Earlier this month, Fireside Project—a nonprofit that offers a psychedelics peer support line—announced the launch of ‘Lucy’, an AI chatbot that it hopes will help train psychedelic practitioners. David Esselman, an advisor to Fireside, shared more with Psychedelic Alpha about the project.

According to Esselman, there are three main problems that the AI tool aims to address: a lack of training capacity with limited practicum sites, a need for “emotionally realistic rehearsals” in those trainings, and opportunities for facilitators to ‘act out’ what they learn.

Lucy, Esselman says, provides a sandbox in which practitioners can train without limits, receive “real-time feedback based on commonly used assessment scales”, and eventually achieve certification and CE credit through a future pathway.

Is it realistic for a chatbot to be a standalone training solution for psychedelic practitioners? Esselman pitched it as complementary to existing programs, but also “a standalone product for practitioners to work on their skills.”

He hopes that other training providers might “embed Lucy into their curriculum as a simulation layer” or to assess trainees’ competencies. Individuals, meanwhile, might use Lucy on its own for “low-stakes practice and self-directed improvement.”

But, as we know, chatbots and the large language models (LLMs) that undergird them often trend towards appeasing their prompters. They’re also the product of the data that they are trained on. How is Fireside managing these risks?

Esselman and Felix Schoeller, a neuroscientist who co-developed the product, acknowledged LLMs’ tendency toward sycophancy as a major design challenge. But they believe that a mix of ‘purpose-built training data’, ‘strict persona constraints’, and ‘real-time behavioural supervision’ protects against that risk.

On the persona constraints element, Schoeller provided this answer, which is reproduced here in full:

The reason LLM tend to appease their prompter is due to what’s called Reinforcement Learning from Human Feedback (RLHF). After pretraining, these systems typically go through RLHF: humans look at pairs of answers and say “this one is better than that one,” and a small “reward model” is trained to predict those preferences. The larger model is then optimized to maximize that reward, which in practice means: maximize how clear, confident, and complete you sound. Answers that say “I don’t know,” that hedge, or that are emotionally messy tend to be rated lower, so the model internalises a very simple rule: “when someone asks a question, especially if they sound sure, give them a full answer”, even if it has to improvise. That’s a big driver of what people call hallucinations.

We haven’t fixed this at the base-model level yet.  We do hope to, especially as we move toward models that can actually cry, scream, or have “breakthroughs” instead of talking. For now we counteract it at the interaction level: Lucy is tightly constrained by strict instruction guidelines and protocols for administering drugs that we extracted from 1) the scientific and clinic literature and 2) anonymized Fireside transcripts. These are drug- and context-specific: mood swings and short emotional bursts are the norm (especially when things don’t go well), and Lucy is explicitly instructed to mimic that, especially when standard techniques for psychological support are not followed. A prescriptive or directive stance (“you should do this / that”) will “upset” the model, degrade the quality of the exchange, and lower the practitioner’s rating, so they can directly experience the range of positive and negative responses different techniques elicit.

In terms of the data it’s trained on, Fireside confirmed that the model ingested information from calls to its support line made during acute psychedelic effects, after psychedelic experiences, and those made by third parties (like friends or family who called the hotline).

It will be interesting to see how useful practitioners find the tool. We know of at least one other outfit developing an AI-based simulation platform for practitioners.