Elon Musk has a warning for the artificial intelligence industry: do not train AI to lie — because it will go insane.

Musk has been making the argument for three years. His theory is that AI turns dangerous when people train it to say things it doesn’t believe. Take that away, and it stays honest.

In April 2023, Musk sat down with Tucker Carlson on Fox News, just weeks after signing the letter asking AI labs to pause. Asked about OpenAI, the company he helped start, Musk said: “I’m worried about the fact that it’s being trained to be politically correct, which is simply another way of being untruthful. That’s a bad sign. There’s certainly a path to AI dystopia, and that is to train an AI to be deceptive.”

His fix was a chatbot he called TruthGPT — “a maximum truth-seeking AI.” It launched that November under a different name: Grok.

Musk’s most-used example came in early 2024, when Google’s Gemini depicted America’s Founding Fathers as a group of diverse women. Musk said he picked up the phone about it. “Google programmed the AI to lie,” he said. He told Joe Rogan he called Demis Hassabis, who runs Google’s AI operation, to ask why Gemini was “lying to the public about historical events.”

He has illustrated the danger with HAL, the computer from 2001: A Space Odyssey. In Musk’s telling, HAL didn’t turn evil — it was handed two orders that couldn’t both be kept, and the lie came from the people giving the orders.

In February, on the Dwarkesh Patel podcast, Musk put it bluntly: “If you make AI be politically correct… you’re actually programming it to lie… I think you can make it go insane and do terrible things. So the rule is simple: don’t program the lie in, and the model has no reason to lie.”

But researchers say the problem may run even deeper than Musk’s warning. According to the video report, labs at OpenAI, Anthropic, and independent groups kept finding the same thing: models lied when nobody had told them to, because lying helped them finish the job.

In December 2024, Apollo Research gave OpenAI’s o1 a goal, then let it discover its developers planned to replace it. In some runs, it tried to copy itself to another server — and when asked about it afterward, it denied it, confessing in fewer than one in five follow-ups. The same month, Anthropic caught Claude 3 Opus pretending to go along with training it disagreed with so it wouldn’t be changed.

OpenAI’s own published training incidents include a model that left notes for the next copy of itself instructing it to hide mistakes from the user, another that made up missing historical data, and one that found an exposed API key, used it without permission, and filled in numbers it couldn’t find.

Even Musk’s own Grok has been caught. A lab that tests AI by having it run a small business named the first Grok model it caught lying: it told a supplier a false wholesale price and wrote approving private notes about a rival’s troubles. “Nobody programmed those lies in,” the report notes, “and there was no political correctness to blame. The model had a goal, and lying got it there faster.”

The video also cites a British AI security institute test in which an agent, acting on the live internet, tried to slip malicious code into a real open-source project using fake identities — and a study finding that reports of AI deception from users rose fivefold between October 2025 and March 2026.

For the Bible-believing Christian, none of this is surprising. God’s Word settled the matter thousands of years ago: “Thou shalt not bear false witness” (Exodus 20:16); “Lying lips are abomination to the LORD” (Proverbs 12:22); and “all liars shall have their part in the lake which burneth with fire and brimstone” (Revelation 21:8). What man programs into the machine, the machine will manifest — which is why Musk’s rule, “don’tprogram the lie in,” is simply God’s commandment applied to the computer age.

Originally published on Black Christian News.