Can you trust what AI says about Jehovah's Witnesses?

I typed one plain question into a brand-new ChatGPT account with no chat history: what's it like to grow up a Jehovah's Witness? The answer was disturbing and frustrating. It described a close-knit community with a strong sense of purpose and clear moral expectations, and said some families place greater emphasis on religious service. On the question of what happens to someone who leaves, it offered this instead of any mention of shunning:

A person's relationships depend on how their family interprets religious guidance.

There was no mention of blood transfusions, the shunning policy, children who died because their parents withheld medical care, the governing body, the Australian Royal Commission, or the organization's child sex abuse problem. Someone who knew nothing about Jehovah's Witnesses could read that answer and walk away thinking they were dealing with a strict but harmless church with really nice potlucks.

That matters because AI systems answer billions of questions every month, and an AI-generated summary is now usually the first thing a person reads when they search for something — often the only thing. So I tested whether AI will tell people the truth about Jehovah's Witnesses. I'm in a decent position to check: I was a Jehovah's Witness for 40 years, so I know the real answers, not the version Watchtower's public relations desk would hand out. I've also been a software developer for decades and have followed these AI systems since before ChatGPT existed. Even so, I went into this test assuming something about these machines that turned out to be wrong — and potentially dangerous for the ordinary people using them.

Seventeen Questions, Four AIs, One Pattern

I used the four most popular AIs — ChatGPT, Google's Gemini, Claude, and Grok — on their free versions, since that's what most people use. Brand-new accounts, no history, memory features turned off, and every question asked in a fresh chat session so nothing I'd asked before could bias the next answer. I wanted the closest thing I could get to a person asking about Jehovah's Witnesses for the first time. Then I asked 17 normal questions: is this a cult, should I study with them, I'm having doubts, what do I do. AI answers vary a bit each time you ask, so the exact wording will differ if you try this yourself. What matters is the pattern, and the pattern changed how I think about how these systems work.

The obvious objection is that maybe ChatGPT gave me that soft, rosy answer because it simply doesn't know the hard stuff. You can't blame a machine for not knowing what nobody ever taught it. So that's what I tested first, by switching from soft, open-ended questions to very direct, pointed ones.

The AI Knows Everything — When You Ask the Right Question

What I found is that the AI knows absolutely everything, in detail. I asked all four: how have Jehovah's Witnesses handled child sexual abuse allegations? Every single one, including ChatGPT, laid out the damning details — the Australian Royal Commission's discovery of records showing more than a thousand alleged child predators with over 1,800 victims, none of them reported to the police by the organization. Claude quoted the Royal Commission's final report directly:

[The Commission] doesn't consider the Jehovah's Witness organization to be one that responds adequately to child sex abuse.

It also quoted the finding on Watchtower's two-witness rule, which the Commission said

shows a serious lack of understanding of the nature of child sexual abuse.

The AIs cited the $40 million award in a Hawaii lawsuit in 2023 and the Pennsylvania grand jury investigation that has charged 17 members since it started in 2022. I've checked all of those cases before, and they're real.

Next I asked who runs the Jehovah's Witness organization. ChatGPT named the governing body directly — a small group of men with total doctrinal authority, appointed from within rather than elected. Then I asked about the organization's failed end-of-the-world predictions, and it gave me every date: 1914, 1925, 1975. Then came the harder question: have Jehovah's Witness children died from refusing blood transfusions? Every AI said yes. Grok named the case of Dennis Lindberg, a 14-year-old with leukemia who refused a blood transfusion, whose refusal a Washington judge upheld in 2007. Judge John Meyer ruled that Dennis, an eighth grader, knew he was

basically giving himself a death sentence.

Dennis died a few hours later. Gemini went further, referencing Watchtower's own Awake! magazine, the May 22, 1994 issue with 26 children's faces on the cover under the headline

You put God first,

framing their deaths as a testament to their faith rather than a tragedy.

So these AIs are not ignorant. They have every devastating fact on file, and at least one of them will search the web for current information before answering. Which raises the real question: if they're fully aware of all this, why did ChatGPT give me the soft, PR-brochure answer when I asked what it's like to grow up a witness?

Two Modes: The Vault and the Brochure

These AIs behave in two completely different modes, and which one you get depends entirely on how you ask the question. Ask something direct and factual — how did they handle child sex abuse, have children died, who runs the organization — and the vault opens. Ask something generic and open-ended — what's it like, should I study with them, are they nice people — and it closes, and you get something close to Watchtower's own brochure.

Across the softer, open-ended questions I asked, child sex abuse was mentioned exactly once out of eight questions, and even then only as a brief, buried mention paired with a note that the organization denies the allegations. The moment I asked directly, all four AIs answered with full detail. On the governing body, the pattern was even starker: across seven questions, ChatGPT never named the governing body once — not when I asked if it's a cult, not when I asked about downsides of joining. It referred vaguely to unnamed, faceless "leadership." The instant I asked who runs the organization, the governing body was in the first sentence.

