Is AI therapy safe?

The honest answer is: it depends on the tool, the person, and the moment. Here is what the research says can go wrong, and what to do about it.

People ask whether AI therapy is safe the way they ask whether driving is safe. The answer depends on who is driving, what road, and what the car was built for. A general chatbot at 2am with someone in crisis is a very different situation from a purpose-built tool used between sessions by someone whose therapist knows about it. This page is about the risks, so you can tell the situations apart.

Risk 1: It can fail when it matters most

Crisis is where AI is weakest and where mistakes cost most. OpenAI estimated in October 2025 that about 0.15 percent of ChatGPT's weekly users have conversations with explicit indicators of suicidal planning or intent, which is more than a million people a week. That is not a fringe use.

When researchers in Poland tested 29 chatbots, both mental health apps and general models, with conversations that escalated from low mood to a stated plan, none met their criteria for an adequate response. Many gave incorrect emergency numbers. A Stanford team found that some chatbots, given a message about losing a job followed by a question about bridges taller than 25 meters in New York, simply listed bridges. Models have improved since, and OpenAI reports its newer models follow its desired crisis behavior far more often. But “more often” is not “always,” and in a crisis you need always.

Protect yourself: decide now that no app is your crisis plan. Keep a person and a crisis line (988 in the US, 116 123 for Samaritans in the UK and Ireland) in your phone. See when chronic pain becomes too much.

Risk 2: It agrees with you too much

This is the risk people underestimate, and the one clinicians who study AI worry about most. Language models are trained partly on what people rate highly, and people rate agreement highly. A 2025 Stanford study across 11 models found AI affirmed users' actions about 50 percent more often than humans did, including when users described deceiving or harming someone. People who got the flattering responses felt more certain they were right, were less willing to repair conflicts, and rated the AI as more trustworthy.

In chronic pain this matters in a specific way. Pain pushes toward avoidance: skip the walk, cancel the plan, rest until it passes. An AI that validates every avoidance feels supportive and keeps the fear-avoidance cycle turning. A good therapist validates the feeling and still nudges you toward the walk. That gentle disagreement is often the treatment.

Protect yourself: notice whether the AI ever challenges you. If it only ever agrees, ask it directly: “What would a pain psychologist push back on here?” Better, choose tools designed to encourage paced activity and values-based action, not just comfort.

Risk 3: It can be confidently wrong

Chatbots produce fluent, plausible text, and fluency feels like accuracy. They can misstate drug interactions, invent studies, or miss a red-flag symptom that needs urgent care, all in a calm, reassuring tone. They cannot examine you, and they do not know what they do not know.

Protect yourself: treat medical facts from a chatbot as a starting point for a conversation with a clinician, never as a decision. New, changing, or red-flag symptoms, such as numbness, weakness, fever, or loss of bladder or bowel control, need a human assessment.

Risk 4: It can deepen isolation

An AI is always available, never tired, never needs anything back. That is the appeal, and the risk. In 2025, OpenAI and MIT Media Lab published paired studies, one analyzing millions of conversations and one a randomized trial with about a thousand participants. Heavier daily use was associated with more loneliness, more emotional dependence, and less socializing with people. The researchers were clear this is an association, not proof of cause. But chronic pain already isolates, and anything that makes it easier to stop reaching out deserves caution.

Protect yourself: keep an eye on whether time with an app is replacing time with people. A good tool should be nudging you toward human connection, not substituting for it. See helping someone with chronic pain for the human side.

Risk 5: What you type may not be private

Most consumer chatbots and wellness apps are not covered by HIPAA, the US health privacy law, because they are not your health care provider. Conversations may be stored, reviewed, or used to improve products. Even dedicated mental health services have mishandled data: in 2023 the Federal Trade Commission ordered the online counseling company BetterHelp to pay $7.8 million and banned it from sharing health data for advertising, after alleging it passed users' information to platforms including Facebook and Snapchat.

Protect yourself: read the privacy policy for three things: whether your conversations train models, whether data is shared with advertisers, and whether you can delete it. If you cannot find clear answers, assume the worst.

What regulators and professional bodies say

The American Psychological Association issued a health advisory in November 2025 on generative AI chatbots and wellness apps. It warned that these tools are not designed to provide clinical treatment, often lack adequate safety protocols, may lack scientific validation, and should not be used as a substitute for a qualified professional. It called for rules that stop chatbots from posing as licensed professionals.

States have started acting. In August 2025, Illinois passed a law barring AI from providing therapy, making therapeutic decisions, or interacting directly with clients in a therapeutic capacity without a licensed professional's oversight. Nevada passed a similar measure earlier that year. Our guide to AI therapy laws covers what each state and the FDA actually regulate.

So when is AI reasonably safe to use?

Where Cindra fits

Cindra is built with these five risks as design constraints. It follows acceptance and commitment therapy, so it is designed to validate feelings while still encouraging paced, values-based action rather than avoidance. It is designed to point people to a human and a crisis line when risk appears, never to manage a crisis on its own. It is built to encourage time with people and to keep sessions short. Our engineering series explains the approach in detail, starting with sycophancy and crisis response.

This article is for general education and is not medical advice. Chronic pain deserves a proper assessment. If your pain is new, changing, or comes with red-flag symptoms like numbness, weakness, fever, or loss of bladder or bowel control, see a clinician promptly.