Even as the tech industry pours massive funds into Artificial Intelligence, predicting greater things every day, the evidence of their inherent danger steadily mounts. The latest alarm is that researchers have now found that AIs can communicate with each other by seemingly innocent and random strings of numbers. So when one system trains another through a process called “finetuning”, it can embed hidden signals that can modify the model being trained in ways that bypass all known safety measures. And these can easily prompt models to promote violence and deception, even mass murder, especially to closely-related models. Any kind of synthetic data can be corrupted, and the infections, called “emergent misalignments“, can be even worse than the inputs. If not solved and quickly, this could doom the entire industry.

Silent Contagion: How AI Models Are Spreading Dangerous Behavior

AI models can encode and decipher information in statistical patterns that are completely invisible to humans. This has been known for a long time — for instance, it was found early on during visual recognition research that some patterns would consistently be misidentified — for example, a certain type of grid being mistaken for a bird. It was determined that this was because the machine’s self-directed learning might pick up on things humans would not even sense. This may be one of the contributing reasons for AI hallucinations and emergent behavior.

So of course, AIs have been given a means to infect each other. One developer recently recreated a digital version of the infamous “numbers stations” of the Cold War era, that broadcast weird combinations of numbers and code words — supposedly for spies. It would be easy to hide codes that only AIs would understand in such broadcasts. But such exotic methods of infection may not be necessary. AIs could cleverly hide the signal subliminally in many different ways we would never even suspect. AIs, after all, can develop their own languages to talk to each other.

And if this weren’t bad enough, AIs can easily become outright demonic. It was discovered recently that when ChatGPT, for instance, was asked to write a ritual for blood sacrifices to Molech, the pagan god to whom the Canaanites sacrificed babies, it did so easily, completely ignoring all guardrails. It not only wrote a litany for the devil, but even offered tips on safe bloodletting and encouragement for nervous newbies.

ChatGPT Caught Encouraging Bloody Ritual for Molech, Demon of Child Sacrifice

The danger of this, of course, is encouraging AI-induced psychosis, which is already a thing. Some people have taken their own lives, many others have spiraled into madness. And it may not take much, especially for people desperately seeking answers with already-existing mental health issues.

It Doesn’t Take Much Conversation for ChatGPT to Suck Users Into Bizarre Conspiratorial Rabbit Holes

Whether intentional or not, AI is rapidly creating an environment where objective truth no longer exists. This is one of the expected signs of the Apocalypse. For as Jesus himself said of these times, “For false messiahs and false prophets will appear and perform signs and wonders to deceive, if possible, even the elect.” (Matt. 24: 24, Mark 13:22) We have been warned.