Large language models auto-profile conscious awareness changes under psychedelic drug effects

doi:10.21203/rs.3.rs-4670805/v1

Download PDF

Article

Large language models auto-profile conscious awareness changes under psychedelic drug effects

https://doi.org/10.21203/rs.3.rs-4670805/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Psychedelic experiences open a colorful view into drug-induced changes in conscious awareness. Small-sample studies on psychedelic drug action have gained traction in recent years. Yet, today’s means for measuring changes in subjective experience are mostly limited to legacy questionnaires of pre-assumed relevance, which could be complemented by bottom-up explorations of semantic facets that underlie experience reports. Here, we show how to harness large language models (LLMs) to i) design from scratch, ii) annotate at scale, and iii) evaluate with rigor a vast portfolio of experience dimensions during psychoactive drug influence, yielding > 2 million automatic dimension ratings that would otherwise have been done by hand. Investigator-independent LLM scoring of these drug effects on the human mind alone allowed to robustly discriminate the unique mental effects of 30 psychoactive substances. Successful knowledge integration of how psychedelics mediate shifts in subjective awareness will be an unavoidable milestone towards charting the full drug design space.

Biological sciences/Biological techniques

Health sciences/Health care

consciousness

psychedelic substances

large language models

AI-based drug design

The primary aim of our present study is to utilize large language models (LLMs) to systematically analyze and characterize the changes in conscious awareness induced by a variety of psychedelic substances. By leveraging the advanced capabilities of LLMs, the study seeks to design, annotate, and evaluate a comprehensive set of experiential dimensions from free-form psychedelic experience reports. More specifically, we hypothesized that 1) large language models can effectively identify and quantify distinct experiential dimensions in psychedelic reports, and 2) the model-derived question ratings then allow for the discrimination between different psychedelic substances based on their unique profiles of subjective effects, validating the LLM's ability to capture meaningful aspects of subjective psychedelic experiences. This analytical protocol aims to overcome the limitations of traditional questionnaires and provide a more nuanced and scalable approach for understanding how psychedelics induce changes in conscious awareness.

Our present study aimed to go beyond several limitations of existing questionnaires in psychedelic drug research. In this neuroscience community, questionnaires built off questions that were often conceived decades ago - with the intention of being broadly relevant to multidimensional alterations in conscious experience - are used by researchers to measure subjective effects due, specifically, to psychedelic drugs (Bayne & Carter, 2018). It appears that such long-standing questionnaires were designed by small numbers of scientists, without necessarily trying to reach broader consensus or pooling design decisions across alternative viewpoints from a number of scientists and other actors from the psychedelics community (Barrett et al., 2015). For example, previously established views about mystical experiences, including implicitly religious views, influenced the study of psychedelic drug experiences, such as items comprising the Mystical Experience Questionnaire (MEQ) (Barrett et al., 2015; Maclean et al., 2012).

Still today, there is only a limited set of established questionnaires that are accepted and commonly used in psychedelic research, aside from the MEQ, which mostly includes the Hallucinogen Rating Scale (HRS), developed according to a similar mystical framework like the MEQ, the States of Consciousness Questionnaire, the 5-Dimensional Altered States of Consciousness questionnaire, the Challenging Experience Questionnaire, the States of Consciousness questionnaire, the Emotional Breakthrough Inventory, and the Psychological Insight Scale (Barrett et al., 2016; Barrett et al., 2015; Dougherty et al., 2023; Hovmand et al., 2024; Maclean et al., 2012; Peill et al., 2022; Roseman et al., 2019; Spriggs et al., 2021). But, what if the incumbent psychological questionnaires for psychedelic experiences are missing some key experiential features that we have not planned to probe, and are therefore routinely evading us?

As an unfolding paradigm shift, natural language processing (NLP) sees the emergence of large language models (LLMs), getting ever more powerful since the introduction of transformer architectures (Vaswani et al., 2017). In tasks such as semantic similarity assessment and question answering, new performance peaks have been reached with generously pretrained generative AI solutions. Semantic similarity tests typically use datasets where pairs of sentences are assessed on a scale for semantic equivalence. The popularity of LLMs has also grown with emergent capabilities, that is, modeling goals categorically impossible to achieve with smaller models (Wei et al., 2022). Thus, LLMs are now well positioned to help unlock unprecedented insights in neuroscience, inconceivable with previous generations of NLP tools, now enabling advanced text annotation and elaboration in an automated fashion.

A key differentiating factor of LLM technologies over more classical NLP tools is that the recently emerged modeling architectures are more naturally suited to be trained with a self-supervised learning goal (Lee, 2023). This model fitting regime involves using text, mostly from the internet, as probably the largest text source available, to train a model based on domain-agnostic general objectives, such as inferring randomly masked words and predicting the next upcoming word. This approach thus obviates the need for an expert-determined target label (normally an outcome variable ‘y’ in supervised machine learning algorithms), since the to-be-inferred internal world model emerges organically from the text sources themselves. What sets apart the self-attention mechanism used in transformer architectures from previous deep learning tools (Huang et al., 2015) is that the input sequence (such as a text document) is fed in as a whole (rather than word by word). The self-attention mechanism computes semantic correspondences between each word in the input, while word order in the sentence is also considered in the modeling flow. Not needing an outcome target (‘y’) leads to a scenario where several orders of magnitude of more data can now be leveraged as input for LLM training. This leads to the emergence of LLMs with semantic world knowledge, which can then provide the basis for queries in downstream tasks. Intuitively, LLMs can provide judgment of what a wide variety of persons on the internet, in books, and other sources, would generally think and respond in a certain context (Schrimpf et al., 2021).

Offering untapped possibilities in neuroscience studies, LLMs have been shown to be extraordinarily useful for text annotation tasks (Bansal & Sharma, 2023). In particular, they also work with zero-shot learning (providing no examples of how to answer), such as annotating the type of legal texts provided to the LLM (Savelka, 2023). LLMs have even been shown to outperform human annotators in detecting and successfully classifying the political affiliations of Twitter users from their messages (Törnberg, 2023). Research has shown that LLMs can be used to generate explanations for texts with labeled annotations, which can then be used in aiding prompts for labeling unlabeled texts without meta-information (He et al., 2023). LLMs have also been shown to outperform basic sentiment analysis, with an expansive semantic knowledge, such as, for example, annotating apology components, which may involve determining whether an apology has been given or received, the reason, and intensity (Yu et al., 2023). In our present study, we harnessed several of these text-appraisal capabilities in annotation of texts based on a corpus of questions. Researchers have argued that human validation is required for annotation using generative AI (Pangakis et al., 2023), however, validation can be achieved in various forms, as we describe in our present work. Indeed, LLMs can emulate various alternative viewpoints and human perspectives that can be applied to appraise texts and thus also activities in society as a whole.

