How 'Artificial Intelligence' Is Slowly Drowning In Its Own Droppings

© 2026 Peter N. M. Hansteen

A website designed to appeal to humans' vanity serves to reveal how the large language models degrade by ingesting each others' unverified output

In early July 2026, I was made aware of the website intheweights.com via a fediverse post. The idea is simple, and appeals to human vanity: You type in your name, and the site checks the various large language models for any data it can find on you in the models.

The post I responded to appears to be no longer available, but I checked out my own results and posted the results with screenshots:

Post by @pitrh@mastodon.social

View on Mastodon

If for some reason the embedded post is not readable, I also saved screenshots of the site output.

First up, the main result, with what seems to be the highest confidence score:

Fair enough, I advocate OpenBSD but did not write any part of the operating system, I wrote The Book of PF, and I work in the tech industry.

I addition, there were two entries that I was pleased to see the site labeled as likely hallucinations:

That is definitely not me. For context, according to the Name Search service on Statistics Norway's (Statistisk Sentralbyrå) website, we are only 140 persons who have Hansteen as their surname, and Fewer than four people or no one has the name Peter Nicolai Mathias Hansteen, and to my knowledge I do not share first names with any living relative.

I am no longer active on the site that used to be called Twitter, which is where I assume Grok finds a significant subset of the data it ingests. But while there, since the late noughties until shortly after the site was sold, I may very well have posted about global warming with links to actual studies that involve solar phenomena.

The next entry I found a bit more bizarre, to wit:

That one is a bit strange. While I do have a living close relative who is a geologist with publications in academic journals, this entry seems to have taken most of its data from the biography of Christopher Hansteen. We are related, but as can seen from the Wikpedia article, separated by several generations.

But at least, the site considered both of those to be likely hallucinations.

But give it a couple of weeks

For whatever reason (er, actually my laziness), the intheweights.com tab stayed open in my browser, and I went back to check the results on July 21st.

The first entry was much the same,

Oddly enough, the confidence had risen a bit, and I was now in the top four percent. I still did not write that operating system, I only use it and like it enough that I will likely talk about it to strangers.

The second had apparently been moved up the ranks and was no longer considered a likely hallucination:

So what happened here is that Grok had grown more confident in its most likely imaginary physicist. Another close but no longer living relative was a physicist, but the description does not map well to their actual published work.

Then thankfully, the last item was still considered a likely hallucination:

So the conclusion here seems to be that while the confidence in the largely fact based first item had risen overall, so had the rating for one of the likely hallucinations.

The explanation here is likely that for my material, it is very much available online, some of it under copyright but free, some under copyright but without technologically enforced restrictions, and some is actually under a free license for anyone to use as they wish.

For my relatives who are actual academics, both the living and the deceased, more than likely they have published more material than I have in total, but a large part of that material either predates the Internet or is hidden behind paywalls. Then again my relative the geologist insists that the bulk of their material is indeed freely available online with sufficient information to separate their work from mine.

So there we have it: The large language models can only ingest what is placed within reach, and they work in parallel to generate material based on what their inner, opaque, and possibly quite tangled, logic considers probable. Then it seems that the output generated soon after is fed back to the large language models as training data for themselves.

This means that more likely than not, more material will be attributed to the more visible person in the training data, based on their version of probability, than is actually appropriate. We saw an early version of this in Those Titles the AI Bot Thought I Had Written (also here).

With the large language models seeming to grow more confident of the actual existence of their own fictional creations, how likely is it that they will soon sink out of sight, buried in layer upon layer of unverified output from themselves and their peers?

My best hope is that we will be able to clean the Internet of their nonsense before it is too late.