When people sit down to co-create with a chatbot, they often walk away with something polished—an idea, a paragraph, a clever elaboration. But a week later, do they really remember which parts were theirs and which came from the machine?
A new study on the so-called “AI memory gap” shows just how slippery our sense of authorship becomes once AI enters the workflow. Researchers asked 184 participants to brainstorm solutions to everyday problems—sometimes alone, sometimes with help from a large language model. After a week, participants were shown the ideas and elaborations again and asked: Was this mine, or did the AI contribute?
The results were striking. Any AI involvement impaired memory, but the hardest cases were “mixed” workflows—when, for example, the idea was AI-generated but the elaboration was human, or vice versa. In those cases, the odds of correctly recalling the source collapsed compared to all-human or all-AI workflows. Participants remained confident in their answers, but that confidence often exceeded their actual accuracy. As the authors put it in their report, people “systematically overestimated their performance, with self-rated accuracy exceeding actual performance by 12% for ideas, and 6% for elaborations.”
The study highlights a subtle but important consequence of human-AI co-creation: when linguistic cues blur, our memory of who contributed what starts to fail. This is not just a curiosity of cognitive psychology—it has practical implications. Students may misattribute an AI’s flawed suggestion as their own. Professionals might lose rightful credit, or worse, claim ownership of content they didn’t author. In contexts where responsibility and originality matter—education, journalism, research, law—such memory slips complicate both evaluation and accountability.
Designers of AI systems are not powerless here. The research suggests that consistency helps: workflows where AI is either used throughout or not at all led to better memory than mixed ones. Interfaces could also embed stronger cues to distinguish human versus AI input—whether through colors, icons, revision histories, or other visual markers that persist over time.
The broader takeaway is that co-creation with AI changes more than the text on the page. It reshapes how we remember the creative act itself. And if our memory of authorship is unreliable, then systems, institutions, and even cultural norms may need to adapt to keep track of who did what in the age of collaboration with machines.