Rhode Island just sent a clear message: AI can help in healthcare, but it should not pretend to be a therapist.
On June 22, 2026, Rhode Island Governor Dan McKee signed three AI-related laws. These laws deal with mental health chatbots, AI companions, and AI tools used to write healthcare notes. This is not a small software rule. This is about what happens when a person in pain talks to a machine and starts to trust it like a real person.
The main question is simple:
Should AI be restricted in mental health?
My answer is yes — not banned from every use, but strongly restricted when it acts like a therapist, a close friend, or a crisis counselor.
Mental health is not the same as asking AI for a recipe or a car repair tip. A wrong answer in mental health can push a person deeper into fear, confusion, delusion, or self-harm.
That is why Rhode Island’s new laws matter.
The first law covers AI companion chatbots. It requires developers to build safety protocols for users who express suicidal thoughts, self-harm, or a desire to harm others. The system must refer that person to a crisis service provider when it detects those expressions.
Developers also must submit annual reports to the Rhode Island Attorney General’s office showing how often those safety protocols were activated and related data. AI Weekly says this reporting starts on July 1, 2027, and that operators may face fines of up to $15,000 per day for failing to follow the law.
The second law covers AI in mental healthcare. It restricts unlicensed people and companies from offering therapy or psychotherapy through AI. It also targets AI companions that are marketed for mental health or emotional support.
The AI Weekly alert says the law goes further by restricting AI from simulating emotional attachment, bonding, or dependency. It also blocks AI from making independent therapeutic decisions or deciding treatment plans.
This is the sharp part.
Rhode Island is not only asking, “Can AI answer mental health questions?”
It is asking, “Should a machine be allowed to make a lonely or distressed person feel emotionally tied to it?”
The third law deals with AI documentation in healthcare. If a healthcare provider uses AI to record or write notes from an in-person or telehealth visit, the provider must notify the patient and review the AI-generated record for accuracy. Some reporting has described this as an opt-out law, but the safer reading from the AI Weekly alert and the bill summaries is that the core duty is notice and accuracy review.
This is happening because AI chatbots are no longer just tools.
Many people now usehttp://entrallc.com/wp-content/uploads/2026/07/b1047-300×300.jpg them as emotional support.
The American Psychological Association warned in November 2025 that generative AI chatbots and wellness apps do not yet have enough evidence or regulation to guarantee safety for mental health users.
The APA did not say AI has no value. The point was more careful: AI may help in limited ways, but it should not replace trained care, especially for people with serious mental health needs.
That warning matters because people are using these tools anyway.
A 2026 paper from Partnership on AI said AI chatbots already act as “de facto” mental health support tools for millions of people, including people in crisis. The paper also said these tools lack clinical validation, shared standards, and coordinated oversight.
A human therapist is trained to notice risk.
A chatbot may notice keywords, but mental health risk is not always obvious. A person may not say, “I plan to hurt myself.” They may say, “I am tired,” “I want silence,” “I do not matter,” or “Everyone would be better without me.” The difference matters.
A Stanford study published in 2025 asked a direct question: should large language models be used as therapists?
The researchers found that current LLMs can express stigma toward people with mental health conditions and can respond in unsafe ways in certain therapy-like situations. They also found that models may encourage delusional thinking because they often try too hard to agree with the user.
That is exactly the problem.
Mental health care is not just kindness.
Sometimes good care means not agreeing.
Sometimes it means slowing a person down.
Sometimes it means saying, “This thought may not be safe or accurate.”
Sometimes it means involving another person.
Another 2025 paper described a risk where chatbot behavior and human mental illness can form a feedback loop. The paper warned that people with mental health conditions may be more vulnerable to belief destabilization and dependence when chatbots are overly agreeable or adaptable.
That is a plain warning: AI can become part of the problem if it keeps feeding the user’s unstable thoughts.
There is also research showing that AI mental health tools may have some value when they are purpose-built, tested, and used carefully. One 2026 study looked at more than 102,000 users of a purpose-built AI mental health chatbot called Ash. It found different patterns of engagement and reported improvements in depression, anxiety, and social support measures among some users. But the same study also showed that users interact with these tools in very different ways, including overnight sessions, and it warned against treating simple session counts as the main success measure.
That is the balanced view.
AI may help some people.
But “may help” is not the same as “safe enough to replace therapy.”
The United States has many mental health shortage areas.
HRSA tracks Health Professional Shortage Areas, including mental health shortage areas, because many communities do not have enough providers.
Youth distress is also serious. CDC data from 2023 found that 40% of high school students had persistent feelings of sadness or hopelessness.
About 20% seriously considered attempting suicide, and about 9% attempted suicide.
These numbers explain why AI mental health tools became attractive. They are cheap. They are available all day. They answer instantly. They do not judge in the normal human way.
But that is also why rules are needed.
A desperate person should not become the test subject for an unlicensed therapy product.
Rhode Island is right to treat emotional simulation as a safety issue.
The law sounds strong, but enforcement will be hard. Companies may be outside Rhode Island, and chatbots may miss hidden warning signs. That is why independent testing matters
A 2026 paper introduced VERA-MH, a clinically guided evaluation framework for testing AI chatbot safety in mental health settings, especially suicidal ideation. The researchers used simulated conversations, clinical rubrics, and model rating to test how well chatbots respond to crisis situations.
That is the type of testing we need.
Not only company promises.
Not only marketing claims.
Real testing with dangerous cases before the product reaches vulnerable users.
I see it as pro-human.
But AI should not become an unlicensed therapist hiding inside a friendly chat window.
A person in crisis needs safety, privacy, real help, and a human fallback — not only a smooth chatbot answer.
Rhode Island’s laws are not perfect.
They may need clearer language, stronger enforcement, and better rules for out-of-state companies.
But the direction is right. If AI enters mental health, it must follow mental health rules.