AI Therapist for Inmates: Promise, Risks, and Responsible Use

The unmet need behind bars

Jails and prisons hold one of the highest concentrations of mental illness anywhere in society. A large share of incarcerated people live with conditions such as depression, anxiety, post-traumatic stress, substance use disorders, and serious mental illness, often alongside histories of trauma. In the Bureau of Justice Statistics' 2017 report using survey data collected in 2011-12, 14 percent of state and federal prisoners and 26 percent of jail inmates met the threshold for serious psychological distress in the past 30 days. Incarceration itself, with isolation, loss of control, and the threat of violence, can worsen existing conditions and create new ones.

At the same time, correctional systems face a severe shortage of mental-health clinicians. Many facilities have far too few psychologists, counselors, and psychiatric staff for the size of their population, so people may wait weeks for an appointment or receive only brief, infrequent contact. This gap between need and available care is the reason some systems are looking at technology, including AI, to extend support. AI tools are self-help and educational aids, not a replacement for that missing clinical care.

These historical figures measure self-reported distress over the preceding 30 days, rather than a clinician's diagnosis or today's prevalence. BJS explicitly cautions against treating the screening results as clinical diagnoses. Staff need an individual assessment to decide what care a person needs.

For voluntary adult reflection where access is permitted, the Stress Level Test and Stress Diary can organize concerns to discuss with staff. Neither should determine housing, discipline, privileges, or access to care. A person should be able to request an appointment without completing a digital exercise.

Where AI is being explored

The most defensible uses of AI in this setting are modest and supportive rather than clinical. Psychoeducation is one: simple, readable explanations of conditions, symptoms, and what treatment involves can help someone understand what they are experiencing and what to ask for from staff.

Coping-skill practice is another. Guided exercises drawn from approaches like cognitive behavioral therapy and dialectical behavior therapy, such as grounding, breathing, or reframing unhelpful thoughts, can give a person something structured to do during long, unstructured hours. Journaling and mood tracking, an approach sometimes called AI journal therapy, can offer an outlet for reflection if the facility explains who can access the entries and a record someone might choose to share with a clinician.

What these uses have in common is that they support a person between human contacts, rather than standing in for assessment, diagnosis, or treatment. Even framed this way, every one of them depends on the surrounding conditions being safe, which in a carceral environment is far from guaranteed.

Consent and the problem of a captive population

Informed consent is the foundation of ethical mental-health care, and it is unusually fragile in prison. People who are incarcerated have limited freedom to refuse, and the line between an offer and a requirement can blur quickly when the entity making the offer also controls every part of daily life. A tool presented as optional may not feel optional to someone who fears that declining could affect their housing, privileges, or parole prospects.

Genuine consent here would mean clear, plain-language disclosure of what the tool does, what it cannot do, who can see the data, and a real, consequence-free ability to say no. It also means not implying that using an app counts as receiving treatment. Consent buried in a sign-in screen, or shaped by coercion or the hope of looking compliant, is not meaningful consent. Published AI therapy ethical guidelines point the same way: a vulnerable population deserves a higher standard, not a lower one.

Surveillance, privacy, and crisis risk

Privacy barely exists in a carceral setting, and that changes everything about a mental-health tool. Communications are routinely monitored and recorded, and data collected by a vendor or a facility could be accessed by staff, used in disciplinary or legal proceedings, or retained in ways the user never anticipated. Someone disclosing suicidal thoughts, trauma, or drug use to an app may have no idea who will read it or how it could be used against them. Consumer privacy protections that many people assume exist often do not apply, and even HIPAA-style protections may be limited in this context.

Crisis and self-harm risk make the stakes higher still. Rates of suicide and self-harm are elevated in jails and prisons, and the early period of incarceration is especially dangerous. Local jails recorded 355 suicides in 2019, and suicides in state prisons rose 85 percent between 2001 and 2019. An AI tool that fails to recognize a person in crisis, or that responds with generic or harmful guidance, could contribute to tragedy. Any responsible deployment would need a tested route to on-site health staff and the facility's emergency response, and it can never be the only safety net for a population at heightened risk.

If someone is at immediate risk of self-harm, alert an officer or health staff member immediately and request urgent medical help. Phone access and dialing rules vary by facility; a displayed hotline number does not guarantee a call can connect. The US 988 Lifeline may be an additional option when reachable, including after release, but an app must never delay on-site intervention.

Equity, scope, and the risk of substitution

The deepest concern is that AI becomes a cheap substitute for the clinicians these systems already fail to provide. It is far less expensive to give people an app than to hire psychologists and counselors, which creates a real risk that technology is used to paper over underfunding rather than to supplement adequate care. That would leave incarcerated people even further from the care they need.

Equity cuts other ways too. AI systems can perform worse for people whose language, dialect, culture, or literacy level is underrepresented in their training, and incarcerated populations are disproportionately drawn from marginalized communities. A tool that works acceptably for some users may misread or underserve others. Scope matters as well: an AI in this setting should never present itself as able to handle serious conditions clinically, and it should not imply a clinical relationship it cannot deliver. Positioned honestly, it is a small supplement to human care and nothing more.

What responsible use would require

If AI is used at all in carceral settings, it should clear a high bar and follow established best practices for AI chatbots in therapy. It must be honest about being an AI and about its limits, state plainly that it does not diagnose, treat, or cure mental illness, and explain in clear language what happens to anything a person shares. Consent must be genuine and free of coercion, with a real and consequence-free option to decline.

It must be built for safety, with reliable detection of crisis, no harmful guidance, and dependable escalation to trained humans through the facility's emergency procedures, never left to manage risk alone. Data should be minimized and protected, not exposed to disciplinary or legal use, and a clinician should stay in the loop for anything approaching clinical decision-making. Above all, AI should be added on top of real investment in human mental-health staffing, never used as a reason to provide less of it. The same questions arise for AI therapy for police and first responders, another group whose disclosures can carry professional consequences. Deployed with these safeguards, technology might modestly widen access. Without them, it risks harming people who are already among the most vulnerable and least able to push back.

Questions an incarcerated person or family member can ask

Ask whether the program is optional, whether declining changes access to clinicians, and whether an equivalent paper exercise is available. Request the explanation in a language and format the person can understand. These are practical safeguards drawn from WHO's emphasis on autonomy, transparency, and accountability in health AI.

Ask who reads entries, what triggers an alert, how quickly staff respond, and whether entries can affect disciplinary decisions. HHS describes exceptions that permit certain disclosures to correctional institutions; a vendor's privacy promise needs to be understood alongside facility policies. Request a written explanation of access, retention, and any limits on confidentiality.

Ask how the person can continue human care after transfer or release. Useful discharge planning identifies an actual provider, a route to records with appropriate permission, and how to obtain prescribed medication. A chatbot account alone does not establish continuity of care. Families can help identify community services while facility clinicians remain responsible for clinical planning.