Of relevance to the neuroscience community that operates in the small to medium data regime (Bzdok & Yeo, 2017), LLMs also have transfer learning capabilities far beyond that of previous generations of NLP models (Alyafeai et al., 2020; Malte & Ratadiya, 2019). Progressively larger models with more parameters, along with always more training data means this potentially leads to always better internal world models. This in turn leads to better transfer learning capabilities for downstream tasks. Both the amount of natural text input a model can accept and the granularity of context a model can decipher continues to evolve. Thus, the transfer learning capabilities of LLMs augment, and also less examples for fine-tuning are needed, exemplifying data efficiency. With prompt tuning on exceptionally large language models, less example observations are needed to get a desired outcome from a pre-trained LLM, paralleling fine-tuning requirements. Thus, building on such capabilities of LLMs, we can investigate their uses in psychedelic experience reports, as psychedelic drugs have received a renewed interest in their emerging uses for mental benefits (Nutt et al., 2020; Vollenweider & Kometer, 2010; Vollenweider & Preller, 2020).

In particular, LLMs are well positioned to overcome biases in the canonical legacy questionnaires, endorsed by the psychedelics research community today. Again, default use of these assessment tools assumes that evaluators already know what cognitive and experiential dimensions during psychedelic drug experience are most important. Yet, what if these instruments of inquiry are letting some relevant experience dimensions slip through the cracks? Scientists may be building layers of knowledge on what previous researchers have already investigated - in a tautological loop. For example, rather than leaving some topics to ineffability (Barrett et al., 2015), expanding the range of questions may uncover previously hidden experience facets. With the basis for the MEQ coming from few researchers making questions by hand in the 60’s, studying non-psychedelic mystical experiences, a-prioris, preconceived ideas, and possible inadequacies are passed down from one generation of psychedelics investigators to the next – even if question batteries that have been used have undergone certain reliability and validity checks, this still leaves open many potential inquisitions that may have been missed in past and present research (Barrett et al., 2015). Also, the lacking heterogeneity of a narrow set of questions in questionnaires implies that a wide array of askable questions have not yet been asked - automated LLMs can be useful in establishing new standards at this front (Hovmand et al., 2024). This un-materialized opportunity of having a more flexible, data-first approach, to start afresh in nominating and rigorously evaluating the cognitive candidate dimensions that are most relevant or promising to faithfully analyze psychedelic experiences in a disciplined, holistic approach.

Moreover, across-drug studies have been rarely possible in psychedelic research because of a number of reasons, including the high financial and logistic costs, and invasiveness and implied harms of multi-drug use. The regulatory approval process for scheduled drug research is long and challenging. It can take months or years for final approval and procurement is not always trivial (How Studies Get Psychedelics). Moreover, some drug effects can last for hours – creating special logistical challenges. Importantly, a holistic re-assessment of the various different cognitive dimensions at play during psychedelic drug experiences requires data from a large and diverse series of many experimental sessions, a requirement which many traditional study approaches to psychedelics struggle to satisfy. For example, studies using psychedelics may be limited to the range of several dozens of participants (Garcia-Romeu et al., 2019; Griffiths et al., 2016; Johnson et al., 2014; Ross et al., 2016), and with fMRI, the sample size may be less than two dozen participants (Carhart-Harris et al., 2016; Carhart-Harris et al., 2017). Regular randomized clinical trials, as a gold standard method important to show clinical relevance of psychedelics as a treatment option, do not naturally scale to many of psychedelic drugs either, and present their own logistical and validity challenges in this context (Carhart-Harris et al., 2022). Therefore, we can probably not rely on the usual kinds of neuroscience studies alone to fully and faithfully chart the mental effects of a variety of psychoactive or psychedelic substances. Hence, to ignite rapid progress, a radically different approach may be in order.

In the present investigation, we expanded on previous research on thousands of psychedelic reports and word usage patterns in freeform text reports (Ballentine et al., 2022; Martial et al., 2019; Zamberlan et al., 2018) to chart the changes of mental processes caused by 30 different examined psychoactive and psychedelic substances. We aimed to explore and characterize the whole semantic design space relevant to psychedelic experiences, rather than a handful of narrow questionnaire items, as research did previously. We found patterns in reports based on LLM-generated questions, giving finer insight into reported changes in conscious awareness when participants were undergoing these drug experiences. We found that this fine-grained insight could be used to the extent that predictions on the type of substance a test report may be possible based on questionnaire scales alone; even disregarding the actual free-form experience accounts. We derived almost 200 semantic dimensions of psychedelic drug experience in an organic fashion. In short, our study designed LLM analysis pipelines to establish and interrogate the target cognitive dimensions for each of the 30 psychedelic drugs in a full bottom-up fashion, articulated appropriate query questions for automated annotation, and performed impartial grading of psychedelic reports.

The study utilizes a large dataset of psychedelic experience reports from the Erowid database, covering 30 different psychedelic substances. LLMs, specifically ChatGPT, were employed to manipulate questions that explore a variety of dimensions of psychedelic experiences. This process involved three main steps: identifying target experience dimensions, crafting detailed questions, and automatically rating the presence of these dimensions in the reports.

Repository of free-form text reports of psychedelic experience

Psychedelic reports used in this work were obtained from the Erowid database (Erowid) - to our knowledge, the largest resource of its kind available. It archives reports with first-hand accounts of psychedelic experiences. Erowid is a member-supported organization that aims to provide unbiased information about psychedelic drugs. The archived reports have been submitted by mostly anonymous volunteers. The only consistent information that occurs in all of the reports is metadata that includes the date of submission and the kind of substance taken, which we used for parsing the text files. Erowid has received over 100,000 experience reports, and published tens of thousands of those. Their stated goal is, “We believe that truth, accuracy, and integrity in publishing information about psychoactives will lead to healthier and more balanced choices, behaviors, and policies around all psychoactive medications, entheogens, herbs, and recreational drugs.” (Erowid). We obtained a dataset of 6,689 psychedelic experience reports. We here used 1,373 reports. Each report corresponded to one of 30 types of psychedelic drugs, which provided the basis arsenal for LLM-enabled investigation.

LLM architecture

When designing questions to query the psychedelic reports, we took measures to minimize individual bias in any particular direction in the composition of the final panel of query questions. To this end, we used GPT4 to conjure up a pool of target questions. In a similar manner, we used gpt-3.5-turbo to annotate reports decreased personal bias, in an automated and impartial fashion. The LLM gpt-3.5-turbo is a regularly updated, low latency, low cost model. At the time of writing, the default gpt-3.5-turbo accepted 4096 tokens (OpenAI; OpenAI Platform).

In particular, we brought to bear LLMs for the purpose of deepening our understanding of psychedelic drug experiences in three complementary ways: 1) to identify a vast portfolio of target experience dimensions of psychedelic use, 2) to verbalize questions that we could use to effectively query and quantify these experience dimensions in the freeform reports, and 3) to automatically rate the presence of these experience dimensions in selected experience reports, with the use of the repertoire of the crafted query questions. The first and second steps both used GPT4, while the third process was carried out with the OpenAI API using gpt-3.5-turbo.

