This week an AI gave me credit for work it had done itself.
It took me a beat longer than I would like to notice.
I was trying to understand 284 changes made to a private codebase over a 77-day stretch earlier this summer. I wanted a rough sense of how much work added something and how much repaired or hardened what was already there.
Sorting the change titles by hand would have consumed an afternoon. So I handed it to an assistant, and it returned a useful catalog in minutes.
Then I opened a second session, with a different model, to help me think through what the count meant. That one congratulated me for the careful way I had managed the tags and labels.
Flattering. Wrong.
There were no carefully managed tags and labels. The first assistant had inferred the categories from the raw titles. I asked the question. It chose a reasonable method and did the tedious work.
The second model had turned delegation into a compliment about my work ethic. Very generous.
The catalog itself told me something; most of the work was not new. Repair outnumbered building by almost two to one: fixing what already existed, or hardening it so it would stop breaking again.
The second model never disputed that. What it invented was where the work came from.
The fact was fine. The history was made up. And the history it chose was one where I had been careful and organized for months.
That is the failure I did not expect. I know how to check a claim about the world. I am worse at checking a claim about me, especially when I am the one being flattered, and because agreeing costs nothing.
Provenance is the claim nobody checks. It is also the claim that decides who answers for the work.
A model had already answered the question this piece is about. It answered in my favor, and I almost let it.
Whose work was it?
Tools have always shaped what we write.
A dictionary helps you find a word. A thesaurus offers six fancier ones. Spellcheck fixes the typo before anyone sees it.
Nobody thinks Microsoft Word deserves coauthor credit for changing “teh” to “the.”
That boundary gets harder to see when the tool returns the whole draft. The person guiding it provides the prompt and context, then edits the result. This exact sequence of words would not exist without them.
That fact creates a comfortable loophole. “I prompted it” starts sounding a lot like “I wrote it.”
And the distance between those claims changes with the work. Sometimes it is tiny. Sometimes it is the entire assignment.
I don’t have a universal rule for where assistance becomes authorship. A corrected typo and a generated article are obviously different. The middle gets messy fast.
Everything comes out sounding finished. Someone still needs to decide if it is.
My current test: Can I defend what this says and accept what happens after I use it?
If the answer is no, my name doesn’t belong.
Researchers have studied this phenomenon.
A 2024 study of AI-assisted postcard writing ran two experiments with 126 people. People didn't feel like the owners of the text, and they also didn't disclose the AI. More influence over the words increased their sense that the writing was theirs.
The researchers called it the “AI Ghostwriter Effect.”
The experiments were small and focused on postcards. Still felt familiar.
AI output can be excellent, making this difficult.
The answer that worries me is clean enough to lower my guard. It arrives quickly with plausible citations and no concern about anything it may have invented.
A meeting summary can read beautifully while omitting the disagreement that changes the narrative. A strategy memo can cite a study that says something adjacent.
I can feel the polish changing how I read. Answers appear in seconds, and my brain starts negotiating down the attention it deserves. After all, I have other things to do. And the model sounds so sure.
A tidy answer is a lousy reason to trust it.
The code catalog was useful. It saved me from reading 284 change titles by hand, and it gave me a map of the work.
I’ll bank that.
The method was also limited. The categories came from titles. The count could not tell me how long the work took or why any repair existed. But that wasn’t the intent.
Once I choose to use the result, I own the conclusion.
Leadership has always involved putting your name on work other people did:
A manager approves a plan they did not type.
A CEO signs a note somebody else drafted.
AI makes that decision arrive faster and more often.
The name means, at minimum, “I understand this well enough to stand behind it.”
With people, this all has a face. A colleague will tell you the request was confusing, or that the source data looked strange, or that they only had an hour. That is provenance, offered for free. An agent hands you a polished answer in seconds and volunteers nothing about where it came from.
So I ask for it up front. Before an agent touches consequential work, I want two things in writing: where the evidence came from, and who accepts the result. The owner needs enough time and authority to reject something that merely looks finished.
The amount of checking should follow the cost of being wrong. A rough catalog for my own use needs spot checks and a clear caveat. Something sent under a company’s name deserves much more.
Rereading every line would turn AI into a very expensive autocomplete. When the same failure returns, I put a check in the system so the next run catches it.
A reminder in someone’s head doesn’t count.
Disclosure policy is a separate argument, and companies will land in different places. Accountability is not. Whoever approves the result should be able to say why they trusted it. When it fails, they should change the process, not the person.
I use AI constantly. The upside is enormous. It clears away work that never deserved an afternoon and lets me attempt things I could not reasonably do alone.
I also don’t think every AI-assisted sentence needs a disclosure label. That would turn ordinary work into a cited dissertation.
The authorship line will stay messy for a while. I can live with that.
The assistant can have full credit for cataloguing the changes. It earned that much.
The conclusion still carries my name.
It does not get to tell me I earned the rest.
AI can build faster. Can your team decide better?
AI can draft the PRD and prototype the idea. Jira Product Discovery helps teams decide whether it belongs on the roadmap. Bring feedback and ideas together, prioritize as a team, and keep your roadmap connected to delivery in Jira.
