AI Art Generation: What DALL-E, Midjourney, and Stable Diffusion Do

A fifth-grade teacher told me about the morning she walked into her classroom and found three students huddled around a tablet, generating increasingly elaborate fantasy illustrations. “They figured it out over the weekend,” she said. “The pictures were stunning. I didn’t know whether to be impressed or concerned.”

Both, probably. AI image generation has gone from research curiosity to everyday creative tool in under three years. The results are genuinely remarkable. The questions they raise — about what “creating” means, about whose art got used to train these systems, about what kids should be allowed to generate — are genuinely difficult.

Understanding the technology doesn’t answer the ethical questions. But it makes them much easier to discuss honestly.

Key Takeaways

• Diffusion models work by learning to reverse a noise-adding process — starting from random static and progressively refining toward an image matching a text prompt
• The training data for most commercial image generators included billions of images scraped from the internet without artist consent, which is the basis of ongoing litigation
• AI image tools are not “looking up” existing images — they generate new images from mathematical representations of patterns in training data
• Copyright for AI-generated images is an unresolved legal area; the US Copyright Office has consistently declined to grant copyright to purely AI-generated works
• Kids can use these tools for genuine creative work, but they need to understand what they’re building on

How Diffusion Models Actually Work

The first thing to understand is what image generators are not doing. They’re not searching a database and returning a matching image. They’re not combining existing images like a collage. They’re generating new pixel arrangements — genuinely new images that never existed before.

The mechanism is called diffusion. Here’s the core idea.

Take a real image. Add a small amount of random noise (static) to it. Then add more. Then more. After enough steps, the image is just pure noise — completely unrecognizable. A diffusion model is trained to reverse this process. Given a noisy image, it learns to predict what the image looked like with slightly less noise. Apply this prediction repeatedly, and you can go from pure noise all the way back to a coherent image.

Once trained, the generation process starts with pure random noise and iteratively denoises it — but guided by a text prompt. The text conditioning (added via a separate component called a text encoder) steers the denoising toward images that match the description. “A red fox in a moonlit forest” — the denoising process is constantly pulled toward pixel arrangements consistent with redness, fox-shaped forms, moonlight, and forest textures.

This was first demonstrated effectively in the 2020 paper “Denoising Diffusion Probabilistic Models” by Ho et al. at UC Berkeley, and was the foundation for DALL-E 2, Stable Diffusion, and Midjourney (Ho et al., 2020).

What “Latent Space” Means

A complication: most modern diffusion models don’t operate on full-resolution pixels. They operate in “latent space” — a compressed mathematical representation of images. A variational autoencoder compresses an image down to a much smaller representation (the “latent”), the diffusion process happens in that compressed space, and then a decoder expands the result back to full resolution. This is why these models are called “Latent Diffusion Models” — Stable Diffusion specifically got its name from this architecture (Rombach et al., 2022).

Latent space is also the reason the models can be run on consumer hardware. Doing diffusion at full resolution would require far more compute than doing it in a compressed latent space.

The Training Data Controversy

This is where the technical discussion intersects with genuine ethical complexity, and parents should understand it clearly.

The major image generation models were trained on datasets containing hundreds of millions to billions of images scraped from the internet. LAION-5B, a publicly available dataset used to train Stable Diffusion, contained approximately 5.85 billion image-text pairs sourced from web crawls. These images included artwork by living professional artists — often without their knowledge and never with their explicit consent.

Artists have filed multiple class-action lawsuits. Getty Images sued Stability AI over the use of its licensed photo library. A class action on behalf of visual artists alleged that Stable Diffusion reproduced their copyrighted works without compensation. As of mid-2026, these cases remain in litigation.

The central legal question: is training an AI on copyrighted images “fair use”? This is genuinely unsettled. Proponents argue training is transformative — the model doesn’t store copies, it learns statistical patterns. Opponents argue that the model can demonstrably produce images in specific artists’ styles, and that this constitutes reproduction of the creative elements of their work.

For parents, the practical implication: the AI art tools kids are using were trained on a foundation that is legally and ethically contested. That doesn’t mean kids shouldn’t use them, but it’s worth discussing honestly.

Tool	Primary Diffusion Model	Training Data Source	Known Controversies
DALL-E 3 (OpenAI)	Proprietary	OpenAI dataset (undisclosed)	Opt-out program for artists, limited transparency
Midjourney	Proprietary	Web-scraped + curated	Multiple artist lawsuits, no opt-out until 2023
Stable Diffusion	Open-source	LAION-5B (web-scraped)	Getty lawsuit, artist class action, fully open weights
Adobe Firefly	Proprietary	Adobe Stock (licensed)	Designed explicitly with licensed data; pays contributors
Google Imagen	Proprietary	Google dataset (undisclosed)	Limited public access; Google’s AI principles applied

Adobe Firefly is notable because it was specifically trained on licensed Adobe Stock images, with explicit compensation mechanisms for contributors. This is the closest thing to a “clean” training data approach in a major commercial product.

What AI Art Means for Creative Careers

Is AI replacing artists? The honest answer is: it’s changing what artists do, and the impact is uneven.

Illustrators who produced large volumes of commercial work (game assets, stock illustrations, book covers) are experiencing real competitive pressure. Stock illustration platforms have seen declining revenue per image as clients use AI generators for projects that previously required custom illustration.

