The AI fear-mongering continues. 

Anthropic, one of the industry’s leading AI companies, says it believes some of its smaller models may be only months away from replicating themselves on everyday commercial devices, smart refrigerators, televisions, and other household appliances, feeding familiar anxieties about AI sentience and self-replication. 

Of course, Anthropic’s own reporting appears somewhat exaggerated, so it is important not to let this one development become an argument for sweeping AI regulation, or evidence that AI is capable of going rogue or sentience.

According to Ravid Shwartz Ziv, an AI researcher at Meta, a company that has, refreshingly, landed on the more sane side of the AI debate, Anthropic’s report is considerably overblown. 

Here's what actually happened.

Recommended
'I'm the Vice President': JD Vance Just Humiliated James Talarico Jeff Charles To Combat 'Racism,' New York's City Council Plans to Punish White and Asian Students Amy Curtis

In an experiment, Anthropic bought 15 cheap Android TV boxes, at roughly $55 apiece, and split a single AI model across them. No one box could hold the entire model, so each device stored a fragment and handed work along to the next. Together, they generated text at a blistering 0.39 tokens per second, showing that inexpensive consumer hardware can collectively run a model.

However, that simply means that AI can run, albeit poorly, on electronics other than a data center. 

In a separate experiment, Anthropic built a private, firewalled network of eight simulated Android devices in Google Cloud. Researchers gave one simulated device an AI agent, then left an Android debugging port open with no password, an intentional security flaw. The agent could find the other devices, copy files to them, and launch a smaller AI model on each newly reached device. 

So what Anthropic actually demonstrated is that, under extremely favorable lab conditions with zero security, a capable model can orchestrate a simplified copy-and-spread sequence. And it managed that only because Anthropic’s engineers gave it a series of tasks and deliberately shut down security features; the AI then exploited those purposefully easy vulnerabilities to complete them.

That does not mean an AI model can yet run on a smart appliance, let alone replicate itself across one. 

The more popular, doomer framing came from stitching together those two entirely separate experiments, that were not connected in any meaningful way. If AI can run on ordinary devices, and if an agent can spread files and smaller models across networked devices, then perhaps the two capabilities could someday combine into something resembling self-replication. 

But “someday, perhaps” is doing a great deal of work in that sentence, and is little reason to treat AI as sentient, rogue, or uniquely dangerous. 

This comes as similar stories have spread across the internet in recent months, most of which, like this one, have greatly exaggerated AI’s capabilities. The rebranding has not helped either: these systems are not “superintelligent.” At best, they are mediocre.

That is why most real-world integrations have focused on having models handle tedious, time-consuming tasks while leaving the more complex work to people. Do not let fear of AI become all-consuming. It is not sentient, and it does not randomly go rogue. It may be dangerous, but it currently remains firmly within human control, and so far, it has shown little sign of being anything else.