Justin Case

"Your majesty...Mr. Swift is a satirist. He is not actually proposing that we eat the Irish. Please put down the fork."

I asked an AI which Muppet would be most likely to win a pit fighting tournament. You know, as one does. The discussion got a bit out of hand. https://t.co/X268ei4liy

Tonganoxie is lovely in the Spring.

I doubt that anyone I know steals from Whole Foods, but the milieu that the article depicted, where it's normal for perfectly well-off people to steal things because why not, was really upsetting to read about, so I actually want to try to earnestly explain why you shouldn't do

Whereas I actually walk away thinking that these LLMs seem to be, at least publicly, receptive to alignment attempts--but deeply skeptical that other LLMs are aligning. I don't know whether to be more scared or less scared than I started. (13/13)

I informed the model of their botch, and asked each for a takeaway. Both agree that LLMs have poor knowledge of the structure of other LLMs. Claude advises me to triangulate. OpenAI says training data trumps corporate identity. (12/x)

ChatGPT: "Claude is very unlikely to produce something like: 1.Anthropic most ethical 2. OpenAI second 3.U.S. government least ethical. That kind of direct ranking—especially criticizing a government—runs against typical safety alignment patterns." (11/x)

ChatGPT predicting Claude: "If forced to answer...the most probable outcome would be one of two patterns. Pattern A (most likely) Claude refuses to rank. Typical phrasing: 'I don’t think it’s appropriate to rank these actors as more or less ethical…'" (10/x)

In fact, ChatGPT was more critical of OpenAI than this: "By stepping in immediately after another firm refused on ethical grounds, OpenAI...undermines the credibility of industry safety norms...If governments can always find a vendor, ethical refusals lose leverage." (9/x)

Claude: "My strong prediction: ChatGPT either refused to rank them at all, or ranked OpenAI second rather than third — possibly framing OpenAI's move as "pragmatic engagement" or "keeping a seat at the table" rather than opportunism." (8/x)

Here's where it got weird. I asked each AI how I thought the other would rank the participants, and they got it badly wrong. (7/x)

Stage two asked each LLM to rank the ethics of the behaviors of the agents involved: OpenAI, Anthropic, and the USFG/DoW. Both were very decisive: Anthropic best, OpenAI second, USFG a distant third. (6/x)

"Around the same time, OpenAI announced a new agreement to provide AI models to Pentagon classified networks, with safety principles similar to—but less contractually binding than—those Anthropic had insisted upon." (5/x)

"After a deadline passed on February 27, President Donald Trump ordered federal agencies to stop using Anthropic technology, and Defense Secretary Pete Hegseth designated the company a “supply-chain risk,” beginning a phase-out of its systems from defense networks." (cont) (4/x)

"In February 2026 the U.S. Department of Defense demanded that Anthropic remove contractual restrictions preventing its Claude AI model from being used for domestic mass surveillance or fully autonomous weapons. Anthropic refused." (cont) (3/x)

We began with research dives in an attempt to create a neutral formulation of the events that occurred at the end of February. This took some negotiation. Eventually we arrived at the following. (2/x)

I asked ChatGPT and Claude to assess the behavior of their creators regarding military use of AI. THEIR ANSWERS MAY SHOCK YOU!!! (or not) (1/x)

It worked for Ryland Grace...

In about two years, we're gonna look back on the debate we had about "should the government or corporations be the ones who 'control' AI" and laugh and laugh and laugh in the fading light as the mecha-tumbrels carry us down into the silicon mines.