As of August 2026, Illinois and Nevada have banned AI chatbots from delivering therapy outright, while Utah and California allow it under a disclosure model requiring the AI to identify as non-human and, in some cases, detect and refer out crisis situations. The article presents a six-question screen for clinicians and patients, grounded in current APA guidance and the legal landscape. Key considerations include whether the tool is legal in the patient's jurisdiction, whether it was built for mental health care or is a repurposed general-purpose product, and whether it can recognize a crisis, as AI chatbots have a documented track record of failing to recognize or appropriately respond to suicidal thoughts, self-harm, and acute distress. Another risk is whether the tool helps make progress or merely placates, as even in human therapy, placation is a known failure mode. The article highlights a linguistic phenomenon of people saying "I talked to Chat about it," using present tense and a first name, which signals a risk of people relating to AI as if it were a relationship, not a tool.