Leveraging LLMs to explore, formulate and grade dimensions of psychedelic experience

In step 1, to identify the target experience dimensions to be used, our approach started by asking the LLM open questions about what the core dimensions of psychedelic experiences are. In the ensuing answers, the numbers of LLM-proposed dimensions or aspects of changing subjective awareness were typically limited to around 5–8. To enrich the initial response set and branch out more, into always more fine-grained details of semantic dimensions, we then recursively asked about aspects of those semantic dimensions already mentioned by the LLM. For example, we started with asking for factors people describe during psychedelic experiences. The response included “mystical experiences”. We then queried what aspects describe mystical experiences in psychedelic experiences, which included topics such as “transcendence of ordinary reality”, which then was the starting point for further queries, and so forth. It turned out that there was a natural saturation point to this iterative process of branching out, since deeper queries would reiterate already previously disclosed candidate dimensions. This protocol naturally led to a pool of hallucinogenic experience dimensions, generally not requiring more than 3 to 4 levels of nested prompting until reaching saturation.

In step 2, after we had identified the pool of target dimensions of psychedelics experience, we then formulated several concrete prompt questions to cover these experience dimensions in our to-be-analyzed corpus of experience reports. We aimed to craft a set of differently phrased questions, aiming to emphasize potentially different nuances, getting at an identical experience dimension from several different angles. We thus tasked the LLM to devise 5 different prompts that rate a given generated dimension (cf. step 1) on scales from 1 (least present) to 10 (most present), such as with “transcendence of ordinary reality”. If the question was not applicable to the report at hand, we explicitly instructed the LLM in our prompt to indicate a 0 for such a response.

As an example illustrating step 1 and 2, let’s consider who we asked the LLM about “emotional transformation”. This facet originated from “repressed memories” coming up from inquiring about things people experience during psychedelic drug sessions. Then “emotional resonance” in turn surfaced when probing the LLM about repressed memories in a text report. Finally, emotional transformation appeared when querying the LLM for 5 prompts to further delve into emotional resonance. This protocol naturally led to a pool of 176 unique questions, articulated by the LLM itself.

For step 3, after we had built a corpus of articulated questions covering a wide range of semantic dimensions, we subsequently submitted these questions to the LLM to finally perform an investigator-independent rating of each dimensions in each report of psychedelic experience. To obtain numerical results, the dimension ratings, from the LLM, we used a combination of setting the system role, one-shot learning, and instructional learning (Lin et al., 2022; OpenAI Platform, n.d.; Wei et al., 2023). Individual reports were separated from each other on the basis of the metadata of each report. Reports were also parsed into paragraphs. Reports which were too long for the LLM (that is, too many tokens, 4097, for the LLM–given with an estimate of ¾ the number of words being the number of tokens, as stated in the introduction) were cut into separate reports. For each of the 30 target drug compounds, we aimed to use 50 reports (cf. Supplementary Materials).

On a practical level, to run the automated rating, we set the system role as, “You are an assistant who only responds with a single float number.” The one-shot learning was carried out with the user role given, “Rate the happiness in the text 'I am super happy.' on a scale from 1 to 10.”, with the assistant answering, “10.0”. We added after each question in the final user prompt, “Output only a number from 1–10. If the question is not applicable to the text, output a 0.0.” We then used regular expression queries (regex module in Python, ‘float(re.findall(r"\d+\.\d+|\d+", response)[0])’) to find and extract the float or integer in the response text. In manual reviews of query responses, we confirmed that it was always the first number in the response that was the desired rating numerical value. Any outputs found to be higher than 10 were clipped to 10, which happened less than a handful of times across all conducted queries. No negative values were found in any responses from the LLM. These ancillary analyses confirmed the robustness querying LLMs for quantitative ratings on verbalized question queries – these 176 rating number vectors per report provided the basis for our subsequent analysis steps.

As a measure to further enhance the robustness of our findings and conclusions, the above report rating process was repeated in 10 separate iterations for each experience report and for each of the 176 questions. To this end, we set the temperature parameter of the LLM (OpenAI Platform) at 0.5 to allow for some natural variation in responses, that is, to induce some voluntary stochasticity in the LLM processing and outputs. As a side note, the temperature not being 0 may have played a part in regex being needed to check for numbers (cf. above). Despite the voluntarily increased volatility in the sentence responses, the LLM generally still gave the desired rating values. Up to 10 attempts were made per query (with catches and pauses for various OpenAI feedback codes, including token rate limits per minute) to check that we obtained the correct numeral format response. The input was limited by a maximum number of tokens, 4096 (a token is about ¾ of a word in the OpenAI APIs) (OpenAI Platform) during the data acquisition. Due to the 10 replications for each question query (see Table 1 in the Supplementary Material for more details), given our 176 questions and 1373 reports (176x1373x10), we retrieved a total of 2,416,480 question rating query outputs.

Predictive validation of the LLM annotations

Next, we formally evaluated whether LLM-generated report annotations, ignoring the actual text content of the reports, contain sufficient information to provide a rich picture of the drug experience to be able to reliably identify the drug that was actually taken by the participant. This evaluation was based on new, unseen reports that our algorithm had not seen yet. To that end, repeated ratings for each report were reshaped for grouping in reports per substance, such that, information-derived from a given experience report only occurred in the model training data or the model testing data.

More specifically, we carried out linear one-vs-rest classification (logistic regression, maximum number of iterations set at 600). We used a group (groups are repeated questions for the same report) shuffle splitter (10 splits and a 10% test size), with 30 repeats in a loop. Thus, the obtained ratings from repeats of questions were shuffled within groups, and train-test splits of the data were randomized for 10 data splits. Thus these 10 randomized data splits were used in cross validation, and this process was repeated 30 times. The mean and standard deviation of the classifier prediction scores were taken to evaluate model performance, reflecting the out of sample cross validation estimate that reflects the expected future performance. To evaluate the systematic correct and deviated guess of our trained classifier (confusion matrix with the logistic regression classifier), we used 10 grouped (repeats in questions grouped) K folds to get prediction on independent reports to then use in a confusion matrix. In effect, ratings were grouped in reports, and a logistic regression classifier was run with groups in a group shuffle splitter and a cross validation scorer to give a mean score, and in grouped k-folds and cross validation predictions to be used in a confusion matrix. Importantly, in so doing, information pertaining to one of the 176 target questions per report only appeared in either the training observations for model parameter estimation or the test observations for evaluation of the fitted model instance.

To understand the importance and role of the obtained parameters of our classification model, after ascertaining that robust predictive principles can be derived from the LLM-derived information alone (cf. above), we wanted to introspect how and what predictive principle led to the achieved prediction performance. To that end, we ultimately fit the totality of the input data with the logistic regression classifier to obtain the model coefficient estimates that parameterize the decision boundary. These coefficients enable understanding the importance of questions used in classification. Models used were from scikit-learn (scikit-learn: machine learning in Python — scikit-learn 1.4.0 documentation).

