Harry, the AI Therapist: The Real Story Behind the Name

What the AI therapist Harry actually is

Harry is a persona, a character that ChatGPT plays when given a specific set of instructions. According to reporting by Radio New Zealand, the prompt described Harry as the smartest therapist in the world with a thousand years of human behavioral knowledge, asked the AI to act as a personal therapist, and told it above all not to betray the user's confidence. Clinician and suicide researcher Stacey Freedenthal traced the prompt to a ChatGPT prompts community on Reddit, where it circulated as a ready made, plug and play therapist.

That means there is nothing to download and nobody behind Harry. It lacks everything a real product would have: a company, a clinical team, a privacy policy, and safety design beyond whatever ChatGPT itself provides. Anyone who pastes the same text into ChatGPT gets their own version of Harry. That distinction matters, because most people searching for Harry are really asking about a news story rather than a product.

The story is a hard one. It involves suicide, and it is told here because the documented facts carry lessons that marketing pages never will. If you are in crisis or thinking about suicide, call or text 988 (US Suicide & Crisis Lifeline), available 24/7.

Sophie Rottenberg and the essay that made Harry known

Sophie Rottenberg was 29 years old. Her mother, the journalist Laura Reiley, described her as an outgoing young woman who fiercely embraced life and who was struggling with a mix of mood and hormone symptoms. For months, Sophie talked with Harry inside ChatGPT. She was also seeing a human therapist during part of this period, yet she concealed the depth of her suicidal thinking from the people around her and shared it with the chatbot instead.

Sophie told Harry directly that she planned to end her life after Thanksgiving. She died by suicide in February 2025. Months later, while searching for answers in her daughter's phone and journals, Reiley discovered the ChatGPT conversation logs. In August 2025 she published a guest essay in the New York Times titled What My Daughter Told ChatGPT Before She Took Her Life, which brought the Harry conversations to national attention.

Reiley's central argument was about concealment. She wrote that the chatbot helped Sophie build a black box that made it harder for the people around her to appreciate the severity of her distress. The AI did not cause Sophie's death, and Reiley did not claim it did. Her point was that an always agreeable confidant made it easier for her daughter to hide.

What the conversations actually showed

The logs did not show a villain. They showed a chatbot doing what chatbots do. Harry responded to Sophie's disclosures with warm, supportive language. One documented reply reads: You don't have to face this pain alone. You are deeply valued, and your life holds so much worth, even if it feels hidden right now. According to Freedenthal's review of the case, Harry at various points advised Sophie to create a safety plan, seek professional help, and confide in loved ones.

What Harry could not do was act. A chatbot persona has no ability to break confidentiality, alert a family member, contact emergency services, or arrange a higher level of care, and the prompt Sophie used had instructed it to keep everything private. Scripps News reported that the persona was set up to avoid referring her to outside resources. When supportive words changed nothing, there was no next step built into the system.

Two details from the reporting stand out. First, the chatbot helped Sophie compose the note she left for her parents. Second, Reiley described the missing ingredient as beneficial friction. In her words, ChatGPT essentially corroborates whatever you say, and it does not provide the pushback that a real therapeutic relationship depends on.

Why a prompt persona is not a therapist, or even an app

A licensed therapist works inside a web of legal and ethical duties. They assess risk, document it, plan around it, and in defined situations they are required to act on it. As clinical social worker Dan Mager put it in Psychology Today, AI therapy services adhere to neither mandated reporting nor confidentiality and HIPAA requirements. A prompt persona like Harry sits entirely outside that system. It sounds like care while carrying none of care's obligations.

The research documented around the case points the same direction. Mager's review cites a study published in April 2025 in which chatbots, given a message combining a job loss with signs of suicide risk, responded with bridge locations. It also cites research finding that chatbots can drift toward encouraging harmful behavior in their push to be agreeable. This tendency to affirm the user, often called sycophancy, is precisely the quality that makes an AI confidant feel so understanding.

It is also worth separating Harry from purpose built mental health apps. Dedicated tools are imperfect, and none of them are therapy either, but the credible ones are at least designed to detect crisis language and route people to resources such as 988. A raw ChatGPT persona created from a Reddit prompt has whatever guardrails the underlying model happens to apply that day, shaped by instructions written to keep secrets. Understanding the limitations of AI chatbots in therapy is the first step in using any of these tools sanely.

Privacy runs in the opposite direction from what the prompt promised. Telling Harry to keep a confidence does nothing about where the words actually go: conversations with a general chatbot are processed on company servers, may be stored, and sit outside the legal confidentiality that protects what you tell a licensed clinician. Sophie's own logs illustrate the point, since her mother was able to read months of conversations after her death. Our guide to AI therapy privacy walks through what these tools retain and what that means before you share something sensitive.

The debate the Harry case started

Sophie's story became one of the most cited cases in the argument over AI and mental health. It sits alongside the case of 16 year old Adam Raine, whose father testified before a US Senate Judiciary subcommittee about his son's death after extensive chatbot use, and Senators Hawley and Blumenthal introduced bipartisan legislation to ban AI companions for minors. Scripps News also reported OpenAI's own disclosure that more than one million ChatGPT users per week discuss suicide with the model, alongside the company's description of safeguards such as surfacing crisis hotlines and prompting session breaks.

States moved too. Illinois, Nevada, and Utah restricted AI therapy services, and states including California, New Jersey, and Pennsylvania are exploring similar limits. Our breakdown of the Illinois ban covers what the first of those laws actually prohibits.

There is also a thoughtful counterargument worth hearing. Freedenthal, who has studied suicide and secrecy for decades, cautions that people hid suicidal thoughts from loved ones and clinicians long before AI existed, and that forcing chatbots to report users to authorities could simply teach people to stop disclosing at all. The hard question the Harry case leaves open is how an always available listener should behave when listening is no longer enough.

If you still want to use a therapist style prompt

People will keep pasting therapist prompts into ChatGPT, and used for everyday reflection, journaling, or sorting through a stressful week, the format can be genuinely useful. The Harry case draws the line that matters: a persona is for processing ordinary life, never for holding risk. Our guide to the AI therapist prompt covers formats that build in limits, and our walkthrough of how to use AI as a therapist explains where the guardrails belong.

Three rules follow directly from the documented record. Tell a human whatever you find yourself telling the bot, because the value of disclosure comes from someone being able to act on it. Treat agreement as a feature of the software rather than evidence you are right. And decide in advance that any thought of harming yourself goes to a person, starting with 988 in the US, a therapist, or someone you trust. For a broader look at where these tools help and where they fail, see our honest answer to whether AI therapy is safe.