around the internet. But why do AI tools struggle so much with prompts that seem obvious to humans? There are a few different reasons — let’s explore some of them.

Limited understanding

AI art tools, like language models, rely on large datasets to generate responses. However, they don’t truly understand the context and meaning of the prompt the way humans do.

Instead, they recognize patterns in the data and generate responses based on those patterns. If the prompt is ambiguous or contains unusual phrasing, the AI might misinterpret it and produce an incorrect or nonsensical output.

For example, if you type in “basketball shoes” as a prompt, you might expect AI to generate a pair of sneakers. However, it’s just as likely you’ll get an image of a pair of shoes made of basketballs.

Lack of real-world experience

AI models are typically trained on datasets from the internet, books, articles, and other sources. However, they lack the real-world experience humans have, which can lead to a disconnect between what the AI generates and what is contextually appropriate in certain situations.

For example, a prompt like “a mouse eating cheese” could produce a picture of a mouse being eaten by a piece of cheese — instead of the other way round, as your human brain might expect.

Bias in training data

AI models can pick up and replicate biases present in their training data. If the training data contains biased or misleading information, the AI may generate outputs that reflect those biases. This can lead to politically incorrect or offensive responses.

Buzzfeed made headlines for all the wrong reasons in 2023 when it published and then deleted a post about “Barbies around the world.” The article was heavily criticized for perpetuating harmful stereotypes and exhibiting colorism — a form of discrimination that favors lighter skin tones over darker ones.

Overfitting or underfitting

AI models can suffer from overfitting (memorizing the training data without understanding the underlying concepts) or underfitting (failing to capture the complexities of the data). Both of these issues can result in inaccurate responses to prompts.

For instance, if an AI model learns about dogs using pictures that are 90% brown dogs, it may not recognize a black dog as a dog because, to the model, “brown” is an indispensable part of being a dog. Underfitting is the opposite — for example, if a model learns that a dog is anything with four legs, it’s going to produce some strange images.

Lack of common sense and reasoning

AI models lack inherent common sense and reasoning abilities. While they can generate impressive outputs in many cases, they struggle with tasks that require reasoning or understanding beyond the surface level.

So if you write a prompt like “a woman and a cat on a bed together,” you might be picturing a heartwarming scene of a female relaxing with her feline friend. However, the word “together” could confuse the AI, causing it to merge woman and cat into one monstrous being.

Prompting 101: how to ensure AI doesn’t utterly fudge your designs

Crafting AI prompts that give you the result you’re looking for is both an art and a science. Here are our top 10 tips for prompting like a pro:

1. Understand the limitations

Familiarize yourself with the limitations of the AI tool you’re using and recognize that AI models can sometimes produce inaccurate or nonsensical outputs — even with well-formed prompts.

2. Use clear and specific prompts

Frame your prompts in a clear and specific manner to reduce the chances of the AI misinterpreting or generating irrelevant responses. Avoid ambiguous language or open-ended questions that might lead to unexpected results, and re-read your prompt back to yourself before entering it to foresee possible alternative interpretations.

3. Experiment with the context

Test the AI tool with various prompts and evaluate its responses to understand its strengths and weaknesses better — this will help you get a sense of the type of content it can handle well.

4. Check for coherence

Ensure that the AI-generated output is coherent and follows a logical flow. If the response seems disjointed or lacks continuity, the AI might not have fully understood the prompt or may be struggling with the context.

5. Avoid sensitive or controversial topics

Be cautious when using AI tools to generate content related to sensitive or controversial subjects. AI models can inadvertently reproduce biases present in their training data, which may lead to inappropriate or offensive responses.

6. Review generated content

Always review the AI-generated content before using or sharing it. If the output seems off or nonsensical, consider rephrasing the prompt or trying a different approach.

7. Use reputable AI tools

Choose AI tools from reputable sources or platforms that have been well-tested and have a track record of producing reliable results. Look for user reviews and feedback to gauge the tool’s performance.

8. Report issues and provide feedback

If you encounter persistent issues with the AI tool generating nonsense, provide feedback to the developers or support team. Reporting problems can help them improve the model and address specific issues.

9. Combine AI with human review

When using AI-generated content for critical tasks, consider having a human reviewer double-check the outputs. Human judgment can catch inaccuracies or nonsensical content that the AI might miss.

10. Be cautious with creative tasks

For creative tasks like AI-generated art or storytelling, be aware that the output can be highly variable. AI models may produce interesting and imaginative content, but they can also generate bizarre or nonsensical creations.

Make AI work for you

Hopefully, by now, your mind is at ease that AI isn’t going to take your job anytime soon. However, that doesn’t mean it doesn’t have its uses.

For graphic designers, AI tools can streamline workflows, automate repetitive tasks, and even accelerate the creative process. Playing around with prompts and upleveling your prompting skills can generate interesting new ideas that can spark your imagination.

Of course, you’ll need a place to store all of your incredible AI art — and all the prompts you use to create it. And that’s where Playbook comes in.

With Playbook’s AI features, you can now create and store your AI art and prompts directly on the platform using Dall-E or Stable Diffusion. And with 4TB of storage, there’s (virtually) no limit to how many versions you can generate.

Learn more about Playbook’s AI features here.