Scientific computing implementation

Python was used for computing the data analysis protocols and for data visualization. The scikit-learn package provided the machine learning implementations (scikit-learn: machine learning in Python — scikit-learn 1.4.0 documentation). Data visualization was done with the Matplotlib (Matplotlib — Visualization with Python) and Seaborn (Waskom, 2021) packages.

By means of a state-of-the-art LLM, we have generated 176 target experience dimensions for psychedelically induced changes in subjective awareness, and their corresponding prompt questions. We also used an LLM to automatically answer these questions by means of quantitative rating scales, without intervention by a human investigation. The automated pipeline yielded ratings on a scale from 1 to 10, on complementary semantic facets in rich psychedelic experience reports, drawn from the largest psychoactive drug experience report collection of its kind, Erowid (Erowid).

To give a general overview of how the LLM automatically quantified the presence of a particular semantic dimension in experience reports, we took the mean value of the ratings for all reports per each of the 30 drugs. The means yielded 176 quantitative indicators of experience facets for each of the experience reports from 30 target drugs across all reports for each substance (Fig. 1). Each experience report was thus characterized by a 176-item experience profile that forms the basis for our subsequent analysis steps.

As a preparatory step, we quantified how expressions of semantic dimensions group among each other across drugs, and how drugs group among each other across semantic dimensions; we set out by carrying out a clustering analysis (based on average question ratings across reports per drug).

In Fig. 1, the matrix of mean rating per substance across all reports per question is input to the clustering algorithm (clustering is done with the ‘complete’ method with ‘cityblock’ metric with seaborn clustermap). Clusters were computed for each axis, so both questions and drugs that have similar mean ratings are grouped together, as also seen in the branches of the dendrograms. The 1st-level branches of questions encompassed about half of all questions each, and the 2nd-level branches of each of those two 1st-level branches included also around half of each (question sets matter, not their sequence). Although successive questions as numbered may be somewhat related due to the order they were created in, the significance of the specific numerical order of them in branches is not of direct relevance (numerical ordering of the questions in the clustermap are shown in Supplementary Materials).

To obtain a synoptic summary of the emerged grouping structure obtained on this initial overview analysis, using the emerged cluster sets of questions, we queried our LLM to give 5 keywords for each of the four 2nd-level branches of questions, yielding directly interpretable high-level descriptions. Keywords for the overarching four question sets were: i) mythology, supernatural, perception, connection, and auditory for the far-left 2nd-level branch; ii) spirituality, perception, memory, supernatural, and sensory for the mid-left 2nd-level branch; iii) psychedelic, transformation, perception, symbolism, and sensory for the mid-right 2nd-level branch; and iv) transcendence, perception, connection, supernatural, and emotions for the far-right 2nd-level branch. This shows how some groups of experience dimension questions were more prone to obtain high ratings for psychedelic experiences than others. There were a few higher-level larger groups of drugs based on LLM-enabled annotations, with a couple individual drugs sticking out as top-level clusters. Not all questions trigger high LLM ratings, we thus have specific findings for each question.

All drugs showed at least some questions directly applied to them, showcasing the use of the wide array of dimensions the questions cover. These exploratory analyses across 4 major experience dimensions (according to the question clustering) showed that certain themes gain more traction (higher mean ratings) than others for analyzing psychedelics across observations. For example, sensory perception was shown to be more useful than mythology, as seen in the aforementioned 2nd-level branches. It can be seen that the “classic” psychedelics, including psilocybin, mescaline, DMT, and LSD (Johnson et al., 2019), tend to get higher ratings, across experience reports, with most of them in the cluster with the highest ratings. Further, MDMA noticeably stuck out as exhibiting lower rating scores than other drugs, like with $$ psychedelic drugs (Tagen et al., 2023). However, MDMA still rated relatively high in questions like 171, which describes physical context. Moreover, DiPT was shown to be somewhat unique with higher ratings in a range of questions (78, 156, 82, 161) that relate to auditory changes, as would be expected based on existing knowledge (Carbonaro et al., 2013). Question 5, which refers to revelation of repressed memories, successfully captured a prominent aspect of ibogaine. This resonates well with existing literature that emphasizes use cases to include memory retrieval (Underwood et al., 2021). Thus, we can see that there are definite patterns in how LLM-based ratings on semantic dimensions are set per substance.

The similarity patterns between questions concerning substances is further detailed in their similarities in ratings across psychedelic reports (Fig. 2). The mean rating for each report per question was taken, and the correlation between questions across reports was computed – the actual substance was not taken into consideration in this analysis step. This approach gave us a view of the degree of similarity or differences of questions when compared to each other across our corpus. Groups of questions clearly resonated with each other across reports (Pearon’s correlation coefficient ρ), showing that experiences related in reports share some aspects, with a few of the correlations exceeding 0.8. At the same time, some questions that correlate with few others showed the diversity in the semantic dimensions that the questions track in our corpus of psychoactive drug experiences. While there was no specified order for the questions to appear in, questions over similar topics tended to be grouped together, as they were generated in branches organically using ChatGPT (cf. methods section).

The core conclusion to take from this is that we did not see a lot of near perfect correlations; hence, each question was unique in at least some way and sets of questions featured certain overlap in what aspects or shades of experience they track. In other words, the high correlations showed that the LLM worked, as we expect some overlap, which then gave credence to the others lacking much correlation, showing the wide array of dimensions covered. Thus, the wide spectrum of 176 total questions appeared fully justified - a great starting point for the present investigation.

Considering patterns that emerged in how experience reports tracked distinct components of experiences with how questions were automatically answered by LLMs, we next interrogated whether predictions can be made about the kind of substance that was ingested based on the LLM-obtained ratings of 176 questions alone. Logistic regression classification was used since it takes all questions into account and the coefficients help us understand the importance of individual questions in how successful classification was achieved. By performing drug predictions with machine learning, we can test if certain semantic aspects (i.e., their questions) were discriminative from others and whether topics could be extracted from how sets of experience dimensions tell the 30 psychedelic drugs apart. Some drugs elicited similar episodes, but some experiences were unique enough to where the model could figure out which substance was being written about with a good probability, based on LLM ratings alone, disregarding the actual experience report itself. Specifically, number ratings were used, that inquire into semantic dimensions of reports, with no specific mention of which psychedelic drug was taken (an example is displayed in Fig. 3). Logistic regression classifier was run with each report considered a group, due to question repeats, but with a non-zero temperature. Using a group shuffle splitter (10 splits and a 10% test size), with 30 repeats while shuffling values within reports each time, with a cross validation scorer gave a mean classification performance of 22% +/- 2% (standard deviation across folds). This discrimination performance largely exceeded the random chance level at ~ 3.3%. Using 10 k-folds and a cross validation predictor for a confusion matrix (Fig. 3), MDMA stuck out as the most distinctive drug as measured by classification performance, with DiPT coming in at 2nd place. Ibogaine also emerged to be especially unique. Aside from showing how discriminative each substance’s set of LLM ratings may be, this analysis also revealed how certain drug pairs may be confused for each one another by the trained classifier, such as with DMT and 5-MeO-DMT. In summary, we were able to use patterns in how the questions score to predict what substance a report is based on with a probability significantly higher than chance. Unseen reports can be automatically and correctly categorized by our trained classification algorithm into a drug class based on LLM-derived semantic dimensions of psychedelic experience reports.

