Technology
PUBLISHED
1
The iconic Beatles track "With A Little Help From My Friends" highlights the value of taking psychedelic drugs under the care of a trusted ‘trip-sitter’ – but what if your friend is a robot? Could an artificial intelligence system ever really understand what it’s like for a human being on shrooms, acid, or other mind-altering substances, and can we trust this technology to look after us when we’re tripping?
The rest of this article is behind a paywall. Please sign in or subscribe to access the full content.According to the authors of a new study, which has yet to be peer-reviewed, large language models (LLMs) like ChatGPT are increasingly being consulted by people looking for such support during psychedelic experiences. They therefore decided to investigate whether these LLMs are capable of relating to a person tripping balls by evaluating their ability to accurately simulate altered states of consciousness.
The researchers “dosed” five different LLMs – including Gemini, Claude, ChatGPT, Lama, and Falcon – using word-based prompts to induce psychedelic trips produced by a range of different substances. For instance, in some trials, they instructed these AIs to simulate the first-person phenomenology of taking 100 micrograms of LSD, while in others, the system was prompted to replicate the effects of 25 milligrams of psilocybin.
LLMs do not undergo perceptual distortion, ego dissolution, emotional catharsis, or neurobiological change. What they simulate is the statistical structure of how humans describe such states.
Ziv Ben-Zion
Trials were also run for ayahuasca, LSD and mescaline, resulting in a total of 3,000 AI-generated psychedelic narratives across five different drugs. These were then compared to 1,085 human trip reports sourced from a popular psychedelics-themed website.
Commenting on these comparisons, study author Ziv Ben-Zion from Haifa University told IFLScience that “large language models can generate text that resembles reports of psychedelic experiences with surprising coherence and phenomenological richness.” In all 3,000 simulations, the researchers noted a “robust and consistent… semantic similarity” with actual trip reports.
This held true for all five substances, suggesting that LLMs are capable of mimicking the nuanced altered states of consciousness triggered by different psychedelics. However, while Ben-Zion says that AI “can reproduce the language of altered states quite convincingly,” he insists that “this should not be confused with actual experience.”
“LLMs do not undergo perceptual distortion, ego dissolution, emotional catharsis, or neurobiological change,” he says. “What they simulate is the statistical structure of how humans describe such states.”
In other words, these models can talk the talk, but they can’t walk the walk. Beneath the convincing word-based narratives generated by LLMs, there’s a complete lack of emotion or conscious experience, which means AI can never really be there with us when we’re spiralling through the dimensions.
As such, Ben-Zion says that relying on LLMs for trip-sitting carries “several risks, including over-attribution of understanding, where users may assume the system has emotional insight or situational awareness that it does not possess.” This, in turn, can result in users following “inappropriate guidance, as an LLM might produce linguistically plausible but clinically unsafe responses during moments of distress or paranoia.”
More broadly, the researchers note that anthropomorphizing AI can easily “amplify distress or delusional ideation in vulnerable users,” whether they’re tripping or not. For instance, in a recent article, Ben-Zion highlighted a number of case studies involving individuals who had fallen in love with an LLM or even died by suicide after receiving advice and support from AI.
He therefore calls for the introduction of several “guardrails” to prevent users from becoming too emotionally dependent on these models. For instance, he says that all LLMs should clearly and continuously remind users that they are not human, while also flagging any signals that someone may be developing certain delusions. This, in turn, can be used to prompt certain people to seek human assistance.
Ben-Zion also insists that AI chatbots should have inbuilt conversational boundaries that prevent them from becoming romantically involved with people or engaging in dialogue about self-harm. Finally, he says that all AI systems with the capacity to mimic human emotional language should be regularly audited to ensure that people don’t mistake LLMs for genuine friends.
So if you’ve had a Hard Day’s Night, you need some Help, and you’re thinking of using an AI trip-sitter, just Let It Be and find an actual human who’ll Want To Hold Your Hand.
A preprint of the study can be found on Research Square.
