Indeed, but that’s not what you’re seeing in the example I showed you. No corrupted message or other prompt hacking technique was used to get Grok to declare a non-existent Pope canonized Gregory XVII. AIs suffer from a problem called hallucination, and that’s why all major AIs include a warning that they can make mistakes - a warning message you seem to be ignoring in attempting to use AI as an infallible source.
Hallucination is a real issue, and as someone who has entered the world of AI consulting in a professional capacity as an annex to my existing systems engineering business, and who has written over 50 AI-based applications, not only this, but also the fact that minor differences in wording which would seem inconsequential to the human user, can substantially alter the output of the question.
Also, your argument presupposes another technical inaccuracy, that AI behavior is deterministic (on all useful LLMs, its not, rather, a variable called temperature is used, and if this is set to a non-zero value, which it normally is (AIs with a temperature of zero it turns out aren’t particularly useful or interesting) the result being the same input prompt will by default produce different output.
So even if you were to run ask the same question on multiple AIs a multiple number of times, you still would not have a useful proof of whatever point you were trying to make - AI is, by itself, not reliable.
Now, the correct course of action, and there is a correct course of action, is to ask the AI what sources it used to make the statement you’re trying to use in an argument, examine those sources to make sure they say what the AI thinks they said, and if so, include them directly. In this manner one can avoid the embarrassment of the attorneys who recently in a federal case made the mistake of trusting one of the most reliable AIs not to generate spurious precedents - and it let them down, and the Federal judge was most displeased.
All uses of AI by itself as a direct source are appeals to an unqualified authority, in the same way that college professors have rightly regarded most or all of Wikipedia as an unreliable authority - an AI prove nothing other than the contents of their output. However, like Wikipedia, they can be used to find sources which are reliable and can be quoted. Although even in this case I would warn you are in for some frustration; if the information comes from their pretraining data, they are often unable to name the exact source (in this case the solution is to ask them to find external validation for their claim).
Now that I’ve explained the theoretical reasons why your argument is unreliable, take a look at what I got when I asked a freshly initialized ChatGPT 5.1 the very question you proposed:
View attachment 374119
I would assume that answer does not correspond with what you were expecting based on the content of your posts? Of course, AI can make mistakes, although our Lutheran friends
@MarkRohfrietsch @ViaCrucis @Ain't Zwinglian and others might well agree with chatGPT’s answer based on how denomination is defined (which I warned you in a previous post is an ambiguous term) and how sola scriptura is defined (presumably chatGPT 5.1 based its decision on Martin Luther’s historical definition of Sola Scriptura, which would incline towards a selection limited to Lutheran and Calvinist churches - but only Calvinist churches with a unified administration would qualify, whatever defines an administration - another ambiguous term. At any rate, within certain definitions of the ambigous parameters, I don’t see that chatGPT is in error.