But fine art, character design, concept art in collaborative media, and art that carries an artist’s specific creative vision remains in demand. The 2023 survey by AIGA found that 67% of professional designers had used AI tools in their workflow, most often for ideation and asset generation in early stages — not final deliverables (AIGA, 2023). The pattern emerging is similar to what happened with photography vs. painting: AI doesn’t eliminate artists, but it changes what professional artists do and reduces demand for certain kinds of routine commercial work.

For kids who love making art: these tools are a remarkable addition to a creative toolkit, not a replacement for learning to draw, compose, or design. Understanding both is more valuable than knowing only one.

How to Teach Your Kid About AI Art Generation

Ages 5–8: What Is “Creating”?

Don’t start with the technology — start with the question. Show your child a piece of art they love. Ask: “What do you think the artist was feeling when they made this?” Then show an AI-generated image. “Can a computer feel anything? If not — is this the same kind of creating?” There’s no right answer. The goal is to start developing the vocabulary for distinguishing human creative expression from pattern generation.

Ages 9–12: Use the Tool, Question the Tool

Let your child use an AI image generator (DALL-E via ChatGPT, or Adobe Firefly which uses licensed images). Have them create something they’re genuinely excited about. Then discuss: “What did you contribute? What did the AI contribute? Whose art went into the training?” Research one of the artists who is involved in a lawsuit against an AI company together and read their argument. This isn’t to make kids feel guilty — it’s to develop the complexity of thinking they’ll need as this technology evolves.

Ages 13+: Style Mimicry as a Learning Experiment

Have your teen try to use an AI image generator to replicate a specific artist’s style. Then have them try to achieve the same visual effect by studying and manually applying the elements of that style (composition, color palette, brushwork). The comparison is instructive: what did the AI capture? What did it miss? What did the manual practice teach that the AI couldn’t? This exercise develops both technical skill and critical understanding of what style actually is.

The question to ask: “If an AI learns from a million artists’ work to make a picture — who deserves credit for that picture?”

What to Watch For Over the Next 3 Months

Month 1: Watch for legal developments in the AI art lawsuits. Several cases are scheduled for rulings in late 2026 that could significantly reshape how training data is handled. The outcomes will directly affect which tools are legally sound to use.

Month 2: Try using Adobe Firefly as a comparison to tools with contested training data. Notice any quality differences. Ask your kid whether the training data provenance matters to them in their creative choices.

Month 3: If your kid is interested in digital art, encourage them to learn fundamentals independently of AI tools. Basic composition, color theory, and figure drawing are skills that make someone better at directing AI tools, not obsolete because of them.

Frequently Asked Questions

Is it legal for kids to use AI image generators?

Most major tools have terms of service requiring users to be 13+ (or 18+ without parental permission). Adobe Firefly requires an account with age verification. DALL-E via ChatGPT follows ChatGPT’s 13+ terms. The content generated may also be subject to restrictions — most tools prohibit generating explicit, violent, or deceptive content. Parents should review the terms of any tool their child uses.

Can kids copyright AI-generated art they created with a prompt?

The US Copyright Office has consistently ruled that purely AI-generated images cannot be copyrighted, because copyright requires human authorship. However, images where a human makes significant creative choices (editing, compositing, heavily directing the generation) may be eligible. This area is actively evolving — the 2023 Zarya of the Dawn case established that AI-generated elements of a comic book could not be copyrighted while the human-created elements could (US Copyright Office, 2023).

What’s the difference between Midjourney, DALL-E, and Stable Diffusion?

Midjourney is a commercial service accessed via Discord, known for particularly high-quality aesthetic outputs, especially for painterly and fantastical images. DALL-E is OpenAI’s tool, integrated into ChatGPT. Stable Diffusion is an open-source model that anyone can download and run on their own hardware, making it the most customizable but also the least restricted. Adobe Firefly is specifically designed for professional creative workflows.

Does using AI art tools hurt real artists?

For certain categories of commercial illustration, there’s real competitive impact. For fine art and concept work requiring distinct creative vision, the impact is much less clear. The debate is real, and both “AI is just another tool” and “this is theft at scale” contain elements of truth. Families can hold both realities simultaneously and make considered choices about which tools to support.

---

About the author Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Ho, J., Jain, A., & Abbeel, P. (2020). “Denoising Diffusion Probabilistic Models.” NeurIPS 2020. https://arxiv.org/abs/2006.11239
2. Rombach, R., Blattmann, A., Lorenz, D., et al. (2022). “High-Resolution Image Synthesis with Latent Diffusion Models.” CVPR 2022. https://arxiv.org/abs/2112.10752
3. U.S. Copyright Office. (2023). “Copyright and Artificial Intelligence: Zarya of the Dawn Decision.” https://www.copyright.gov/ai/
4. AIGA. (2023). “AI and Design: How Professional Designers Are Using AI Tools.” https://www.aiga.org/resources/ai-design-survey-2023
5. Schuhmann, C., Beaumont, R., Vencu, R., et al. (2022). “LAION-5B: An Open Large-Scale Dataset for Training Next Generation Image-Text Models.” NeurIPS 2022. https://arxiv.org/abs/2210.08402
6. MIT Technology Review. (2023). “The Race to Define Who Owns AI-Generated Art.” https://www.technologyreview.com/2023/09/14/1079102/ai-art-copyright/