What Image AI Can and Cannot Do

The concept

Image generation tools are improving quickly, but four specific weaknesses are still common enough across major tools that you should check for them every time an AI-generated image is headed somewhere that matters — a client deliverable, a published post, anything representing your organization.

Text rendering. Ask for an image containing readable words — a sign, a label, a book cover — and even strong models frequently produce garbled, misspelled, or nonsensically-shaped letters. This happens because the model is generating pixels that look like text based on visual patterns, not composing actual characters the way a word processor does. The fix is not to keep regenerating and hoping: it's to generate the image without text (or with a blank sign/label area) and add real text afterward in a design tool, where it will render correctly every time.

Hands and fine anatomical detail. Hands have a notoriously high number of plausible configurations (fingers overlapping, bending, holding objects at odd angles), and models trained on images where hands are frequently partially obscured or awkwardly posed sometimes produce an extra finger, a fused knuckle, or an anatomically impossible grip. This has improved substantially in recent models but still needs a visual check before anything with visible hands goes out, especially in close-up or detailed compositions.

Exact likeness and logo reproduction. Models can approximate a recognizable style associated with a brand or a public figure, but cannot reliably reproduce an exact logo, trademark, or a specific person's precise likeness on demand — results tend to be close-but-off in ways that are legally and practically risky to use as-is. If you need an exact brand mark, use the brand's real asset; don't ask an image model to recreate it.

Licensing and provenance. Most image generation tools cannot guarantee that a given output is free of identifiable influence from copyrighted training images, and the legal landscape around this is still unsettled and varies by tool, jurisdiction, and use case. For anything commercial — client work, products for sale, ad creative — check the specific tool's commercial-use terms and licensing guarantees before relying on generated images, rather than assuming "AI-generated" automatically means "safe to use commercially."

A fifth, subtler weakness worth watching for is consistency across a set. If you generate several images meant to go together — a product shown from different angles, a recurring character across a comic strip, or a set of icons for one product — each generation is an independent pass, and the model has no built-in memory of exactly how it rendered the subject last time. Small details (the exact shade of a color, the precise style of a line, a character's proportions) can drift from image to image even with an identical prompt, because of the randomness in the model's diffusion process. Some tools offer "seed" values or reference-image features specifically to reduce this drift, but even then, plan to review a full set side by side rather than approving each image in isolation.

Treat all five as a pre-publish checklist, not a reason to avoid image AI — for the vast majority of internal, conceptual, and non-text visual needs, these tools are fast and genuinely good. The skill is knowing exactly where the seams are so you catch them before a client or audience does.

Worked example

A learner generates a promotional image for an event that includes a sign in the background reading "GRAND OPENING." The AI's version renders the text as a garbled approximation — letters that are almost right, with one word partially fused and a letter backwards. Rather than regenerating repeatedly hoping for a clean result, the learner generates a second version with the sign left blank, then uses a basic design tool to add the real "GRAND OPENING" text in a clean font, matching the sign's perspective and lighting by eye. The result is a clean, usable final image where the AI handled the scene and a two-minute manual step handled the text.

The learner also generates an image of two people shaking hands for the same project and notices, on close inspection, one hand has an extra knuckle crease that reads oddly at full size — a flaw easy to miss at thumbnail size but visible once the image is used at full resolution. They flag it for a quick retouch rather than publishing it as-is.

For a third asset in the same campaign, the learner generates four icon-style images meant to sit side by side on a landing page. Comparing them, two use a noticeably thicker line weight than the other two, even though the prompt was identical each time. Rather than publishing the inconsistent set, the learner regenerates the two outliers with an added phrase specifying the exact line weight used in the other two, bringing the set into visual alignment before publishing.

Exercise

Generate one image that includes legible text (a sign, a label, a product box) and separately generate one image of a hand or a small group of people. Zoom in on both at full resolution and identify the specific flaw in each — garbled text, an anatomical oddity, or both.

For the text image, note what you would do instead (regenerate blank and add text manually, or accept it if the text is small/decorative and not meant to be read). For the hand image, note whether the flaw is subtle enough to need a close look or obvious enough to require a full regeneration. Finally, generate a small set of 3–4 images meant to look consistent together (a set of icons or a recurring object) and check them side by side for drift — note anything that would need a regenerate to match the rest of the set.

Recap

• Five known weak spots — text rendering, hands/fine detail, exact logo or likeness reproduction, licensing/provenance, and set-to-set consistency — need a human check before any AI image goes out commercially or client-facing.
• The fix for garbled text is usually not "regenerate again" but "generate without text, then add it in a design tool."
• "AI-generated" does not automatically mean "safe to use commercially" — check the specific tool's licensing terms, and review any multi-image set side by side rather than approving each image on its own.

Resources

• OpenAI: "DALL·E usage policies" — https://openai.com/policies/usage-policies
• Midjourney: Terms of Service (commercial use) — https://docs.midjourney.com/docs/terms-of-service
• U.S. Copyright Office: "Copyright and Artificial Intelligence" — https://www.copyright.gov/ai/