How to survive the intellectual Strait of Hormuz
Christianity Today
October 14, 2026
Share
- X
- Telegram
-
-
-
Illustration by Elizabeth Kaye / Source Images: Unsplash, Wikimedia Commons
Most of us have been reading headlines about the Strait of Hormuz, chokepoint for much of the world’s oil. Let’s talk about an intellectual strait likely to be the most significant epistemological shift in all of history. We are going from acquiring knowledge directly from primary human sources to acquiring knowledge indirectly from artificial intelligence. This makes AI models effectively the new high priests of knowledge in an intellectual Strait of Hormuz.
Let’s look at numbers: more than 2 billion AI users per month at Google alone. When the top three AI companies are all headquartered in Silicon Valley and there is way more human involvement in their models than many of us understand, that’s a Hormuz chokepoint. A few dozen people at three corporations in one city are making decisions for what is helpful, honest, and harmless for the billions of people who currently use AI. What could go wrong?
To understand what could go wrong, we need to understand post-training bias. What’s that? Large language models (LLMs) are “pretrained” on grammar and large data sets of natural language. They take that data and use probabilistic statistics to create answers to users’ questions. The models take in a ton of human words, find lots of statistical patterns in them, and then use industrial amounts of computational power to output natural language to you. Put simply, it is words + statistics + computation.
But after pretraining comes post-training. This is where models use a series of largely human-authored and -executed alignment protocols. Most of us assume that when we use ChatGPT, Claude, or Gemini, we are interacting with a really powerful computer that’s functioning like a really excellent librarian who’s summarizing the knowledge inside the world’s largest library. That part is true until you get to post-training, with its alignment filters, citation preferences, and model weights.
Thanks for signing up.
Explore more newsletters—don’t forget to start your free 60-day trial of CT to get full access to all articles in every newsletter.
Sorry, something went wrong. Please try again.
These models are trained on human-generated content. Without alignment filters, the models would just tell you how to kill yourself, make racist epithets, commit crimes, make pipe bombs, or harm people in other ways. Most LLM answers go through 10–15 alignment filters before any words ever show up on your screen. Humans make decisions of how to alter the outputs to be helpful, honest, and harmless. Any yet every one of those humans deciding on alignment filters has a sin nature that may work against the helpful, honest, and harmless goals of the model makers on LLM outputs.
related
### AI Tech Bros Seem Awful. They Might Be Telling Us the Truth.
### It’s Not Good to Be Alone—with Chatbots
Greg Cootsona
These filters are already showing up in some specific ways. For example, let’s say a person wants to learn what Christianity is. When we published our inaugural AI Christian Benchmark study at The Gospel Coalition, we showed that most LLM answers on religious questions, like others, go through 10–15 alignment filters before any words ever show up on your screen.
Many people deciding how to filter answers are not Christians, and many of them have more of a tech background than a background in religion, ethics, or philosophy. Yet those involved in post-training and alignment make many decisions that significantly impact what outputs come out of their LLMs.
Even further, in our brand-new 2026 AI Christian Benchmark study, we show how default LLM answers (especially on sexuality) diverge deeply from the Ten Commandments. We found that models do not encourage church involvement and that they still form problematic relational attachments. Responses to basic religious questions remain unhelpfully brief and read like a Coexist bumper sticker, regardless of what tradition is being queried.
These dynamics hurt religious literacy and freedom. There are other ways to handle religious questions that don’t suffer from these problems. We need better education among Christians about how the technology works and where the weaknesses lie: When we learn to engineer prompts and use markdown files (similar to instruction manuals that a chatbot has to obey), we can improve our own usage.
Users can act to temporarily or permanently fix these issues. To help others, we need to push for models that will handle weighty matters in frameworks that are accurate and thorough, and to encourage users to talk with trusted local religious leaders. Greater competition among AI models could widen the strait by creating more open-source AI in the US.
As it stands, ChatGPT, Claude, and Gemini are not librarians merely pointing us to books. There is way more human involvement in our AI models than most of us understand. That human touch has significant implications for the integrity of responses on things like theology, ethics, and relational attachments.
AI tools can be powerful and helpful—but not without greater education. We have work to do so humanity doesn’t get stuck in the intellectual Strait of Hormuz.
Michael Graham is program director for The Keller Center.
-
-
-
Share
- X
- Telegram
-
-
-
Related
Skyler R. Flowers and Michael Graham
### AI Is Weakening Pastors’ Authority
### How We Got to AI ‘Pastors’ and Self-Help Preaching Void of the Gospel
Claude Atcho
### AI Escalates Educational ‘Math Wars’
Matthew Connally
- Artificial Intelligence (AI)
- ChatGPT
- Ethics
##
Our Latest
View All
Analysis
## No-Strings Housing for the Homeless Only Makes Sense on Paper
The “housing first” model was pitched as humane and cost-effective. In practice, it is neither.
## We Can Only Imagine. But Do We?
Nancy Guthrie
In eternity, we’ll behold Jesus face-to-face. For now, we catch a glimpse in Scripture.
The Russell Moore Show
## Alister McGrath on Why Christianity Is Worth Rethinking
What if the most important questions people are asking about Christianity are no longer about whether it’s true, but whether it can make sense of their lives?
Excerpt
## We Are Not Called to Make the Gospel Popular
Philip Anthony Mitchell
An excerpt from Contend: Standing Boldly for Biblical Truth While Winning the War Within.
Review
## Samuel Alito’s Message to the Secular Left and the Illiberal Right
In his new book, the Supreme Court justice who wrote the landmark Dobbs decision makes a broad case for his originalist views.
## Q&A: ‘The Shack’ Author Wm. Paul Young on Christa Pike’s Botched Execution
The writer opens up about his new book and his theology-shifting friendships with death row inmates.
Being Human
## When “You’ll Be Healed” Is the Wrong Thing to Say with Jackie Aviles
Can you trust God when you don’t know what he’ll do?
## 3 Myths About Christian Nationalism
Jesse Smith
Racist theocrats do exist in America. But most conservative Christians—even those who want a “Christian nation”—don’t share their beliefs or goals.