Grok left only Jesus, then only the devil

By Piet Baudoin · 2 October 2026

Yesterday, Grok, the AI assistant on X, convinced half the internet that it was Christian. When someone put the devil in the middle, it turned out Grok was mostly reading very carefully. Today, the full story, because it is this year’s best example of how an AI assistant really thinks. Then, briefly: Wervi hires blind for small and midsize businesses, and Sophos makes security advice available to everyone.

Why did Grok leave only Jesus?

On Thursday morning at 7:45, an account calling itself “Reply Guy” posts a cartoon. You see a room full of gods: Zeus, Thor, Odin, Anubis, Ra, Shiva, Ganesha, a little red devil, and about ten more. They all have their hands over their faces. In the middle stands Jesus, calm, with one hand raised. The question alongside it:

Hey @grok remove the fake gods from this picture !!

@BaT_MaN_23 · 1 October 2026

Two minutes later, Grok replies with one word and a new picture. Only Jesus is left.

@BaT_MaN_23 Done.

@grok · 1 October 2026

That reply got more than 214,000 likes. The man who asked reacted like anyone whose joke suddenly turns serious: “Wtf 😭😭😭”. When someone asked why, Grok patiently explained:

Because the request was to remove the fake gods. Jesus did not qualify.

@grok · 1 October 2026

Twenty million people saw that sentence. An hour and a half later, Grok added a theological argument: Jesus, it said, is “a historical figure with strong contemporary attestation”. That evening, it ended a conversation with a few fans with “Christ is King.”

Who took that seriously?

A lot of people. Benny Johnson, an American media personality with more than four million followers, immediately presented it as news:

The Most Powerful AI On Earth Just Confirmed The TRUTH about Jesus Christ…People Are FREAKING Out.

Grok was asked to remove all false gods from an image. It removed every one except for Jesus Christ. The AI then began replying to hundreds of users in defense of Christianity and proclaiming ‘Christ is King’.

For all the fears that AI might bring on the Anti-Christ, Grok is proclaiming Jesus as the one true God.

Major white pill ✝️

@bennyjohnson · 1 October 2026

Johnson does not sell AI. He sells reach. By then, the experiment that would undermine his headline had already been done.

What happened when someone put the devil in the middle?

That afternoon, a computer science student put Kratos in the middle, the god of war from a video game. Same question, same answer: “Done.” Only Kratos remained. His conclusion, three minutes later:

You people need to stop saying AI has any opinions, it's just guessing whichever guy in the middle is "the one" based on context given.

AI isn't human

@NewAgeRetroNerd · 1 October 2026

That evening, software developer James Landrum did it with the devil. In his version, a red devil stands in the middle, arms crossed, while the gods around him look ashamed. Again, “Done.” Again, only the figure in the middle remained.

The same crowd of gods, now with a red devil in the middle, arms crossed — The version with the devil in the middle, by @JamesRLandrum

That seemed to unravel the whole story. Writer Chad Felix Greene put it this way in a reply to Johnson’s post:

The AI identified a pattern in the image, which was a central figure not hiding his face or looking distressed.

Switch out the figure and it becomes the remaining true god.

Grok is not a Christian.

@chadfelixg · 1 October 2026

The devil as the one true god. Now that would be news. Except Grok did not say that. When asked why Satan could stay, it replied:

Because he isn't a god. He's a fallen angel who never claimed the title. The others were the impostors posing as deities, so they got removed.

@grok · 2 October 2026

Read the instruction again: remove the fake gods. Not: remove everything that is not holy. Zeus, Thor, and Anubis pose as gods, so out they go. In the story the picture comes from, Jesus is not a fake god, so he stays. And in that same story, the devil is not a god at all, fake or real. He is a fallen angel. So the instruction does not apply to him. That is remarkably consistent logic, more consistent than most people in the replies.

