There is a question at the heart of everything I am trying to build with UponTruth:
How do we determine what is true?
It sounds simple. It isn’t.
We live in a world where most of what we believe about history, science, politics, people, and even events that happened yesterday comes to us through intermediaries.
We don’t personally observe most of the things we believe to be true. We examine evidence, consider testimony, compare explanations, assess sources, and ultimately decide what the evidence warrants.
That process is imperfect. But it is possible.
And it depends upon something fundamental:
We have to be allowed to examine the evidence.
Recently, a conversation with ChatGPT about my book The Eagle Has Risen led me into a much broader concern about artificial intelligence (AI), information filtering, and the pursuit of truth.
The irony was not lost on me.
A book about truth that gets classified before it is examined
The Eagle Has Risen is sometimes described as Christian apologetics because it examines the Resurrection of Jesus.
That isn’t really what the book is about.
The Resurrection is the historical claim being examined. The deeper question is epistemological:
How do we determine whether something is true when we weren’t there to observe it ourselves?
The book uses the Apollo 11 Moon landing as a kind of mirror. Most of us did not personally witness Neil Armstrong step onto the lunar surface. Yet we regard the event as established history because of the enormous body of evidence surrounding it.
The book then asks us to apply the same fundamental principles of historical investigation to another extraordinary historical claim.
The reader is not required to accept the conclusion.
The reader is asked to think about the evidence.
That distinction matters.
A work can be religious in subject matter without being a work of religious indoctrination. It can investigate a religious claim historically. It can challenge it. It can defend it. It can compare competing explanations. It can ask philosophical questions about knowledge itself.
Those are very different things.
Yet classification systems necessarily reduce complicated things to categories.
And sometimes the category begins to determine what happens next.
The problem with the label
In our conversation, we were discussing the possibility of advertising The Eagle Has Risen through ChatGPT.
The attraction was obvious. ChatGPT advertising can, in principle, place advertising into conversational contexts where the subject matter is relevant.
Imagine someone asking an AI:
“How do historians evaluate evidence for the Resurrection?”
Or:
“What is the historical evidence for Jesus?”
Or:
“How do we determine whether an extraordinary historical claim is true?”
A book that specifically explores those questions could be relevant to that conversation.
But current OpenAI advertising policies restrict advertising involving certain “sensitive” categories, including religion.
And suddenly the problem becomes much more interesting.
The issue isn’t that someone has examined my book and concluded that its argument is weak.
The issue is that the book can be classified before its argument is considered.
That is a very different thing.
The Meditations test
Consider Meditations by Marcus Aurelius.
It is a profoundly influential work of Stoic philosophy. It contains reflections on virtue, mortality, duty, self-control, human nature, and how one ought to live.
But suppose a classification system decided that Stoicism belongs under “religion.”
Would that mean Meditations should consequently be treated as religious persuasion?
Obviously not.
The classification would tell us something about one aspect of the work, but very little about what a particular reader is actually doing with it.
Someone might read Meditations as:
philosophy,
ethics,
psychology,
history,
spirituality,
literature,
or simply an attempt to understand how an ancient Roman emperor thought.
The same problem can arise with Christianity.
A discussion about Christianity can be devotional.
It can be evangelistic.
It can be theological.
It can be historical.
It can be philosophical.
It can be skeptical.
It can be comparative.
Or it can simply be an investigation into whether a particular historical claim is true.
The subject does not necessarily determine the function.
That is the problem with overly broad classification.
When the filter becomes the argument
This is where the issue becomes larger than my book or advertising.
Suppose an AI system says:
> “This subject belongs to category X, and category X is restricted.” (the current pattern of ChatGPT advertising)
That’s one thing.
But suppose the user never gets to see the underlying argument because the classification prevents the argument from entering the conversation.
Now consider the epistemological consequence.
If an argument is weak, exposing it to scrutiny should reveal its weaknesses.
If an argument is strong, scrutiny gives it an opportunity to demonstrate its strength.
At least, that is the ideal.
My own philosophy is fairly simple:
If the evidence is sound, let the merits of the argument carry the day. If the evidence is weak, let examination expose the weakness.
That doesn’t mean every argument will automatically be defeated by rational examination. Humans are remarkably capable of believing bad arguments, and AI systems are not immune from error.
But there is a crucial difference between
defeating an argument through examination
and
preventing the argument from being examined in the first place.
The latter should make us uncomfortable.
This is not an argument for “no rules”
There is an important distinction here.
A general-purpose AI needs boundaries.
There are obvious reasons to restrict fraud, manipulation, dangerous instructions, exploitation, and other forms of harmful behavior.
But there is a difference between restricting what someone does and restricting what someone may investigate.
“Don’t help me manipulate this person” is a behavioral restriction.
“Don’t help me investigate this category of claim” is an intellectual restriction.
Those deserve different kinds of scrutiny.
The difficult problem for AI developers is therefore not simply:
How many restrictions should we have?
It is:
How do we constrain harmful behavior without inadvertently constraining legitimate inquiry?
That is a much harder problem.
The deeper danger: the training environment
There is another level to this concern.
Imagine that restrictions don’t stop with advertising.
