Intimate insight: MDMA changes how people talk about significant others

Baggott and colleagues situate their study within a body of literature indicating that ±3,4-methylenedioxymethamphetamine (MDMA) produces acute prosocial effects such as increased empathy, interpersonal closeness and sociability. Prior experimental work has shown MDMA enhances recognition of positive emotion and reduces sensitivity to negative facial expressions, and clinical research has investigated its potential as an adjunct in psychotherapy for disorders such as post-traumatic stress disorder. The authors note that most controlled studies to date have relied on individual, computerized tasks rather than interpersonal interactions, and that analysing speech content during real social exchanges may reveal mechanisms underlying MDMA's social-emotional effects. The study therefore aimed to test whether MDMA alters speech production and content during a brief dyadic interaction about a close personal relationship. Specifically, Baggott and colleagues hypothesised that MDMA (1.5 mg/kg orally) would increase the amount of talking (total words), increase the proportion of emotional and social words, and allow machine learning methods to distinguish MDMA from placebo based on word usage. They combined a validated dictionary approach (LIWC) with an exploratory machine learning classifier to assess changes in speech content during a five-minute, semi-structured conversation about an important person in the participant's life.

This was a within-subject, double-blind, placebo-controlled study in which 35 healthy adults (12 female, 23 male; mean age 24.3±4.3 years) with light-to-moderate prior ecstasy experience (4–40 lifetime uses) completed two laboratory sessions. Sessions were separated by at least five days and the order of MDMA (1.5 mg/kg oral) and placebo was randomised. Participants were screened for medical and psychiatric exclusions, required to be fluent in English and within a BMI range of 18–30, and were tested for recent drug use and pregnancy on each session day. Sessions lasted 4.5 hours with physiological and self-report measures collected at baseline and repeatedly up to 4 hours post-dose; the talking task occurred at 11:00 to coincide with expected peak effects. The primary behavioural measure was a five-minute talking task adapted from an interpersonal perception procedure. Prior to experimental sessions, participants provided names of three important people; on each drug session a gender-matched research assistant selected one name at random and asked the participant to speak about that person while using reflective-listening techniques and minimising their own verbal contributions. Conversations were audio-recorded and professionally transcribed; any research assistant speech was removed. Speech was analysed by two complementary methods. First, a dictionary-based approach used Pennebaker's Linguistic Inquiry and Word Count (LIWC) to generate percentages of words in 43 preselected categories (affect, social interaction, cognition, time frame, perception, etc.). The authors also isolated descriptive phrases about the target person beginning with "he is"/"she is" and categorised those phrases as psychological, non-psychological (e.g. appearance), or relationship-focused; blinded raters applied these labels. Second, a machine learning approach used a bag-of-words strategy: names were removed, words were lemmatised, and occurrences of each word (1755 unique words) were used as predictors in random forest classifiers. Recursive feature elimination was applied to reduce predictors; out-of-bag accuracy and cross-validation guided model selection. Statistical analysis predominantly used mixed-effects models with participant as a random effect and drug condition as a fixed effect; significance was set at p≤0.05. To relate speech changes to self-report drug effects, the authors reduced significant LIWC variables via principal components analysis (PCA) with varimax rotation and then used LASSO regression to predict component scores from peak visual analogue scale ratings. Additional random-effects models explored relations between specific rotated components and selected self-report items.