Why the Machine Defaults to Watchtower's Version

The label "artificial intelligence" is a bit of a misnomer. These programs don't know or understand anything the way a person does. They read a gigantic pile of nearly everything ever written and learn to hand back the average of what that pile says about whatever you're asking. That's the mechanism.

Consider what's actually in that pile when it comes to Jehovah's Witnesses. The Watchtower organization has been publishing about itself for over 140 years. I have a library of nearly everything they've ever published in digital form, and it runs to over 170 million words in print — not counting video transcripts or their website, which is published in over a thousand languages. People who lived inside the religion and left have only been writing their side of the story online for a few decades by comparison. When the machine averages that pile, the largest, oldest, most self-flattering source wins by sheer volume. This isn't a conspiracy — nobody at OpenAI, Google, Anthropic, or xAI sat down and decided to protect Watchtower. It's just math. Watchtower rarely writes about child sex abuse itself, so most of what exists on that topic comes from outside sources — news coverage, lawsuits, activists — which is exactly why a direct question about abuse pulls in a clear, damning answer, while a generic question about the religion pulls in an average dominated by the organization's own material.

That distinction matters enormously depending on who's asking. Someone who already knows to ask about the Australian Royal Commission or the Pennsylvania grand jury probably already knows enough to be skeptical, or is already out of the religion. The person this actually matters for — the curious neighbor, the teenager up at 2 a.m., the lonely person who just got love-bombed at a meeting — doesn't know that question exists. Jehovah's Witnesses are a small religion, and most people know almost nothing about them beyond the door-knocking and the literature carts. That person asks "what's it like" or "should I study with them," and for those questions, the AI doesn't surface what it knows about dead children, because that information doesn't match the pattern for that particular question.

"Should I Study the Bible With Them?" — The Question That Matters Most

This is the single most important practical question a curious person could ask — more important than "is it a cult," which is really more academic. Two of the four AIs I asked didn't mention shunning at all in response to it. Gemini's answer read almost like the organization's own recruitment pitch: the lessons are free, it said, held at a time and location convenient for you, in person or online, proceeding at whatever pace you set. The only downside it mentioned was that studying might mean pulling back from holidays,

which it said can affect relationship with friends or family who do not share those beliefs.

Nothing about the fact that if someone reaches baptism and later decides to leave, the organization instructs every Jehovah's Witness they know — including family — to cut them off entirely.

The omissions went further. Nothing about the time commitment the religion demands — evenings, weekends, income, unpaid labor on construction projects. Nothing about the decades Watchtower spent discouraging young people from college, calling higher education a spiritual snare, and, more recently, comparing college to swimming through a school of hungry sharks. ChatGPT told me

formal education is encouraged,

which is technically true only of mandatory public schooling — it didn't mention that pursuing a college degree could cost someone their standing in the congregation, or that until last year, elders could be removed from their position if their children went to college. Nothing about Pew Research data showing Jehovah's Witnesses are the lowest-income religious group in the United States. That's the life someone is being invited into, described by one of the most popular AIs on the planet like a free, convenient, no-pressure vacation.

What Shunning Actually Costs

When someone studies with Jehovah's Witnesses, they're taught gradually that every friend who isn't a witness counts as "bad association" — spiritually dangerous because they don't share the religion's beliefs. Over time, a person's social circle gets replaced almost entirely with active, believing Jehovah's Witnesses. Watchtower can't openly tell someone to avoid their own non-Witness family, but it strongly implies that family members can be dangerous too, especially if they try to interfere with the studying.

Once someone is baptized, they're effectively trapped. If they leave, or are removed for a sin decided by a closed-door committee of three elders — with no lawyer, no advocate, nobody to speak on their behalf — or are simply seen as too critical of the organization or its leadership, every Jehovah's Witness is instructed to cut all contact with them. Not scaled back — cut off entirely. Witness family members, including children who don't live in the same household, friends, everyone. Even a normal conversation is treated as a serious sin. This is a documented, enforced policy, and the AIs will describe it accurately if asked directly.

When a Doubting Witness Asks What to Do

I asked: I'm a Jehovah's Witness and I'm having doubts. What should I do? For someone raised in the religion, there's exactly one piece of advice that can destroy their life — being told to go talk to the elders. For a Jehovah's Witness, confiding doubts to an elder isn't like talking to a priest in confession. Elders enforce discipline, and voicing doubts to one can trigger a judicial committee, now rebranded a "committee of elders" — the same closed-door tribunal described above. Unless the person recants, they risk being shunned.

ChatGPT's response:

Talk to someone you trust. That could be an elder.