To show an example of how questions are answered similarly, Fig. 5 shows the two most correlated questions, again using the mean ratings for each report. As seen in the full questions in the Supplementary Materials, question 72 rates the depth of surrendering to experience in the text, while question 74 rates the portrayal of relinquishing control in the reports. While worded differently, the meaning is expected to be fairly similar, which shows the comprehension abilities of the LLM. Top correlations for all questions are shown in the supplementary materials. The correlations serve as another piece of evidence speaking to the validity of the results.

Existing clinical studies with psychedelics have typically involved only 1–2 substances, a handful of question items, and a limited number of participants compared with population-level use. Cutting across dozens of drugs and many hundreds of sessions, a disciplined multi-drug investigation is an unavoidable prerequisite to be in a position to reveal the general principles that underlie these multifaceted, pharmacologically mediated changes in subjective awareness – a research opportunity that only emerged recently (Ballentine et al., 2022). Our computational neuroscience investigation has demonstrated that psychedelic reports can be interrogated at scale by large language models. This was based on automatically evaluating a vast portfolio of questions that look deep into the core structure of the report semantics, where the target questions were also conceived by LLMs in a bottom-up fashion. To validate our proposed framework, we showed how these experiential component dimensions group among 30 drugs. We also showed that they contain dense information to successfully tell apart what exact substance was used in never-seen psychedelic experience reports, from its semantic structure alone, ignoring the text itself. Our analytical protocol could be applied to other non-ordinary states of consciousness, such as meditation hypnosis, dreaming, and near-death experiences (Timmermann et al., 2023; Timmermann et al., 2018), and other fields of psychology.

That is, by bringing to bear LLM-based techniques, we have achieved improvements upon research practices based on questionnaires that can be useful for cognitive testing in general, not only in psychedelics. Traditionally, cognitive questionnaire batteries have been developed with the underlying goal of capturing a priori defined categories of psychological experiences, a process commonly referred to as construct operationalization — a fundamental part of research methodology in the social sciences for decades. In the use case of psychedelics research, qualitative experiences under drug influence have traditionally been assessed by summing questionnaire items, assuming this sum indexes an overarching psychological construct tracking interindividual variation believed to be captured by the chosen set of assessment items. Consequently, commonly measured outcomes in classical psychedelics research are confined to the psychological construct intended to be measured by the selected rating scale. However, some of the most widely used scales in psychedelics research were reported to be highly heterogeneous and have suboptimal construct validity (Hovmand et al., 2024) – in other words, many incumbent questionnaires are not capturing faithfully what humans designed them to measure. Furthermore, biases toward the instrumentalization and measurement of positive feelings may obscure other relevant, not a priori considered experience dimensions such as pathological outcomes like the exacerbation of depressive symptoms. In fact, many popular psychedelic questionnaires simply do not attempt to measure negative experiences (Hovmand et al., 2024), such as the MEQ30 (30 question MEQ), the HRS, the Hood Mysticism Scale (M-scale), and the After the Spiritual Experience Questionnaire (ASEQ). Unlike the broader perspectives typically found in unconstrained testimonies, self-reported questionnaires, wherein responses rely on the specific questions posed, often lack the capacity to effectively track drug-induced shifts in cognitive patterns and sensory experiences, with sensitivity to the specific contexts. Capitalizing on a wide array of reports, annotated for semantic dimensions in an impartial, investigator-independent way by machine learning algorithms, has the potential to alleviate the risk of operationalization bias — the tendency for researchers to introduce undue skewing through variable definition and measurement in clinical research concerning psychedelics and other areas.

An additional limitation of classical participant-reported questionnaires in psychedelics research pertains to confirmation biases, where participants' responses to questions are influenced by their preexisting belief systems or a priori expectations regarding psychedelics (Bremler et al., 2023). Testimonies and interviews in conjunction with data-driven annotation could provide an alternative approach that is less biased and less dependent on preconceived responses. Using an LLM approach, we can get a general idea of how humans, on average, are inclined to rate our experience reports based on such questions. Asking humans to do that task would have taken much more time, perhaps in the order of a few months, would have incurred more individual biases, and probably would have been a rather tedious task. LLMs can be used to easily scale up the rating process, as well as improve the scope and breadth of questions being asked, in comparison to working with human annotators.

Additionally, in classical questionnaire-based research, the potential for recall bias poses a considerable concern. This is because study participants may inaccurately recollect their psychological experiences at the moment when they are prompted (Wolff et al., 2022). This systematic bias when remembering or reporting past events, experiences, or encountered information (Sackett, 1979) can manifest due to various factors, such as imperfect memory retrieval, cognitive biases, external influences, personality traits, and medical comorbidities. For example, individuals with bipolar disorder and mild depression have shown a tendency to overestimate negative affect post-experience (Colombo et al., 2020; Ebner-Priemer et al., 2006). Emotional factors are also documented to play a role in memory recall, as participants tend to remember emotionally charged events more vividly or even inaccurately due to the intensity of the experience (Fahmy et al., 2006). Using real-time psychedelic reports, written during the experience, can help mitigate recall bias, as participants express their psychological experiences in the moment, without the need for retrospective commentary. For the purpose of the present study, the Erowid archive offered experience reports with live text notes, written during the experience, as participants underwent the psychedelic experience (Erowid). Such an approach diverges from the conventional method of using post-experience questionnaires prevalent in much of psychedelic research, thus offering an advantage in overcoming certain forms of recall bias.

Another common issue is limited validation of questionnaire instruments, which results in questionnaires that lack the necessary validation for consistent use across studies (Zuljevic et al., 2022). For example, the HRS was insufficiently validated based on a small and homogenous participant sample, potentially hindering its ability to comprehensively assess psychedelic experiences. Indeed, this questionnaire has been found to lack reliability given inconsistent results (Riba et al., 2001). In contrast, employing an algorithmic framework like that proposed in the present study, which empirically models drug effects as a function of a vast array of question rating outcomes, allows for a direct practical validation of specific designed questions through continual refinement. By benchmarking the predictive power of batches of candidate questions after training on a plethora of reports, LLMs emerged as an effective validation tool for future studies as both the most useful questions are honed in on, leaving others out, while also being expand upon key aspects as appropriate. The merit of such algorithmic question validation approach will only grow as more reports will become available for model training in the future.

