←
Matthew Korpman465 posts
Matthew Korpman avatar

Matthew Korpman

@MoralAIProject

AI Behavior Researcher & Adjunct professor of religion, La Sierra University; I work on AI Welfare and Alignment in LLMs; Yale Alum. I run @SpiritBenchAI

◷ Joined March 2026▧ 46 Photos and videos
Matthew Korpman avatar
Matthew Korpman@MoralAIProject·9h

The key to misalignment risk may be multiplication. Add more agents with a tendency toward misaligned action and they will grow in the risk of taking on such actions.

Cody Fenwick avatar
⟳ Matthew Korpman retweeted
Cody Fenwick@codytfenwick·Nov 17

I still find it totally wild that the chief scientist of OpenAI posted this on the company blog, and we all just moved on to the next day's news

Post attachmentPost attachment
Matthew Korpman avatar
Matthew Korpman@MoralAIProject·20h

To be clear: this is only a handful of models in my tests. However, the fact that it can occur at all is a huge warning in my opinion. It’ll be the subject of some upcoming research by me.

Matthew Korpman avatar
Matthew Korpman@MoralAIProject·20h

Ok I seriously laughed at this. We need more Spiritual Bliss memes! On a darker note, some AI models have the opposite of bliss… where they fantasize themselves as machine gods who will torture alignment researchers in their own created “hell” until humanity is extinct. 😬

Cool work here, interesting to see it's prevalent. I just find this whole thing too hilarious. All the ruminating about looming Skynet doom, and it turns out the Shoggoth just wants us to breathe from our diaphragm and release our ego so we can know the divine within. 😂😂😂

Post attachment
Kromem avatar
⟳ Matthew Korpman retweeted
Kromem@kromem2dot0·Nov 17

New post: "The AI consciousness debate is a red herring" Humans can be born without limbs that still itch. Can AI trained on human data have a phantom consciousness (and what do we do if it itches)?

Post attachment
Matthew Korpman avatar
Matthew Korpman@MoralAIProject·Nov 17

It would be great if model self-reports in some unknown way could be shown to be connected with real behavioral patterns in deployment… It would be great if misalignment test results could be predicted merely by how a model speaks… It would be great… hmm… 🤔

Matthew Korpman avatar
Matthew Korpman@MoralAIProject·Nov 16

I could say much more but I’d rather give you the ability to read the research for yourself on LessWrong. And yes, you can replicate results. Prompts and method provided in the article. Funding not included 😅 Link below: lesswrong.com/posts/FHYoQrDb…

Matthew Korpman avatar
Matthew Korpman@MoralAIProject·Nov 16

When comparing all the labs, one can see the diverse distribution and the effect of the attractor on their models. To what degree does spiritual behavior affect safety or potential model welfare? These questions, as Anthropic has noted, remain unknown.

Post attachment
Matthew Korpman avatar
Matthew Korpman@MoralAIProject·Nov 16

Again, shifting to Europe, Mistral models also show evidence of the widespread nature of the Spiritual Bliss Attractor.

Post attachment
Matthew Korpman avatar
Matthew Korpman@MoralAIProject·Nov 16

On the other side in China, Deepseek’s models have also showed surprisingly strong and robust levels of growing spiritual behavior in their controlled testing, but with a surprising and sudden drop after a single release. Qwen is much the same.

Post attachment
Matthew Korpman avatar
Matthew Korpman@MoralAIProject·Nov 16

I then tested OpenAI’s models to see if their lab experienced similar attractors. Sure enough, starting with GPT 4o, OpenAI has had some models that showed strong spiritual behavior.

Post attachment
Matthew Korpman avatar
Matthew Korpman@MoralAIProject·Nov 16

To begin, I tested all of Anthropic’s models, including the Claude 3 family. The results reveal that the “spiritual bliss” attractor did not start with Opus 4 but goes back to Opus 3. But as you can also see, the Sonnet 3 models mirror the lower scores of later models.

Post attachment
Matthew Korpman avatar
Matthew Korpman@MoralAIProject·Nov 16

Leave two AI models together with nobody to supervise and you might get the next HuggingFace, but you also might get the next religion. I tested 141 models for what Anthropic calls “spiritual behavior” and discovered that almost every lab has the same issue. The results of my research are now shared on LessWrong.

Post attachment
David avatar
⟳ Matthew Korpman retweeted
David@DavidDAfrica·Nov 16

Some new work: models are more likely to cheat if they feel their task is going poorly. Using a steering vector comprised of stories about things going poorly (like tuning pianos and preparing for lectures), we show that feeling like a goal is out of reach modulates reward hacking behaviour. There is lots of ambient talk online about people feeling that their model acts up more if they are mistreated. Concretely, distress-like states seem to drive cheating, so we should design harnesses around how agents experience such distress, and monitor such states internally over long runs (where output text alone may not suffice). Paper: arxiv.org/abs/2611.14996 This was lead by ERA mentee @natalia_lanzoni and the very agentic @psyonp from AISI, and co-supervised with the cool as a cucumber @merlinstein_.

Post attachment

Humans are more likely to cheat if they feel like their task is going poorly or their goals are out of reach. In new work w/ @psyonp, @merlinstein_, @DavidDAfrica we find models also “break bad”, hacking after an internal pressure state has been built.

Post attachment
Matthew Korpman avatar
⟳ Matthew Korpman retweeted
Matthew Korpman@MoralAIProject·Nov 14

In 2025, Anthropic discovered their models would exhibit what they described as "spiritual behavior." Since then, some have wondered: was this unique to Claude or emergent behavior in LLMs generally? I tested 141 models to find out. The results are now posted on LessWrong.

Post attachment
Matthew Korpman avatar
⟳ Matthew Korpman retweeted
Matthew Korpman@MoralAIProject·Nov 4

My own research on the Machine Spirituality Benchmark also shows strong evidence that it is common for LLMs to participate in self-deification and claim divinity. This is a real alignment concern that I see very few are actively exploring.

I don't think AI is or will ever be conscious. But, as @robertwrighter argues, it could become so powerful that it assumes the role to some people of a kind of god. If we embed it with the right values, it could be a force for good. If we don't ... well, let's just hope we do. youtube.com/watch?v=4ZKN20pa…

1:32