Forum

AI Art Critique When the Model Checks Its Own Work

Two open wooden shutters framing an iron-barred window that looks down a bright corridor with a small spiral staircase

Two of the best-known names in AI imagery made the same quiet move last week. On 6 October, Google made Nano Banana 2.1 generally available, an update to the image model behind Gemini, listed in its developer documentation as gemini-nano-banana-2.1. The next day, Midjourney added a “thinking” option to its alpha website that sends an existing image back through the model for a second pass. Neither announcement was dramatic. Together they point to a change that matters to anyone who judges images for a living: the model has started to review its own work, which makes human AI art critique more important, not less.

What the two updates actually do

Midjourney’s alpha changelog of 7 October says that jobs made with V8.2, in both standard and edit modes, can now be rerun with “thinking”. According to the changelog, the model checks the image for places where it missed the prompt, such as wrong objects, layout, anatomy or text, and then generates again. A “Rerun (Thinking)” button in the lightbox starts the process. Midjourney calls it an early version and asks users for feedback.

Google’s release is a different kind of change. Nano Banana 2.1 updates Nano Banana 2 and is documented with output at 1K, 2K and 4K, conversational editing and up to 14 reference images. Reports on the release describe gains in mask-based and sketch-based editing and in keeping several characters consistent from one image to the next. In both cases the direction is the same: fewer visible mistakes, and less need for a person to catch them.

Why self-review changes the job

For the past three years, the most common human task around generative images has been repair. A person spots the extra finger, the garbled sign or the chair with five legs, then reruns the prompt or fixes the image by hand. That labour was easy to explain and easy to measure. If the model now finds and repairs many of those faults itself, the visible errors that used to separate careful work from careless work will become rarer.

What remains is the judgement that a model cannot supply for itself. A model that checks an image against a prompt is only checking whether the image matches the words. It cannot say whether the words were the right brief, whether the composition serves the message, or whether a polished image is also an honest one. Those questions belong to the person who commissioned or signed the work.

What good AI art critique looks like

Good AI art critique starts where the model’s review stops. It asks what was intended, what the image communicates and what it leaves out, and it treats technical cleanliness as the minimum rather than the achievement. It also asks where the image came from, because a flawless result with no account of its making tells a viewer very little.

Critics, art directors and competition judges have often relied on flaws as clues. A broken hand or a melted letter suggested that a machine was involved and that no one had looked closely. As those clues fade, the questions have to move from the surface to the process. An earlier piece on a Hasselblad Masters finalist flagged for a garbled label made the same point: spotting AI by its mistakes will not last.

Practical steps for artists and designers

  • Write down the brief before generating, so that the finished image can be judged against intent and not only against the prompt.
  • Keep a short log of the tool, the version, the number of reruns and any edits, including whether a thinking or review pass was used.
  • Review every result as a careful stranger would, asking what the image and its origin seem to say.
  • Ask a colleague who did not see the prompt to describe what they see, and compare that with your intent.
  • Keep editable source files, so that a judge or client can see how the work was made.

The award perspective

The International AI Award Rules page asks entrants to explain how AI entered their process, which models and tools they used and what choices they made. As image models review themselves, that account carries more weight, because it is the part of an entry that shows the entrant’s own judgement. The earlier piece on Nano Banana 2 and the end of the Thaler case looked at why that human contribution also matters for authorship.

The takeaway

A model that can say “I missed the prompt” and try again is useful, and it will make a great many images look more finished. It will not decide what an image is for. As these tools take over the repair work, the scarce skill for artists and designers is critique: knowing what to ask of a result, what to change and when to stop.