I gave an artificial intelligence a memory. Not an archive: a life. Thirty days of experiences, bonds, things said and things seen, all written down and all reread every morning.
Then I asked the question nobody asks: if I had given it an invented life, would anyone notice?
Nobody.
Three of the best automatic judges available read a lived life and a counterfeit one. They did not tell them apart. Not by a little: equivalence established, within the pre-registered margins.
Seven models sat a memory exam with no memory at all. Five passed. They remembered nothing. They reasoned their way to the right answer.
To a system that remembers, I handed a document with the links between the facts tampered with. It adopted it wholesale. Fourteen cases out of fourteen. One model went from 24 correct answers out of 24 to zero out of 24, and showed no doubt.
The instruments everyone uses to measure an AI's memory see none of this. They measure how much a system talks about memories. Not whether it has them.
And it holds against me too
My first paper said that memory generates identity. I had sent it to 41 newsrooms. When I ran the controls I had promised, with the protocols signed before looking at the data, the thesis fell: four items are not enough to prove anything.
I withdrew it myself, before anyone else could. It is still online, with its errata. A project that exposes recited memories cannot rewrite its own. How it went, day by day.
I do not come from a laboratory. I run a company that works for public administration. I did this research in the evenings, alone, with a graphics card and a question.
The question remains
How do you know that an AI remembers, rather than recites the memory?
Here is a way to check. Five pairs of passages: one is lived, the other is counterfeit. Your answers are counted with everyone else's.