What Image Generation Is For

Image generation improved faster than almost any other capability and remains poorly matched to the tasks people first reach for.

Where it works

Illustration and mood. Header images, backgrounds, concept art, atmosphere. Where the requirement is "something that feels like this" rather than "this specific thing".

Iteration on an idea. Twenty variations of a composition in minutes, to find a direction before commissioning real work.

Placeholder content during design.

Textures, patterns, abstract material.

Editing existing images — removing an object, extending a background, changing a sky. This is where the technology is most reliable and least discussed, and it is built into most photo software now.

Style exploration for a project brief.

Where it fails

Anything requiring precision. Text in images has improved and is still unreliable. Diagrams with correct relationships. Charts with real data. Technical illustration where accuracy matters.

Consistency across images. The same character, product or setting rendered repeatedly. Various techniques help and none makes it dependable.

Specific real things. A particular building, product or location will be approximated rather than depicted.

Counting. Ask for five of something and you get four or six.

Hands, still, though much less than before.

Anything where being nearly right is wrong. Instructional images, safety material, product photography.

The legal position, which is unsettled

Copyright in the output varies by jurisdiction. Several, including the US, have taken the position that purely machine-generated images without meaningful human authorship are not eligible for copyright protection. If you need to own an image, this matters.

Training data litigation is ongoing and the outcomes are not settled. Commercial use of generated images carries an unquantified risk that a court decision could change.

Some providers offer indemnification for enterprise customers. That is a meaningful signal about their confidence and a partial mitigation.

Style imitation is contested. Generating in the style of a named living artist is technically straightforward, legally unclear, and ethically objected to by a great many working artists. Prompting by artist name is the practice most likely to be restricted.

Trademarks and likenesses are unchanged. Generating a recognisable brand or a real person's face does not become permissible because a model made it.

For commercial work: check the provider's terms on ownership and indemnification, avoid naming living artists, avoid recognisable people and brands, and get advice if the image is central to a product.

The uses that harm people

Worth naming plainly, because the tools make them trivial.

Non-consensual intimate imagery. The largest documented harm from this technology by volume, overwhelmingly aimed at women and girls. Illegal in a growing number of jurisdictions and unambiguously wrong everywhere.

Impersonation. Images of real people saying or doing things they did not.

Fabricated evidence. Images presented as documentation of events.

Fake product photography for goods that do not exist.

Provider filters block some of this and open models have no filter. The constraint is the user, not the tool.

Disclosure

Norms are forming and are inconsistent. Several platforms now require or apply labels. Some jurisdictions are legislating disclosure requirements.

A reasonable default: label generated images in journalistic, documentary and educational contexts, where the reader's assumption is that a photograph depicts something that happened. Decorative and illustrative use is a weaker case.

Metadata standards exist for recording provenance and are inconsistently applied and easily stripped.

If in doubt, say so. The cost of a label is low; the cost of being found to have passed off a generated image as a photograph is high.

Practical guidance

Describe the image, not the prompt style. Composition, subject, lighting, mood, framing. Long strings of quality keywords matter much less than they did.

Iterate rather than perfecting a prompt. Generate, look, adjust.

Use editing rather than regeneration when one element is wrong. Most tools now support this and it is far more efficient.

Do not use it where precision matters. Diagrams, instructions and technical illustration should be made properly.

Practical prompting now

The elaborate prompt engineering of a few years ago is largely obsolete, and a few things still help.

Describe the image, not the keywords. A sentence about what is in the frame, how it is lit and how it is composed beats a string of quality adjectives.

Specify the framing. Close-up, wide, aerial, eye level. This changes the result more than most adjectives.

Specify the light. Overcast, hard sunlight, backlit, single lamp. Lighting is what makes generated images look convincing or not.

Say what you do not want where the tool supports it.

Iterate on one variable at a time. Changing five things means you learn nothing about which change mattered.

Use editing for local fixes. Regenerating an entire image because one element is wrong is slower and loses everything you liked.

Avoid naming living artists. Legally contested, ethically objected to, and increasingly restricted by providers.