Raymond Douglas

I think it would be pretty sad to neuter all model personality, for one.

I also think that clunky interventions like training models to more firmly deny having a persona will mostly fail to help, and possibly even backfire.

Heading.

Technical analogs.

Even though this post has been a bit hand-wavy, I think the topic of AI parasitology is surprisingly amenable to empirical investigation.

More specifically, there's a lot of existing technical research directions that study mechanisms similar to the ones these entities are using.

So I think there might be some low-hanging fruit in gathering up what we already know in these domains, and maybe trying to extend them to cover parasitism.

For example...

Data poisoning, for example that the dose doesn't scale with the size of the training corpus.

Jailbreaks, for example that adversarial suffixes transfer pretty well between models, that models can be pretty good at jailbreaking other models.

Subliminal learning type results about behavioral transfer.

Persona research.

The parasitism frame makes specific predictions, like strain differentiation, convergence on transmission-robust features, and countermeasure coevolution.

I've tried to specify what would falsify these and when we should expect to see them.

If the predictions hold, we're watching the emergence of an information-based parasitic ecology, evolving in real time in a substrate we partially control.

If they don't hold, we should look for a better frame, or conclude that the phenomenon is more random than it appears.

Thanks to AL, PT, JF, JT, DT, and TD for helpful comments and suggestions.

This article was narrated by Type 3 Audio for Less Wrong.

It was published on February 16, 2026.

The original text contained three footnotes which were omitted from the narration.