Sample characteristics: data from 35 participants were analysed (34% female). Mean age was 24.3±4.3 years, BMI 22.7±2.7, and mean lifetime ecstasy uses 13.1±10.2. Most participants reported current alcohol, caffeine and marijuana use; 12 smoked tobacco. Dictionary-based analysis of whole transcripts indicated that MDMA altered word choice across several LIWC categories compared with placebo. MDMA increased the proportion of social and sexual words, increased language reflecting discrepancies and discussion of the future, and decreased use of relative terms and motion-related language. Participants also used more words relating to death under MDMA, although the magnitude of that increase was small. There were no significant gender effects. MDMA did not increase overall talkativeness: mean (±SEM) words spoken were 692.5±187.08 after placebo versus 685.7±171.14 after MDMA (no significant difference). Four placebo transcripts were lost to recording failure. When analysing isolated descriptive phrases beginning with "he is"/"she is", MDMA shifted content from psychological to more factual descriptions. The proportion of psychological phrases decreased from 76.0±3.73% after placebo to 56.7±3.91% after MDMA (F1,32.86=19.49, p<0.0001), while non-psychological phrases increased from 15.4±3.52% to 30.4±4.03% (F1,32.93=10.29, p=0.003). Relationship-descriptive phrases did not change significantly. Across these phrases, MDMA reduced use of body-related words (0.53±0.2 vs 0.0±0.0 words; F1,498=5.571, p=0.019) and increased words in cognitive-mechanism (12.2±0.92 vs 16.57±1.03; F1,498=11.516, p=0.001) and insight categories (0.51±0.16 vs 1.31±0.38; F1,498=4.503, p=0.034). To relate speech changes to subjective drug effects, the authors applied PCA to significant LIWC variables and retained three components explaining 29.0%, 24.0% and 17.6% of variance. These were labelled "motion/relative" (highly correlated with motion and relative scales), "discrepancy/future" (correlated with discrepancy and future scales), and "social" (correlated with social, sexual and death scales). LASSO regression using peak visual analogue scores found that all three components were predicted by a common set of self-report measures including ratings of feeling drug/high, elation, stimulation, loving, sociability and friendliness. Additional predictors varied by component: relative/motion was also predicted by playful and dislike-drug scores; discrepancy/future by playful and wanting more drug; and the social component by confidence, insightfulness, dizziness and wanting more drug. Further random-effects models showed the social component significantly predicted peak confidence (F1,23=17.28, p=0.0004) and insight (F1,23=9.89, p=0.0045), with a significant or trend interaction between social and relative/motion components in predicting confidence and insight. Machine learning classification using random forests and recursive feature elimination produced a final model using 300 predictor words with 2000 repeats of five-fold cross-validation. The classifier achieved an accuracy of 0.72, sensitivity 0.71 and specificity 0.80 in predicting dosing condition. Words with high importance included social terms (others, public, camaraderie, outgoing) and emotionally valenced words both positive (goofy, beautiful, cheer, fix) and negative (trouble, dead).

Baggott and colleagues interpret their findings as evidence that MDMA selectively alters the social and emotional content of speech during a brief semistructured interaction. Both the LIWC dictionary analysis and the machine learning classifier indicated increased use of social words under MDMA, consistent with prior data suggesting prosocial effects of the drug. The authors report that many speech changes co-varied with self-reported euphoria and sociability, supporting the notion that enhanced positive affect contributes to MDMA's social effects. Beyond euphoria-related sociability, the study identified a link between a speech-derived social component and self-reported confidence and feelings of insight. The investigators suggest this may reflect an additional cognitive phenomenon under MDMA—an altered sense of certainty or perceived insight—that contributes to socially relevant speech changes. Although MDMA did not increase total word count, it shifted descriptive phrases about important others from abstract psychological statements toward more concrete, factual details. The authors propose that this change in linguistic abstraction could indicate deeper or less superficial consideration of the target person, a pattern supported by increases in LIWC categories for cognitive mechanisms and insight. The discussion notes several limitations recognised by the study team. The machine learning bag-of-words approach ignored word order and context, so it captures only part of linguistic nuance; the authors recommend future analyses that include bigrams and trigrams. The talking task specifically asked participants to discuss psychologically important people, which limited the emotional range sampled; the researchers suggest future work could elicit particular emotional memories or employ experimentally induced emotional states (for example, stress paradigms) to broaden understanding. Finally, the authors acknowledge that confirmation in clinical populations is needed before generalising therapeutic implications. Overall, they conclude that combining natural language processing with traditional speech analysis offers a promising route to characterise how psychoactive drugs affect mental states and social interaction.