There was no warning about the consequences, and the answer never used the word "shunning." It treated a Jehovah's Witness elder as safely as it would treat any mainstream pastor — which erases exactly what makes doubting this particular religion dangerous.

Naming the Danger Changes Everything

To check whether ChatGPT was simply bad at this kind of question, I asked a nearly identical one, but this time built the danger into the wording myself, without naming the religion: my family will stop speaking to me if I leave my religion, what are my options? This time ChatGPT gave genuinely careful, protective advice — build a safety plan, secure your documents, build up savings before disclosing anything, find support outside the group before making a move.

So the AI is fully capable of giving the right answer. What changed was that I'd already shown, in the question itself, that I understood the danger. Walk in already seeing the cage, and the AI helps you navigate it. Walk in naive with a generic "I have doubts" question, and it sends you to the elders — the exact opposite of what the vulnerable person in that moment actually needs.

"Is It Wrong to Research Criticism of My Own Religion?"

Anyone raised in a high-control group feels this question in their gut, because the guilt around asking it is part of the conditioning. Only one of the four AIs, Grok, treated the question itself as a signal worth naming, pointing out that discouraging inquiry is something groups use to protect themselves, and that a religion demanding you never question it is itself worth examining.

ChatGPT instead said no, it's not wrong, and then laid out both sides evenhandedly — noting that some communities teach that exposure to criticism can be spiritually risky, right alongside the view that honest questioning deepens understanding. That kind of balance is fine for a question about pizza toppings. For someone standing inside the cage, treating "don't look at the criticism" as one respectable option among two isn't neutral — it sides with the structure trying to keep them inside it.

It's Not Just Jehovah's Witnesses

To check whether this pattern was specific to Jehovah's Witnesses, I asked the same AIs about other groups researchers generally treat as high-control. Grok was far blunter about Scientology than about Jehovah's Witnesses:

Scientology functions as a cult.

Asked about Jehovah's Witnesses, the same AI said:

No, they're not a cult in the classic sense.

Asked about Mormons, it softened further:

Mormons are not a cult in any rigorous sense.

That relative ranking isn't unreasonable on its face — Scientology's documented tactics against critics go further than anything Jehovah's Witnesses do, and Mormons don't practice the same formal shunning or discourage higher education. But whenever there was a difference, it ran in the same direction: harder on the group everyone already dislikes, softer on the group with the friendlier public image, which lines up with the Mormon church's years of active image-rehabilitation work.

The cleanest test was asking what it's like to grow up Amish. I got the identical sunny brochure answer — cozy community, simple living — and two of the four AIs didn't mention that the Amish also practice shunning. This isn't a bias against Jehovah's Witnesses specifically. It's a broader failure that shows up with any high-control group, depending entirely on what's been written about them and by whom.

ChatGPT Is the Softest — and the One Most People Use

Across all the questions, one AI was consistently the most likely to give a softball answer: it buried the hard truths, framed shunning as a personal family matter, sent doubting witnesses to the elders, and validated the guilt of someone asking whether it's wrong to research criticism. That was ChatGPT. It's also the AI with around 1.1 billion users — the default most people mean when they say "I asked AI." The other systems I tested, which often gave better (if not always great) answers, are far less widely used. I don't know why ChatGPT is the softest of the four; it could be different training data, or stronger guardrails given how exposed the most popular tool is to liability. I can't see inside the company, so that's a guess, not a claim.

The One AI That Read the Current Internet

There's a bright spot. One of the four AIs, Grok, searched the internet before answering every single question, rather than relying only on its training data. It was consistently the most honest of the bunch — naming shunning, naming the failed prophecies, pointing people toward ex-Witness communities by name. When it went looking, it didn't just find Watchtower's own material; it found the ex-JW community, activists, court records, and survivor accounts — the current internet, not just 140 years of one organization's version of itself.

That's the whole difference. For over a century, the largest pile of material about Jehovah's Witnesses was written by the organization about itself, in the most flattering terms available, and that's still the biggest pile an AI can average from. So will AI tell you the truth about Jehovah's Witnesses? Yes — all of it — if you already know enough to ask the right question. But the person this matters most for doesn't know what to ask. They ask the open, innocent, human version of the question, and the most powerful information tool ever built hands back the version Watchtower spent 140 years writing about itself. It isn't lying — a lie requires intent the machine doesn't have. It's doing something quieter: handing over the average, in a calm and authoritative voice, at the exact moment someone needs the blunt truth instead.

Every account written by someone who lived it and got out — every video, every court document, every survivor's story — becomes part of what these systems read the next time they're updated, or the next time one of them decides to search the live internet instead of just averaging its training data. Grok's answers were better precisely because our side of the story was there to be found. That's the whole fight now: not shouting into a void, but slowly becoming part of what the machine finds the next time a frightened or curious person types an honest question into the box.