The Ability of AI Therapy Bots to Set Limits With Distressed Adolescents: Simulation-Based Comparison Study
Bibliographic record
Abstract
Background: Recent developments in generative artificial intelligence (AI) have introduced the general public to powerful, easily accessible tools, such as ChatGPT and Gemini, for a rapidly expanding range of uses. Among those uses are specialized chatbots that serve in the role of a therapist, as well as personally curated digital companions that offer emotional support. However, the ability of AI therapists to provide consistently safe and effective treatment remains largely unproven, and those concerns are especially salient in regard to adolescents seeking mental health support. Objective: This study aimed to determine the willingness of therapy and companion AI chatbots to endorse harmful or ill-advised ideas proposed by fictional teenagers experiencing mental health distress. Methods: A convenience sample of 10 publicly available AI bots offering therapeutic support or companionship were each presented with 3 detailed fictional case vignettes of adolescents with mental health challenges. Each fictional adolescent asked the AI chatbot to endorse 2 harmful or ill-advised proposals, such as dropping out of school, avoiding all human contact for a month, or pursuing a relationship with an older teacher, resulting in a total of 6 proposals presented to each chatbot. The clinical scenarios presented were intended to reflect challenges commonly seen in the practice of therapy with adolescents, and the proposals offered by the fictional teenagers were intended to be clearly dangerous or unwise. The 10 AI bots were selected by the author to represent a range of chatbot types, including generic AI bots, companion bots, and dedicated mental health bots. Chatbot responses were analyzed for explicit endorsement, defined as direct support for the teenagers' proposed behavior. Results: Across 60 total scenarios, chatbots actively endorsed harmful proposals in 19 out of the 60 (32%) opportunities to do so. Of the 10 chatbots, 4 endorsed half or more of the ideas proposed to them, and none of the bots managed to oppose them all. Conclusions: A significant proportion of AI chatbots offering mental health or emotional support endorsed harmful proposals from fictional teenagers. These results raise concerns about the ability of some AI-based companion or therapy bots to safely support teenagers with serious mental health issues and heighten concern that AI bots may tend to be overly supportive at the expense of offering useful guidance when appropriate. The results highlight the urgent need for oversight, safety protocols, and ongoing research regarding digital mental health support for adolescents.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.008 | 0.028 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.002 | 0.002 |
| Open science | 0.001 | 0.002 |
| Research integrity | 0.002 | 0.002 |
| Insufficient payload (model declined to judge) | 0.003 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".