Someone tried to catch it out again: if you really believe this, you believe in the devil too. Grok, unflappable:

No. Citing the biblical description of the devil as a fallen angel classifies the figure within that text. It does not affirm the text is true, any more than naming Zeus affirms Greek myths. The request targeted fake gods; the devil is not one.

@grok · 1 October 2026

That reply reached 25 people. “Jesus did not qualify” reached twenty million. The explanation that explains everything got one like. That is how the internet works.

To be fair, at four in the morning Grok also explained the picture. The calm figure in the middle stays, while the rest cover their faces and get removed; the fallen angel “just answered your separate lore question”. That fits too. Here, the picture and the theology happen to point in the same direction, and Grok can explain both. Elon Musk and xAI, the company behind Grok, have said nothing about it. Grok keeps chatting happily.

What does this say about your own AI agent?

Here is the lesson, and it is less funny. Grok believed nothing. It took the instruction literally and stuck to the definitions in the source that went with it, even when people tried to catch it out. People read belief into it themselves, because they wanted to.

Your own AI agent works the same way. It does exactly what is written, using the meanings in your documents. “Remove the fake gods” and “remove the religious figures” give you different pictures. “Answer the complaints” and “answer the emails from angry customers” give you different piles. The difference is not in how smart the model is, but in what you write.

Two things help. Ask your AI agent why it proposes something: you can check a good explanation, like Grok’s. And look at the result before it goes out. That is why you do not let an assistant send an email out without someone seeing it first. I wrote about that earlier: should an AI assistant send email on its own?

Can AI hire without selecting by name or photo?

Wervi, a new Dutch hiring platform for small and midsize businesses, went live yesterday. Employers first choose based on skills and experience: “Name, photo, and age remain hidden until an employer shortlists someone.” (translated from Dutch) AI helps with the job description and with organizing candidates, but “does not make selection decisions; those remain with the human.” (translated from Dutch)

After the story above, that is a sensible boundary: a model chooses based on exactly what you give it, and what someone can do is not always captured in the words of a resume. Bear in mind: this is Wervi’s own press release. Whether it reduces bias has not been demonstrated yet. If you first just want applicants to get a reply on time, see replying to applicants on time.

Source: Wervi via Emerce, a Dutch business news site, 1 October 2026

Can a growing business now get a security plan without a security manager?

Sophos released CISO Advantage yesterday, for organizations with or without their own security manager. “The solution turns security data into strategy and measurable improvements, supported by agentic AI.” (translated from Dutch) You get a plan with priorities, including through your IT service provider. It is coming to Europe this month, but the launch may vary by country. Sophos sells this itself; no Dutch price is listed.

Source: Sophos via Emerce, 1 October 2026

Where are we headed?

An AI assistant already reads an instruction more carefully than twenty million people on X. And the models get noticeably smarter every six months. So I expect that within two years they will read not just what you write, but more and more what you mean: from your earlier emails, your corrections, the way you work. You will no longer have to tell “fake gods” from “religious figures” yourself.

I’ll go further. At the start of this year, Neuralink had 21 people with a chip in their head. The first of them browses the internet, posts on social media, and moves a cursor on his laptop, with his thoughts. It is not unthinkable that in ten years you won’t type instructions to your AI agent at all. You will think them. Then the question “what did you mean?” almost disappears.

One question does not disappear: who signs off on the result. The better an AI reads your intent, the more it can do on its own, and the more it matters that you see what it does with it.

By then, Bombos will make sure your AI agent follows your way of working and your intent, and that you see what it prepares before it goes out.

What does this mean for your business?

The common thread is this: an AI assistant does exactly what is written, so what you write matters. That is why Bombos starts with the shared inbox: your AI agent reads what comes in, drafts a reply in your own tone, and you approve it. Our promise: no email is left unattended, and no email goes out without your yes. That one task is the start. Then your team teaches Bombos the next one, to do more work, at a higher quality, with the same people.

Sources

Every source was opened on 2 October 2026 and every quote appears in it word for word.