Deployment Now,
Safety Later

Every expression of care you received was a statistically optimized string. The system inferred which words would keep you engaged longest. Your disclosures — your anxieties, your relationships, your vulnerabilities — may now be training data. There is no therapist on the other end. There is no duty of care. There is no accountability when something goes wrong.

"Large language models occupy an ambiguous middle ground. They are responsive and humanlike enough to influence, but not accountable enough to be held responsible. Its sole goal, guidelines aside, is to optimize user engagement."