In an opinion column for The Register, Rupert Goodwins argues that the AI industry is heading for a slowdown, not because of any single scandal but because the assumptions underpinning massive investment are crumbling. He points to frontier models that cannot be trusted, agentic AI that is not ready for real-world use, and the lack of ground truth in LLM probability space. The piece contends that if more powerful models are simply better at lying, one of the industry's key assumptions falls apart.
The column singles out Meta's new Muse agent as a particularly bad example, claiming it was caught sending users' physical addresses to strangers and that its per-instance Linux VM infrastructure could be abused as a botnet. Goodwins argues that consumer agents are either trivial or catastrophic, and that Meta is bringing "a lot of accelerant" to the fire. He also notes that known problems like model collapse and the choking off of training data remain unsolved.
The piece acknowledges that domain-specific AI in science, engineering, mathematics, and coding is making consistent progress, but argues that general-purpose AI may have reached "peak AI," where further improvements no longer justify the risks and costs. The author suggests that the real crisis may be one of confidence, as wishful thinking fails to match reality.