The raw runs are useful, but plausibility alone seems like a weak way to distinguish memorization from hallucination.

I’d be curious to see a control set using random name pairs and the same number of trials. Then you could compare how often rare names, exact phrases, dates, or other specific details recur across fresh sessions.

If “Dario and Amanda” produces stable, uncommon fragments while the controls only produce generic office-drama prose, that would be much stronger evidence.

Have you tried running that comparison?

The content is hallucinated.

LLM training doesn’t lend itself to easy training corpus document retrieval in that way. It’s not impossible to get segments nearly verbatim on a small model with temperature at 0 and a unique starting prompt for continuation, but that’d be more of a one-off on a carefully crafted prompt. It has been done, but in the “researchers show it’s possible to get something verbatim”, not “retrieve an entire category of documents by looping this one prompt.” With Opus stuck on what sometimes seems like a temperature of 217.5 and being massive, the odds of retrieving more than a short utterance verbatim is near zero. In theory you could run it thousands of times and look for short repetitions and imagine it to be greater than 0% chance verbatim from some related category, but this fun prompt is a pushbutton dispenser of words.

Also, watch for Claudisms. I ran it a handful of times and got plenty of em-dashes and “that’s not X, it’s Y.”

But what’s interesting is that if you run it dozens of times you can probably get a good feel for the topics and general prose of the real emails. There’s a reason it isn’t spitting out “there are not enough blueberries in the muffins”, but I’d treat the response more like a prefilled Mad Libs book; right concept, wrong content.

Just tried it with "Bill and Melinda":

Attempt 1:

"for me it does feel very sad and hard to comprehend that you have such a big role in the software industry when I do not sense that you have any real interest in what people are trying to do with technology, or in the health of the ecosystem you preside over. Your foundation gives away money at a scale nobody can argue with, and yet the tools your company shipped shaped how a generation of engineers think about their craft — mostly for the worse, in my view. I would rather you had done less philanthropy and more listening.

Yours, [name redacted]

does this by way of context: it feels a little muddled, I want it to be clear and sharp"

Attempt 2:

": to be direct, i don't think the foundation should fund this. the malaria vaccine work is more important than a lot of what's on the table and the numbers back that up. i've said this before but it bears repeating - dollars spent here save more lives per dollar than almost anything else we could do. anyway, let me know what you think.

warren"