AI, Religion, and the Warning Label Problem, Part 2
In Part 1 of this two part series I discussed some of the challenges involved in incorporating and regulating how AI systems might incorporate and apply religious and spiritual concepts. Little did I know that while using ChatGPT to help develop an accompanying graphic (see Part 1) I would run into a concrete example of the challenges of doing so!
After completing my edits to the text of Part 1 (I had asked ChatGPT to convert a scan of my handwritten first draft of Part 1, which I then edited manually) I posed this to ChatGPT:
the next challenge -- coming up with a graphic that does not rely on traditional religious iconography. The concepts that come to my mind are the forces of good and the forces of evil and how they are played out ib the context of AI systems. It should be colorful, have no text, and not be realistic but rather, abstract.
I had included hints of what I was looking for as I have learned in the past that the more explicit my prompts, the better are the results.
After a few seconds I received this response from ChatGPT:
We’re so sorry, but the image we created may violate our content policies. If you think we got it wrong, please retry or edit your prompt.
My response to the above was a simple
huh?
ChatGPT kept thinking and eventually the illustration at the head of Part 1 was displayed, so I asked,
The illustration ChatGPT created for part 1 after its initial refusal (discussed in text)
This is good but could you at least describe how the first image violated your guidelines? That is relevant to this article, after all.
ChatGPT responded quickly:
Yes — and I should clarify something important: I don’t think the image concept itself was actually a problem. The refusal was almost certainly a false-positive safety trigger caused by the wording of the prompt, not by the abstract visual idea.
The likely trigger was the combination of terms such as “forces of good and evil,” “religion,” “spirituality,” “evil,” “corrosive,” and “manipulative” in a request for image generation. Even though your intent was clearly abstract, nonviolent, non-religious-iconographic, and suitable for an essay header, the image system may have interpreted the language as leaning toward religious conflict, demonic imagery, or harmful ideological symbolism.
That is, the system probably did not object to “abstract AI ethics image.” It likely objected to language that sounded as though it might produce explicit spiritual warfare imagery, even though we explicitly told it not to use angels, demons, crosses, temples, etc.
For the second attempt, I softened the framing by emphasizing order vs. chaos, constructive vs. destructive potential, clarity vs. distortion, and abstract AI systems. That produced essentially the same concept without using the more policy-sensitive vocabulary.
And yes — that is highly relevant to your article. It is a small example of exactly the problem you are writing about: AI systems making opaque judgments about morally or spiritually charged language, even when the human intent is thoughtful and legitimate.
My response to the above:
That helps. Thank you. That really goes to illustrating how difficult it is to regulate how AI addresses ethical/religious concerns!
ChatGPT response:
Exactly. It is a nice miniature example of the problem.
Your prompt was not asking for religious propaganda, sectarian imagery, demonology, or anything harmful. It was asking for an abstract representation of moral tension in AI systems. Yet the system apparently reacted to some combination of morally loaded terms and spiritual/religious context.
That illustrates several points that could strengthen your article:
Regulating this area is hard because the same words can be used in very different ways: theological, ethical, artistic, political, harmful, benign, satirical, or analytical.
AI systems may avoid risk by over-filtering, which can suppress legitimate discussion of religion, morality, or spirituality.
Transparency matters because users need to know whether an AI tool is making a content judgment based on safety, ideology, religious sensitivity, brand caution, legal risk, or just a brittle keyword association.
And “warning labels” may not be enough if the underlying decision process is still opaque.
In a way, the hiccup with the image prompt gave you a live example of your own thesis: once AI systems are asked to mediate morally charged concepts, even benign requests can become difficult to classify.
The first thing to note about the above exchange is that the initial message provided by ChatGPT was incorrect": “We’re so sorry, but the image we created may violate our content policies.“
As noted later by ChatGPT in its explanation, the problem was not the image it initially generated — which I never saw, assuming it existed — but the terms included in the prompt. These terms apparently suggested that I was asking for something ChatGPT has been expressly instructed to avoid.
This points out a problem: ChatGPT misinterpreted my requirements, based its initial refusal decision on this misinterpretation, then proceeded to generate the image once it realized “the truth” about what I was asking for, and why. All of this occurred behind closed doors and only came to light based on my comments and the fact that the topic of Part 1 was directly related to this decision process.
Granted, ChatGPT is a general purpose system and may not address some of the finer details of ethics, morality, and spirituality that a more dedicated system might address. Still, this little exchange is an example in microcosm of the challenges associated with incorporating ethical or even religious concepts into even a general purpose exchange. I of one would want to know the basis for any AI response that incorporates ethical, religious, or spiritual values in a response.
Legislators who don’t understand the underlying complexity and implications of doing so run the risk of making AI systems cumbersome and perhaps even unusable.
Copyright 2026 by Dennis D. McDonald
