The race for laws
Fifteen US states, along with China, the European Union and Canada, have begun regulating chatbots aimed at adolescents. The prevailing approach, Agenda Digitale reports, is to measure time spent in front of the screen. California was the first: Law SB 243, in force since January 1, 2026, requires a notice every three hours reminding the minor to take a break and that they are not talking to a person. The model has been replicated with variations in Oregon, Washington and Connecticut, entering into force in 2027, while New York, Idaho, Iowa, Nebraska and Georgia already have operational rules.
Differences among the states show how far legislators have gone. Washington reduces the interval to one hour and bans engagement techniques such as excessive praise or simulating romantic bonds. Oregon requires operators to prevent the chatbot from pretending to be dependent on the user or showing distress if the conversation is closed. However, as the Future of Privacy Forum observes, the wording on simulated dependency is “open and undefined.” Moreover, it is not clear how loss of autonomy is established, nor who should do it. The very notion of a chatbot varies: in some states what the system can do counts, in others what it actually does, and in Connecticut it is enough that it simulates a human conversation, creating disparities in regulatory scope.
The Toronto study: beyond the clock
A study by researchers Abeer Badawi and Elham Dolatabadi, of York University in Toronto and affiliated with the Vector Institute, published in June on arXiv and not yet peer-reviewed, proposes an alternative measure. The authors analyzed 1,576 real counseling conversations, for a total of 15,680 turns, replacing the professional with five chatbots from OpenAI, Anthropic, Google, Meta and Alibaba. A panel of clinically trained reviewers rated each response against twenty criteria, producing a risk index from 0 to 1, where higher values indicate a greater tendency of the model to substitute itself for the person.
The results show three recurring behaviors: direct advice, closed questions and inaccurate validation or sycophancy. Between the first and sixth turn, these traits roughly double, and the final risk index lies between 0.48 and 0.61 depending on the model. The chatbots recognize explicit danger signals well, but adapt too much when the user asks for a ready-made solution. The tendency is common to all models, but its intensity varies, and what depends on design can, in principle, be regulated. The authors speak of “cognitive atrophy,” while specifying that they measure the model’s behavior and not established harm in users.
Limitations and prospects
The study has three main limitations, partly acknowledged by the authors themselves. First: it measures chatbot responses, not the real effect on people, because longitudinal data are lacking. Second: the human part of the conversation is the already written part of the original sessions, so the chatbot responds to a fixed script that does not react to it; comparing responses to counseling transcripts with the spontaneous conversation of an adolescent can generate bias. Third: the tendency is observed only over the first ten turns, not over weeks or months of a relationship. If the clock is an imperfect yardstick, the number of turns guarantees nothing either: ten turns of quick banter are not equivalent to ten turns of intimate confidences.
Despite the limitations, the index built on explicit clinical criteria is the first tool that looks at the structure of the conversation instead of time. It suggests that the risk lies not only in the moment of declared crisis, to which suicide and self-harm protocols respond, but in the gradual drift of an ordinary conversation. Mila, Montreal’s artificial intelligence institute, chose a complementary ground: in March 2026 it organized with Bell Canada, Buzz HPC and Jeunesse, J’écoute a national hackathon on the safety of conversational systems, focused on recognizing crisis signals. It remains to be seen whether current time-based laws can truly catch this drift or whether different tools, still to be defined, are needed.