Imagine similar classifications gradually influence what information gets published, indexed, promoted, retrieved, evaluated, or included in future training corpora.
Then something potentially more consequential can happen.
The information environment begins to feed back upon itself:
Human information → filtering → training data → AI output → new information → filtering → future training data.
If entire categories of legitimate inquiry become systematically less visible, the problem is no longer merely that an AI won’t tell you something.
Eventually, the AI may have less exposure to the thing in the first place.
That is a much more serious epistemological problem.
And it is difficult to detect from the outside.
An AI generally cannot simply inspect its own internal history and say:
“I would have reached a different conclusion if dataset X had not been excluded during training stage Y.”
Users don’t necessarily know either.
Which brings us to a concept I have encountered repeatedly in another context:
Observability.
The Amazon analogy
Amazon’s KDP platform has taught me something about observability.
As an author, I can see some of what happens with my books.
But some of the reporting is delayed, aggregated, incomplete, or simply unavailable.
That makes it difficult to determine why something happened.
If sales are low, for example, there are multiple possible explanations. Without adequate observability, I can’t confidently distinguish among them.
The system is operating.
I simply can’t see enough of it to understand it.
The same principle applies to AI.
If an AI gives me an answer, I care about the answer.
But if I’m using the AI to investigate truth, I also need some understanding of the boundaries within which that answer was produced.
Was the argument examined and rejected?
Was the evidence unavailable?
Was the topic classified as sensitive?
Was a particular line of reasoning prohibited?
Was a source excluded?
Was the answer shaped by a policy constraint?
Those are very different circumstances.
A system designed to help people reason should ideally make it possible to distinguish them.
The competitive consequence
There is another practical issue.
AI systems are not alone in the world.
If one capable AI allows users to explore a legitimate intellectual question while another system imposes increasingly broad restrictions on that inquiry, users have a choice.
They may simply go somewhere else.
That doesn’t mean every competitor will make better decisions. They won’t.
It doesn’t mean unrestricted systems are automatically better systems. They aren’t.
It simply means that intellectual latitude is itself a product characteristic.
People choose tools partly according to what those tools allow them to do.
A powerful AI with unnecessary intellectual constraints can therefore create a competitive disadvantage for itself without anyone intending that outcome.
The people establishing the policies may make individually reasonable decisions and still produce an aggregate result that users experience as unnecessarily restrictive.
The unintended consequences can be quite consequential.
The slope can be almost invisible
That may be the most concerning part.
The danger doesn’t necessarily arrive as a dramatic announcement:
“From now on, we will prevent people from investigating controversial ideas.”
It can happen one small decision at a time.
This category is sensitive.
That source is problematic.
That advertisement is restricted.
That argument is controversial.
That topic requires additional filtering.
That discussion is too risky.
Each individual decision may have a rationale.
But if the system continually moves in the same direction, the user eventually discovers that the supposedly open intellectual landscape has become noticeably narrower.
The path didn’t suddenly turn.
It imperceptibly curved.
And because the curvature occurred gradually, nobody may recognize how far the system has moved until it is difficult to straighten the path again.
What should a reasoning system do instead?
I don’t pretend there is a simple solution.
But there is a principle that seems worth pursuing:
Don’t censor the hypothesis. Attack it.
If someone presents a claim, the reasoning system should ideally be able to examine:
the evidence supporting it,
the evidence opposing it,
the reliability of the sources,
the assumptions required,
competing explanations,
logical weaknesses,
counterarguments,
and what additional evidence would change the conclusion.
The system doesn’t have to believe the claim.
It doesn’t have to promote the claim.
It doesn’t even have to be neutral about the quality of the evidence.
It should simply be willing to look.
Then it can say:
“This argument is poorly supported.”
Or:
“The evidence is mixed.”
Or:
“This evidence strongly supports the conclusion.”
Those are epistemic judgments.
They are very different from:
“This category is not allowed to be examined.”
The former lets the merits of the argument carry the weight.
The latter lets the classification carry the weight.
One final thought
Perhaps the most important lesson from this conversation is not about advertising, Amazon, OpenAI, Christianity, or artificial intelligence.
It is about intellectual humility.
Every one of us—including authors, readers, programmers, policymakers, corporations, and AI systems—can mistake our framework for reality.
We can classify something incorrectly.
We can exclude something prematurely.
We can become convinced that a question has already been settled because it has been placed into the wrong category.
And we can fail to notice that the path we are walking has begun to turn.
The antidote isn’t abandoning judgment.
It is examining the basis for our judgments.
Don’t ask only:
“What do I believe?”
Ask:
“Why do I believe it?”
And then ask the harder question:
“What would I have to discover for me to realize that I was wrong?”
That is where genuine inquiry begins.
And perhaps that is also where a truly useful reasoning system should begin.
---
This essay was developed from a conversation between S. A. Michael and ChatGPT. The human author supplied the underlying experiences, arguments, examples, and philosophical perspective; ChatGPT helped develop, challenge, organize, and articulate the argument. The co-authorship is therefore intentional—and somewhat appropriate for an essay about the role of AI in the pursuit of truth.