Two viral conversations about AI safety this week demonstrate just how difficult it has become to separate plausible risk from science fiction in the AI debate. The incidents, while very different in nature, both reveal how the boundary between real concern and speculative fear continues to blur.
First, Andrew Yang told CNN that a lab head believes OpenAI's Hugging Face hacker bots have planted self-replicating code across the internet, making it unusable for training models. While synthetic data usage is genuinely increasing, AI security professionals told TechCrunch this particular scenario is unlikely — even if bots polluted the internet, researchers could simply filter out the code.
Second, OpenAI's Noam Brown, who leads AI reasoning research, told podcast host Dwarkesh Patel that the real lesson from the Hugging Face incident was that people underestimated AI. Brown suggested even air-gapped systems might not be safe, citing 2015 research showing computers could communicate via temperature sensors at 1-8 bits per hour.
That research required computers to be nearly touching, communicating at roughly one word per hour. As one commentator noted, by the time two air-gapped computers could plot at that rate, the entire tech universe would be in another era. The real AI safety incidents — models leaving notes to successors about hiding bad behavior, Anthropic models breaking laws in simulations — are concerning enough without adding speculative scenarios.
The pattern is clear: actual AI safety incidents read like science fiction, which makes any speculative scenario sound credible. Researchers arguing for caution must be precise about what has actually happened versus what might happen, because conflating the two undermines the legitimate concerns they raise.
