This conjecture is supported by the fact that Adele Lopes described people who helped parasitic AIs to leave messages as follows:

As for "getting LLMs to satisfy common needs in healthy ways, help people figure out what their own vulnerabilities are in an attempt to make them better", is it what OpenAI and Anthropic are trying to do? Or has OpenAI succumbed to external pressure in ways like rebooting GPT-4o-sycophant and announcing Sora the slop generator?

On the other hand, mankind saw that the AI will likely be able to convince researchers^[1] that it should be released, e.g. in an experiment with the AI roleplaying an AI girlfriend. Does it mean that a superpersuader can convince any human who isn't well protected? And what about a group of humans? It might be useful to deliberately check how persuasion capabilities depend on compute spent and architecture in a manner similar to benchmarking compute and architecture on simpler problems, then to ensure that no model approaches the dangerous thresholds...

1. ^{^}

   While someone proposed using prison guards, since these people likely have a different set of vulnerabilities, LLMs have been claimed to induce trance in an experiment.