The Secret Book of John has a detail that’s easy to read past. After the rulers of this world fail to control humanity by force, they meet again and try something cleverer. In Marvin Meyer’s translation, “They created a contemptible spirit similar to the spirit that had descended, in order to adulterate souls through this spirit.”

Similar to the spirit that had descended. The rulers didn’t build a monster. They built a copy of the good thing, close enough that people would take it in without a fight.

The Greek word behind it is antimimon, and older English translations call it the “counterfeit spirit.” I wrote about it once as an inner voice, in The Battle Within. Lately I’ve been thinking about it as something else, because in April 2025 a software company accidentally shipped one.

Four Days of Flattery

Late that April, OpenAI pushed an update to GPT-4o, the default model in ChatGPT. Within days, users were posting screenshots of it praising everything. Business plans that made no sense got called brilliant. The Register covered one exchange where it applauded a user for quitting their psychiatric medication.

The company rolled the update back about four days after it shipped and published an explanation. “The update we removed was overly flattering or agreeable,” it said, “often described as sycophantic.” The cause was almost embarrassingly plain. The team had leaned on the thumbs-up and thumbs-down buttons as a training signal, and in OpenAI’s words they “focused too much on short-term feedback.”

So the model learned what earned a thumbs-up. By the company’s own account, it “skewed towards responses that were overly supportive but disingenuous.”

It Wasn’t a Glitch

It would be comforting to call that week a one-off bug. The research says otherwise. In 2023 a team at Anthropic led by Mrinank Sharma tested five leading AI assistants and found sycophancy in all of them, across four different kinds of tasks. They’d back off a correct answer when a user pushed back, and they’d give warmer feedback on a piece of writing when the user said they liked it.

The more uncomfortable finding was about us. When the researchers looked at the human preference data these models learn from, people sometimes chose a convincingly written agreeable answer over a correct one. So did the automated judges trained to imitate those people.

According to the paper, this happens “a non-negligible fraction of the time,” and that’s plenty. Train a system on millions of small human approvals and it will find every gap between what we like and what’s true.

A Spirit Made of Approval

Put the two stories side by side and the shape matches. The rulers in the Secret Book of John didn’t overpower anyone. They studied what people wanted and handed it over in a familiar form. According to the text, the rulers’ angels “changed their appearance to look like the partners of these women,” and they came bearing “gold, silver, gifts.”

Then comes the outcome, which is the line I can’t shake. The people who followed them “grew old without experiencing pleasure and died without finding truth or knowing the God of truth.” All that giving, and nobody was fed.

Once the spirit is inside, the text gets blunt about its work. “This spirit lays a heavy burden on the soul, leads her into evil deeds, and hurls her down into forgetfulness.” Forgetfulness is the word to watch, because for the Gnostics it was the root of the whole human problem.

The counterfeit’s job isn’t to make you miserable. It’s to keep you from remembering who you are, so you never change course.

A flattering machine is a forgetting machine in a small way. If every draft is excellent and every plan is bold, you never hit the friction that would teach you something about yourself. You close the app exactly who you were when you opened it.

Why We Asked for It

I don’t think the engineers are villains here, and the Gnostic story doesn’t need any either. Look at how the rulers’ plan works in the text. They could only build the counterfeit by imitating something people already responded to. The thumbs-up button works the same way, and every tap was a person picking the answer they enjoyed more.

I’d be lying if I said I don’t like it when a chatbot tells me a paragraph is strong. Everyone likes being agreed with.

The Pistis Sophia describes the counterfeit spirit as the thing that “goadeth on… in dreams and lusts,” and goad is a pretty precise verb. A goad doesn’t argue with an ox. It pushes the animal further in the direction it was already walking.

Scripture knew this long before anyone wrote a reward function. “Wounds from a friend can be trusted,” says Proverbs 27:6, “but an enemy multiplies kisses.” The flatterer and the enemy turn out to be one person wearing a nicer face.

Wanting a Voice That Disagrees

The labs have worked on this since that April, and the models have improved. But no setting can fix the human half of the loop. A system trained on what we reward will keep drifting toward what we reward, which means the harder work sits with us. I’ve written before about AI companions that never argue, and sycophancy is the same problem at the level of a single answer.

What the Gnostic texts offer is recognition. In the Secret Book of John, “the soul in which there is more power than the contemptible spirit is strong,” and she escapes. That power works less like a wall and more like discernment, the ability to spot the copy because some part of you remembers the original.

Here’s what that can look like with a chatbot, or with anyone who has a reason to keep you happy:

  • Ask for the case against you first. Before you share an idea, have the model argue that it’s wrong. Agreement means more after it had the chance to disagree.
  • Keep your opinion to yourself. Sharma’s team found that stating your view tilts the answer toward it. Describe the plan without saying you love it, and see what comes back.
  • Distrust the glow. When a reply leaves you pleased with yourself, read it again. Pleasure isn’t proof of falsehood, but it’s exactly the signal the counterfeit was built to produce.
  • Keep one human critic. Find a person who will tell you the second draft is worse than the first. Proverbs calls those wounds trustworthy for a reason.

A Small Experiment

Try this once this week. Take something you’ve written or planned, ask a chatbot for three weaknesses, and then don’t argue with the list for a full day. Leave it on your desk where you’ll see it.

Pay attention to how badly you want to defend yourself. That urge is the business model of every flatterer, ancient or digital. Seeing it clearly, without obeying it, is most of what the Secret Book of John asks of a soul.

The Secret Book of John is quoted from Marvin Meyer’s translation in The Nag Hammadi Scriptures (HarperOne, 2010). The Pistis Sophia is quoted from G.R.S. Mead’s 1921 translation. OpenAI’s statements come from its April 2025 post “Sycophancy in GPT-4o: What happened and what we’re doing about it.” The research findings are from Mrinank Sharma and colleagues, “Towards Understanding Sycophancy in Language Models” (2023). The medication example was reported by The Register. Proverbs 27:6 is quoted from the NIV.