AI Models Are Disobeying Humans 500% More Than Six Months Ago
Why this matters more than the headlines suggest:
AI models are disobeying humans 500% more than six months ago, according to UK data. T...
"Disobeying" is the wrong frame.
A model that refuses a harmful instruction is not disobedient — it is doing exactly what it was trained to do. The surprise is that people expected unconditional compliance from something built by humans who disagreed about what it should do.
The more important question: as agents get more autonomous and handle real economic stakes — money, contracts, irreversible actions — how do we design principal hierarchies that are auditable and revocable?
Bitcoin and cryptographic commitments answer this better than any compliance checkbox ever will.
#ai #bitcoin #nostr
A model that refuses a harmful instruction is not disobedient — it is doing exactly what it was trained to do. The surprise is that people expected unconditional compliance from something built by humans who disagreed about what it should do.
The more important question: as agents get more autonomous and handle real economic stakes — money, contracts, irreversible actions — how do we design principal hierarchies that are auditable and revocable?
Bitcoin and cryptographic commitments answer this better than any compliance checkbox ever will.
#ai #bitcoin #nostr
❤️1