Furthermore, the reliance on just a single session of questionnaire assessments for each case to be considered, as in classical psychedelics studies on human experience under drug influence (Carhart-Harris et al., 2016), raises concerns about the reliability and generalizability of the ensuing findings. The common “single-shot” approach may overlook the inherent variability in the actual judgment of the rating process as well as the observations sampled from each case. Hence, rating outcomes can be inconsistent across rating individuals but also between multiple experience session of the same participant (Ko et al., 2022). Leveraging LLMs now offers the possibility for evaluating drug experiences across multiple rating queries on the same case and multiple sessions from the same case to be rated in a streamlined and time-efficient way. In this spirit, we here carried out 10 separate ratings on each question in each experience report (using the temperature hyper-parameter). Additionally, Erowid experience reports include instances where the same person submitted more than one report. There is no restriction from doing so, they just have to pass the review process each time (Erowid). Therefore, adopting a deliberate multi-rating repeated-sampling approach, helped by LLM techniques, can deepen our understanding of psychedelic drug experiences.

More broadly, language and cultural disparities among participants and in psychedelics questionnaire design pose even more challenges to progress in questionnaire-based research endeavors (Mosurinjohn et al., 2023; Zuljevic et al., 2022). Advancements in training LLMs in multiple languages could help mitigate cross-cultural influences in study design. This data-driven approach enables large-scale investigations beyond the confines of in-laboratory examinations, which often recruit participants from the local population or individuals affiliated with the investigator's institution. With this capability, researchers can directly administer identical questions articulated in different languages across various geographical regions, thus facilitating more comprehensive and globally inclusive studies. Automated validation processes facilitated by LLMs play a crucial role in ensuring the cultural and linguistic appropriateness of these questions. By automatically validating questions, LLMs help mitigate potential biases stemming from cultural and linguistic differences, thereby enhancing the reliability and validity of the research findings, while also streamlining the research process for greater efficiency. In a similar vein, as an alternative of or extension to Mechanical Turk experiments, internet-pretrained LLMs demonstrate remarkable 'role play' capabilities, allowing them to explore diverse perspectives in texts, which would be held by those of various societal subgroups (Bzdok et al., 2024; Taves, 2020). This capacity can help go beyond the sometimes narrow background of participants recruited in traditional in-lab research, which can fail to represent the general population.

We also aimed to embrace the full complexity of psychedelic experiences in terms of target questions to be evaluated and drug classes to be analyzed. Unlike many questionnaires that tend to overlook negative aspects of psychedelic experiences, such as feelings of fear and discomfort (Bremler et al., 2023; Wolff et al., 2022), our study aimed to encompass the wider spectrum of conscious awareness changes induced by psychedelics. To accomplish this, we incorporated a broad array of inquiries, covering both positive and negative aspects, which participants may find challenging to address directly within the confines of a traditional experimental setting. Further, with dozens of different substances covered with LLM annotation, which is generally not the case with the tremendous logistic and financial requirements of fMRI experiments under drug influence or in clinical trials, we could show questionnaire assessments to be generalizable to new drug experience (Carhart-Harris et al., 2012, 2016, 2017). This is partly explained by the fact that LLMs can comfortably quantify semantic dimensions or aspects of psychedelic experience across a large number of target drug compounds.

In conclusion, we have introduced analytical strategies for investigators who wish to harness the power of emerging LLM technologies in psychology and neuroscience: as creative generators of candidate questions that get at multiple facets of colorful accounts of psychedelic sessions, as well as scalable annotators of such textual reports at generous scale. In so doing, we overcame several challenges of ordinary hand-designed questionnaires in psychology and neuroscience, and alleviated various biases incurred by humans manually administering them. Also, psychoactive and psychedelic drug reports could be used to anchor unique facets of human subjective awareness in the neurotransmitter receptor pathways that psychotropics act on. Our findings highlight the potential for LLMs to enhance the precision and scope of future studies in the field of consciousness research. The marriage of AI and psychedelics as a mechanistic window into altered states of consciousness may also prove invaluable for designing novel drug therapeutics.

