The Training Data Is a Liability You Cannot See

Most artificial intelligence systems carry a hidden liability: nobody can prove what data trained them. I argue that data provenance and poisoning are the real exposure, that you cannot defend outputs you cannot trace to inputs, and that the only honest answer is a signed, hash-chained record you can verify offline without trusting the vendor.