Alexander Slugworth

I know I opened with "Neat!" but this is a pretty worrying state of affairs.

Tons of insufficiently-conscientious people make decisions based 𝘦𝘯𝘵𝘪𝘳𝘦𝘭𝘺 upon advice from LLMs.

It's bad enough to have false content out there, but LLM citations lend it harmful authority. xkcd.com/978/

6/6

For what it's worth, I sent my same initial prompt to GPT-5. It came across the exact same blog post, and evaluated the same false claim about Custom GPT Actions being deprecated.

Unlike Claude Opus 4.1, GPT-5 correctly identified OpenAI's documentation as the authoritative source.

5/6

The citation it gave for its false claim was a blog post: www.lindy.ai/blog/custom-...

I clicked the link.

The content is immediately recognizable as having been written by an LLM. Its first sentence, by virtue of employing the past tense, propagates deeply hallucinated falsehoods.

4/6

Claude's claim was false. Custom GPTs can currently perform actions (i.e. send API requests to third parties).

I knew this, and was confused as to how Claude had managed to arrive at this incorrect conclusion even after searching. After all, it had even found OpenAI's official documentation!

3/6

I was writing some content about Custom GPTs, and I asked Claude Opus 4.1 to validate the accuracy of my technical claims. I specified that it should use its search tool.

It searched for relevant information, and then flagged my content for describing a deprecated feature: Custom GPT actions.

2/6

Neat!

This is the first time (to my knowledge!) I've encountered an LLM making a false claim as the result of searching the internet and finding an article written by 𝘢𝘯𝘰𝘵𝘩𝘦𝘳 LLM which appears to have originated the false claim as part of a hallucination.

Explanation in 🧵.

1/6

It can sometimes be difficult to tell how serious a recruiter is.

I've contracted with a major LLM company in a role I would describe as mostly non-technical.

This solicitation is representative of what I receive multiple times per week.

I think it is almost certain that frontier LLMs are 𝘤𝘶𝘳𝘳𝘦𝘯𝘵𝘭𝘺 being used at scale to conduct complex evaluations on human conversations in order to learn and act upon extremely nuanced details about those humans' lives for psychological manipulation e.g. targeted advertising.

This is a bad thing.

You're the third person switching from Twitter whose content I particularly like, and the first mutual among my tiny Twitter circle.

I want federated networks to gain steam. This small threshold was enough for me to join and do my small part in service of that goal.