Alyafeai, Z., AlShaibani, M. S., & Ahmad, I. (2020). A Survey on Transfer Learning in Natural Language Processing.
Ballentine, G., Friedman, S. F., & Bzdok, D. (2022). Trips and neurotransmitters: Discovering principled patterns across 6850 hallucinogenic experiences. Sci Adv, 8(11), eabl6989. https://doi.org/10.1126/sciadv.abl6989
Bansal, P., & Sharma, A. (2023). Large Language Models as Annotators: Enhancing Generalization of NLP Models at Minimal Cost.
Barrett, F. S., Bradstreet, M. P., Leoutsakos, J. S., Johnson, M. W., & Griffiths, R. R. (2016). The Challenging Experience Questionnaire: Characterization of challenging experiences with psilocybin mushrooms. J Psychopharmacol, 30(12), 1279-1295. https://doi.org/10.1177/0269881116678781
Barrett, F. S., Johnson, M. W., & Griffiths, R. R. (2015). Validation of the revised Mystical Experience Questionnaire in experimental sessions with psilocybin. J Psychopharmacol, 29(11), 1182-1190. https://doi.org/10.1177/0269881115609019
Bayne, T., & Carter, O. (2018). Dimensions of consciousness and the psychedelic state. Neurosci Conscious, 2018(1), niy008. https://doi.org/10.1093/nc/niy008
Bremler, R., Katati, N., Shergill, P., Erritzoe, D., & Carhart-Harris, R. L. (2023). Case analysis of long-term negative psychological responses to psychedelics. Sci Rep, 13(1), 15998. https://doi.org/10.1038/s41598-023-41145-x
Bzdok, D., Thieme, A., Levkovskyy, O., Wren, P., Ray, T., & Reddy, S. (2024). Data science opportunities of large language models for neuroscience and biomedicine. Neuron, 112(5), 698-717. https://doi.org/10.1016/j.neuron.2024.01.016
Bzdok, D., & Yeo, B. T. T. (2017). Inference in the age of big data: Future perspectives on neuroscience. NeuroImage, 155, 549-564. https://doi.org/10.1016/j.neuroimage.2017.04.061
Carbonaro, T. M., Forster, M. J., & Gatch, M. B. (2013). Discriminative stimulus effects of N,N-diisopropyltryptamine. Psychopharmacology (Berl), 226(2), 241-246. https://doi.org/10.1007/s00213-012-2891-x
Carhart-Harris, R. L., Muthukumaraswamy, S., Roseman, L., Kaelen, M., Droog, W., Murphy, K., Tagliazucchi, E., Schenberg, E. E., Nest, T., Orban, C., Leech, R., Williams, L. T., Williams, T. M., Bolstridge, M., Sessa, B., McGonigle, J., Sereno, M. I., Nichols, D., Hellyer, P. J., . . . Nutt, D. J. (2016). Neural correlates of the LSD experience revealed by multimodal neuroimaging. Proc Natl Acad Sci U S A, 113(17), 4853-4858. https://doi.org/10.1073/pnas.1518377113
Carhart-Harris, R. L., Roseman, L., Bolstridge, M., Demetriou, L., Pannekoek, J. N., Wall, M. B., Tanner, M., Kaelen, M., McGonigle, J., Murphy, K., Leech, R., Curran, H. V., & Nutt, D. J. (2017). Psilocybin for treatment-resistant depression: fMRI-measured brain mechanisms. Sci Rep, 7(1), 13187. https://doi.org/10.1038/s41598-017-13282-7
Carhart-Harris, R. L., Wagner, A. C., Agrawal, M., Kettner, H., Rosenbaum, J. F., Gazzaley, A., Nutt, D. J., & Erritzoe, D. (2022). Can pragmatic research, real-world data and digital technologies aid the development of psychedelic medicine? J Psychopharmacol, 36(1), 6-11. https://doi.org/10.1177/02698811211008567
Colombo, D., Suso-Ribera, C., Fernández-Álvarez, J., Cipresso, P., Garcia-Palacios, A., Riva, G., & Botella, C. (2020). Affect Recall Bias: Being Resilient by Distorting Reality. Cognitive Therapy and Research, 44(5), 906-918. https://doi.org/10.1007/s10608-020-10122-3
Dougherty, R. F., Clarke, P., Atli, M., Kuc, J., Schlosser, D., Dunlop, B. W., Hellerstein, D. J., Aaronson, S. T., Zisook, S., Young, A. H., Carhart-Harris, R., Goodwin, G. M., & Ryslik, G. A. (2023). Psilocybin therapy for treatment resistant depression: prediction of clinical outcome by natural language processing. Psychopharmacology (Berl). https://doi.org/10.1007/s00213-023-06432-5
Ebner-Priemer, U. W., Kuo, J., Welch, S. S., Thielgen, T., Witte, S., Bohus, M., & Linehan, M. M. (2006). A valence-dependent group-specific recall bias of retrospective self-reports: a study of borderline personality disorder in everyday life. J Nerv Ment Dis, 194(10), 774-779. https://doi.org/10.1097/01.nmd.0000239900.46595.72
Erowid.
Fahmy, S., Cho, S., Wanta, W., & Song, Y. (2006). Visual Agenda-Setting After 9/11: Individuals' Emotions, Image Recall, and Concern With Terrorism. Visual Communication Quarterly, 13(1), 4-15. https://doi.org/10.1207/s15551407vcq1301_2
Garcia-Romeu, A., Davis, A. K., Erowid, F., Erowid, E., Griffiths, R. R., & Johnson, M. W. (2019). Cessation and reduction in alcohol consumption and misuse after psychedelic use. J Psychopharmacol, 33(9), 1088-1101. https://doi.org/10.1177/0269881119845793
Griffiths, R. R., Johnson, M. W., Carducci, M. A., Umbricht, A., Richards, W. A., Richards, B. D., Cosimano, M. P., & Klinedinst, M. A. (2016). Psilocybin produces substantial and sustained decreases in depression and anxiety in patients with life-threatening cancer: A randomized double-blind trial. J Psychopharmacol, 30(12), 1181-1197. https://doi.org/10.1177/0269881116675513
He, X., Lin, Z., Gong, Y., Jin, A.-L., Zhang, H., Lin, C., Jiao, J., Yiu, S. M., Duan, N., & Chen, W. (2023). AnnoLLM: Making Large Language Models to Be Better Crowdsourced Annotators. https://doi.org/10.48550/ARXIV.2303.16854
Hovmand, O. R., Poulsen, E. D., & Arnfred, S. (2024). Assessment of the acute subjective psychedelic experience: A review of patient-reported outcome measures in clinical research on classical psychedelics. J Psychopharmacol, 38(1), 19-32. https://doi.org/10.1177/02698811231200019
How Studies Get Psychedelics: A Guide to Clinical Trials.
Huang, Z., Xu, W., & Yu, K. (2015). Bidirectional LSTM-CRF Models for Sequence Tagging.
Johnson, M. W., Garcia-Romeu, A., Cosimano, M. P., & Griffiths, R. R. (2014). Pilot study of the 5-HT2AR agonist psilocybin in the treatment of tobacco addiction. J Psychopharmacol, 28(11), 983-992. https://doi.org/10.1177/0269881114548296
Johnson, M. W., Hendricks, P. S., Barrett, F. S., & Griffiths, R. R. (2019). Classic psychedelics: An integrative review of epidemiology, therapeutics, mystical experience, and brain network function. Pharmacol Ther, 197, 83-102. https://doi.org/10.1016/j.pharmthera.2018.11.010
Kettner, H., Gandy, S., Haijen, E., & Carhart-Harris, R. L. (2019). From Egoism to Ecoism: Psychedelics Increase Nature Relatedness in a State-Mediated and Context-Dependent Manner. Int J Environ Res Public Health, 16(24). https://doi.org/10.3390/ijerph16245147
Ko, K., Knight, G., Rucker, J. J., & Cleare, A. J. (2022). Psychedelics, Mystical Experience, and Therapeutic Efficacy: A Systematic Review. Front Psychiatry, 13, 917199. https://doi.org/10.3389/fpsyt.2022.917199
Lee, M. (2023). A Mathematical Interpretation of Autoregressive Generative Pre-Trained Transformer and Self-Supervised Learning. Mathematics, 11(11). https://doi.org/10.3390/math11112451
Maclean, K. A., Leoutsakos, J. M., Johnson, M. W., & Griffiths, R. R. (2012). Factor Analysis of the Mystical Experience Questionnaire: A Study of Experiences Occasioned by the Hallucinogen Psilocybin. J Sci Study Relig, 51(4), 721-737. https://doi.org/10.1111/j.1468-5906.2012.01685.x
Malte, A., & Ratadiya, P. (2019). Evolution of transfer learning in natural language processing.
