Why might teaming up with AI help your writing but actually hurt your image-editing work?

This explores why human-AI pairs might do better on text work than on image work. The honest starting point is that the retrieved notes don't contain a finding that compares text and images directly. The closest evidence is a meta-analysis of over 100 studies, which found that human-AI combinations usually score *below* the better of the two working alone (an effect size of g = −0.23). The one exception was content-creation tasks, where the pairs pulled ahead When do human-AI teams actually outperform either alone?. The authors trace the difference to task structure rather than ability. That suggests a different way to read the question: the medium may matter less than whether the human is *making* something or *judging* something.

That distinction lines up with how text and image tasks are typically set up in studies. Text tasks are often generative: drafting, rewriting, brainstorming. A human can take an AI draft and bend it toward a goal. Image tasks in this research are often classification or detection: is this real, is this a tumor, which category does it belong to? In those tasks the human's job is to accept or override the AI's call. Each override is a chance to replace a correct machine answer with a confident human mistake. Under this reading, the text/image gap is really the creation/decision gap.

A second clue is how badly people judge visual material that AI touches. A 30-study review found that human accuracy at telling AI-made content from human-made content sits around chance across text, images and voice Can people reliably spot content made by AI?. That review covers all three, but the stakes differ. In a writing task you can often get by without knowing where the content came from, because you can judge whether the sentence works. In an image-judgment task, a bad sense of when to trust the machine directly drags down the pair's score. Related work shows that AI text differs measurably from human text even when people can't perceive the difference Can humans detect AI text if machines can measure it?. In other words, intuition is a weak guide to what a model is actually doing.

There is also an asymmetry on the machine side. One recognition system found that the most reliable way to identify an unknown image was to have a vision model *describe it in words* and then search a text database Can describing images in text improve zero-shot recognition?. Language is where these models, and the people working with them, share common ground. When a task stays in words, a person can see the AI's reasoning, argue with it and edit it. When the task is pixels, the AI's basis for its answer is mostly hidden. The person ends up deferring or overriding blindly, which is the pattern that produces losses.

The point you might not have expected: "should humans and AI work together?" may be the wrong question to ask. A better one is whether the human's job in the pair is to create or to adjudicate. The corpus suggests teaming pays off in the first case and often costs accuracy in the second, whatever the medium.