Excellent at fifty options. Poor at the right one.

The genuine shift AI has made in design is the cost of options. Exploring ten visual directions used to be a real investment; now the exploration is nearly free and the expensive part is deciding. That is a useful change, because more exploration generally produces better outcomes — provided somebody competent is choosing.

Where it goes wrong is when generation is mistaken for design. Fifty generated images are fifty guesses at what you asked for, produced by something that does not know your customers, your competitors, or what the piece has to achieve. Selecting well requires knowing all three, and that is the part that has not been automated.

The unglamorous background work is where AI is least controversial and most useful: removing backgrounds, extending an image to a different aspect ratio, cleaning up a photograph, upscaling something supplied too small. Those are tasks with a visible correct answer, which is exactly the condition under which generative tools are reliable.

This is the refusal clients question most, so it is worth explaining properly. A logo is not primarily a picture — it is an asset you intend to own exclusively and, frequently, to register as a trademark. That requires being able to demonstrate origination and to establish that it does not conflict with existing marks.

Generated output complicates both. Ownership of AI output varies by tool and jurisdiction and remains unsettled. More practically, a generative model produces output derived from a large body of existing work, which makes similarity to an existing mark a genuine risk rather than a theoretical one — and you would discover it at the worst possible moment, after printing signage and filing an application.

For a logo, the cost difference between generated and drawn is small and the risk difference is not. So logos are drawn, and AI stays out of that particular file.