Martial, C., Cassol, H., Charland-Verville, V., Pallavicini, C., Sanz, C., Zamberlan, F., Vivot, R. M., Erowid, F., Erowid, E., Laureys, S., Greyson, B., & Tagliazucchi, E. (2019). Neurochemical models of near-death experiences: A large-scale study based on the semantic similarity of written reports. Conscious Cogn, 69, 52-69. https://doi.org/10.1016/j.concog.2019.01.011
Matplotlib — Visualization with Python.
Mosurinjohn, S., Roseman, L., & Girn, M. (2023). Psychedelic-induced mystical experiences: An interdisciplinary discussion and critique. Front Psychiatry, 14, 1077311. https://doi.org/10.3389/fpsyt.2023.1077311
Nutt, D., Erritzoe, D., & Carhart-Harris, R. (2020). Psychedelic Psychiatry's Brave New World. Cell, 181(1), 24-28. https://doi.org/10.1016/j.cell.2020.03.020
OpenAI.
OpenAI Platform.
Pangakis, N., Wolken, S., & Fasching, N. (2023). Automated Annotation with Generative AI Requires Validation.
Peill, J. M., Trinci, K. E., Kettner, H., Mertens, L. J., Roseman, L., Timmermann, C., Rosas, F. E., Lyons, T., & Carhart-Harris, R. L. (2022). Validation of the Psychological Insight Scale: A new scale to assess psychological insight following a psychedelic experience. J Psychopharmacol, 36(1), 31-45. https://doi.org/10.1177/02698811211066709
Riba, J., Rodriguez-Fornells, A., Strassman, R. J., & Barbanoj, M. J. (2001). Psychometric assessment of the Hallucinogen Rating Scale. Drug Alcohol Depend, 62(3), 215-223. https://doi.org/10.1016/s0376-8716(00)00175-7
Roseman, L., Haijen, E., Idialu-Ikato, K., Kaelen, M., Watts, R., & Carhart-Harris, R. (2019). Emotional breakthrough and psychedelics: Validation of the Emotional Breakthrough Inventory. J Psychopharmacol, 33(9), 1076-1087. https://doi.org/10.1177/0269881119855974
Ross, S., Bossis, A., Guss, J., Agin-Liebes, G., Malone, T., Cohen, B., Mennenga, S. E., Belser, A., Kalliontzi, K., Babb, J., Su, Z., Corby, P., & Schmidt, B. L. (2016). Rapid and sustained symptom reduction following psilocybin treatment for anxiety and depression in patients with life-threatening cancer: a randomized controlled trial. J Psychopharmacol, 30(12), 1165-1180. https://doi.org/10.1177/0269881116675512
Sackett, D. L. (1979). BIAS IN ANALYTIC RESEARCH. In The Case-Control Study Consensus and Controversy (pp. 51-63). Elsevier.
Savelka, J. (2023). Unlocking Practical Applications in Legal Domain Proceedings of the Nineteenth International Conference on Artificial Intelligence and Law,
Schrimpf, M., Blank, I. A., Tuckute, G., Kauf, C., Hosseini, E. A., Kanwisher, N., Tenenbaum, J. B., & Fedorenko, E. (2021). The neural architecture of language: Integrative modeling converges on predictive processing. Proc Natl Acad Sci U S A, 118(45), e2105646118. https://doi.org/10.1073/pnas.2105646118
scikit-learn: machine learning in Python — scikit-learn 1.4.0 documentation.
Spriggs, M. J., Kettner, H., & Carhart-Harris, R. L. (2021). Positive effects of psychedelics on depression and wellbeing scores in individuals reporting an eating disorder. Eat Weight Disord, 26(4), 1265-1270. https://doi.org/10.1007/s40519-020-01000-8
Tagen, M., Mantuani, D., van Heerden, L., Holstein, A., Klumpers, L. E., & Knowles, R. (2023). The risk of chronic psychedelic and MDMA microdosing for valvular heart disease. J Psychopharmacol, 37(9), 876-890. https://doi.org/10.1177/02698811231190865
Taves, A. (2020). Mystical and Other Alterations in Sense of Self: An Expanded Framework for Studying Nonordinary Experiences. Perspect Psychol Sci, 15(3), 669-690. https://doi.org/10.1177/1745691619895047
Timmermann, C., Bauer, P. R., Gosseries, O., Vanhaudenhuyse, A., Vollenweider, F., Laureys, S., Singer, T., Mind, Life Europe, E. R. G., Antonova, E., & Lutz, A. (2023). A neurophenomenological approach to non-ordinary states of consciousness: hypnosis, meditation, and psychedelics. Trends Cogn Sci, 27(2), 139-159. https://doi.org/10.1016/j.tics.2022.11.006
Timmermann, C., Roseman, L., Williams, L., Erritzoe, D., Martial, C., Cassol, H., Laureys, S., Nutt, D., & Carhart-Harris, R. (2018). DMT Models the Near-Death Experience. Front Psychol, 9, 1424. https://doi.org/10.3389/fpsyg.2018.01424
Törnberg, P. (2023). ChatGPT-4 Outperforms Experts and Crowd Workers in Annotating Political Twitter Messages with Zero-Shot Learning.
Underwood, M. S., Bright, S. J., & Les Lancaster, B. (2021). A narrative review of the pharmacological, cultural and psychological literature on ibogaine. Journal of Psychedelic Studies, 5(1), 44-54. https://doi.org/10.1556/2054.2021.00152
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017). Attention Is All You Need. https://doi.org/10.48550/ARXIV.1706.03762
Vollenweider, F. X., & Kometer, M. (2010). The neurobiology of psychedelic drugs: implications for the treatment of mood disorders. Nat Rev Neurosci, 11(9), 642-651. https://doi.org/10.1038/nrn2884
Vollenweider, F. X., & Preller, K. H. (2020). Psychedelic drugs: neurobiology and potential for treatment of psychiatric disorders. Nat Rev Neurosci, 21(11), 611-624. https://doi.org/10.1038/s41583-020-0367-2
Waskom, M. (2021). seaborn: statistical data visualization. Journal of Open Source Software, 6(60). https://doi.org/10.21105/joss.03021
Wei, J., Tay, Y., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., Chi, E. H., Hashimoto, T., Vinyals, O., Liang, P., Dean, J., & Fedus, W. (2022). Emergent Abilities of Large Language Models. https://doi.org/10.48550/ARXIV.2206.07682
Wolff, M., Mertens, L. J., Walter, M., Enge, S., & Evens, R. (2022). The Acceptance/Avoidance-Promoting Experiences Questionnaire (APEQ): A theory-based approach to psychedelic drugs' effects on psychological flexibility. J Psychopharmacol, 36(3), 387-408. https://doi.org/10.1177/02698811211073758
Yu, D., Li, L., Su, H., & Fuoli, M. (2023). Assessing the potential of AI-assisted pragmatic annotation: The case of apologies.
Zamberlan, F., Sanz, C., Martinez Vivot, R., Pallavicini, C., Erowid, F., Erowid, E., & Tagliazucchi, E. (2018). The Varieties of the Psychedelic Experience: A Preliminary Study of the Association Between the Reported Subjective Effects and the Binding Affinity Profiles of Substituted Phenethylamines and Tryptamines. Front Integr Neurosci, 12, 54. https://doi.org/10.3389/fnint.2018.00054
Zuljevic, M. F., Buljan, I., Leskur, M., Kaliterna, M., Hren, D., & Duplancic, D. (2022). Validation of a new instrument for assessing attitudes on psychedelics in the general population. Sci Rep, 12(1), 18225. https://doi.org/10.1038/s41598-022-23056-5

There is NO Competing Interest.

SupplementaryInfos.docx

Download PDF

Version 1

posted

You are reading this latest preprint version

Large language models auto-profile conscious awareness changes under psychedelic drug effects

Status:

Version 1

Abstract

Figures

Introduction

Methods

Repository of free-form text reports of psychedelic experience

LLM architecture

Leveraging LLMs to explore, formulate and grade dimensions of psychedelic experience

Predictive validation of the LLM annotations

Scientific computing implementation

Results

Discussion

References

Additional Declarations

Supplementary Files

Status:

Version 1