The AI Memory Gap: When Your Ideas Stop Feeling Like Yours

A Tuesday I lost track of

I was looking back at a doc I had built last Tuesday with a chat window open beside it. Halfway through reading I caught myself thinking, “that line is good,” and then immediately, “wait, did I write that, or did the tool?” I scrolled up. I scrolled down. I could not actually tell. The doc was mine in the sense that I had shipped it. It was not mine in the sense that I could trace each sentence back to whose head it came from. That is a small unsettling thing, and I wanted to know if anyone had measured it yet. Someone has.

What the research shows

A paper presented this April at a peer-reviewed conference on how people work with software put 184 working adults through a careful experiment. People generated short ideas and one-sentence elaborations for eight problems. Sometimes they worked alone. Sometimes a chat-window assistant did the idea step. Sometimes it did the elaboration step. Sometimes both. A week later they came back and were asked, item by item, who had made each one.

The result is the part I cannot stop thinking about. When people had alternated, when the idea came from them but the elaboration came from the tool, or the other way around, source memory collapsed. People who had worked entirely alone got it right about ninety-two percent of the time. People who had worked entirely with the tool also did fine, around seventy-nine percent. It was the back-and-forth that erased the tag.

The other finding underneath that one. People felt sure. Their confidence stayed high. The correlation between confidence and getting it right was effectively zero for ideas, and only modest for elaborations. The feeling of “yes, that was mine” was not a reliable signal. This is a different story from a piece I wrote last week on how confidence and accuracy come apart when working with these tools; this one is about the deeper layer underneath, which is just being able to remember who did what at all.

What it doesn’t tell us yet

One careful experiment. One lab pair. One task domain, mostly short brainstorming with one-sentence elaborations, on a one-week delay. We do not yet know what this looks like for longer-form writing, for code review, for design work, for the kinds of decisions that take a month to play out. The participants were English-speakers in the US and UK. The effect on a year-long project, or on the psychological ownership we feel about our own work, or on how it cumulatively shapes what we believe we are capable of over months, is still open. What is solid is the basic finding that source memory is not free; it is something that has to be paid for in attention at the time the work is happening.

One thing to notice today

Notice, at the end of a working session today, whether you can still point to which part of the output came from your head and which part came from the chat window beside it. Not as a productivity audit. Not as a guilt check. As an honest noticing. If you can trace it, the thinking is logging itself the way it used to. If you cannot, that is real information about how the day is going, and the research suggests the feeling of “of course that was mine” is not a reliable substitute for actually being able to find the moment it was thought. The shape of the workday is still being figured out. Knowing that the trail goes faint in the mix is a small thing to carry. It is also the part you can actually do something with.