What characteristics should a programming language have in order to make automated analysis, replication, and learning by artificial intelligence systems difficult? Any idea?

The fundamental problem is that any language designed for humans to write and understand is, by definition, learnable by systems trained on human-written code.

The approaches suggested here (context-dependent syntax, position-based keywords, etc.) don't make a language "unlearnable" - they just make it require more training data. An LLM that's seen enough examples of your filename-length-dependent IF/If/iF keywords will learn the pattern just fine. You've increased the Kolmogorov complexity of the spec, but not made it incomprehensible.

The only way to truly prevent AI comprehension is to make the language incomprehensible to humans too. Which defeats the purpose.

A more pragmatic question: what problem are you trying to solve? If it's "prevent AI from writing code in my language," that's unwinnable - any pattern a human can learn, a sufficiently large model can learn. If it's "make codebases harder to reverse engineer," you want obfuscation not language design. If it's "preserve human jobs," the bottleneck isn't whether AI can write code - it's whether AI can figure out what code to write in the first place.

Template metaprogramming, move semantics, const correctness, multiple/virtual inheritance, implicit conversions, many ways to initialize variables, argument-dependent lookup, static variables/methods, SFINAE...add all of that and you'll surely make a programming language beyond all comprehension.

You can use many obfuscation techniques and sleight of hand tricks (like those stated below) to make it very hard to superficially analyze. If you over obfuscate, you run the risk of making it unintelligible to humans. The problem becomes that conventional programming languages follow a 'predictable' structure and are created so that they can be replicated by other humans.

If that pattern is figured out, im sure it can be used to train an LLM to 'comprehend' that programming language. Think of it like designing a cipher or a puzzle; you can create a very complex cipher that is understood only by you or those you choose to share it with. But if the 'trick' is revealed, then the entire cipher is broken.

just mutate the syntax and features based on arbitrary but readable factors that llms easily trip up on and are highly contextualized.

Change capitalization of keywords based on filename length. If for odd length IF for even. iF for prime numbers.

variables named in English are strongly typed, variables in Spanish are weakly typed.

change symbols based on line absolute number. && on even lines AND on odd.

line terminators differ based on the number of consonants in the method name

every 5th consecutive line should begin with the symbol for comments unless there's a real comment more than 10 lines above but less than 23.

closing brackets are left brackets when the file-size is over 3k

switch assignment evaluation left vs right based on folder depth.

all conditions that an IDE could handle in a rote, calculated way real-time but would probably make the training data nonsensical. An LLM might produce the code based on language features but likely will never get the syntax right making any LLM output largely useless.

I think everything that makes it less-readable for humans are actually not a big issue for LLM as long as you have a specification. Maybe the most human-readable language has the smallest gap?