Honestly, I wouldn’t be surprised, if various projects adopt a no-AI-merge-requests policy, simply because it makes reviewing absolutely shit.

Human-written code is predictable. When there’s a change in a file, you know that they didn’t make that change for the fun of it. When a change is more complex than you would’ve expected, you can assume that they probably tried the simpler solution first. Just in general, you can assume that they thought about how to structure the code in a logical fashion, so that the structure helps you to understand the contents. And if in doubt, you can always still ask them.

With AI-generated code, none of these apply, which means you can’t look at the code like a story being told, but rather you have to evaluate each line almost in isolation to decide whether it’s good or bad. Which is much harder and incredibly tedious.

Also, while the article states it almost like an axiom, I’d like to emphasize once more how important it is to me that I’m able to teach you something when I review your code. It genuinely feels like “Why in the fresh hell am I looking at this code and explaining what’s wrong, if the recipient has no idea what I’m saying and feeds it back into an AI at best?”.

In the majority of cases, I don’t review your code, because I want that feature to be merged, but rather because I have a deep love for humans and want to make you better.
If you’re not up for that and would rather have me talk to a vapid chatbot, please just don’t. You can create a fork and push whatever AI-generated changes you want to that, without me having to look at it.

@[email protected]

How one can even begin to (properly and non-biasedly) differentiate between AI-generated and human-produced activity?

Not just code, but… everything, from textual to graphic media… Even videos and “computer usage” (i.e. operating a computer, browsing a website) can be done by AI nowadays.

AI Detectors? Ironically enough, they’re as AI as the tool you suspect the person used (so you’re using AI too, how better do you think you are to judge those who use AI to generate content?). Also, they’re far from accurate, yielding both false negatives (e.g. content generated with very specific and tailored prompts, and/or using old/distilled/unknown Language Models and/or post-edited by a human) and false positives (e.g. Neurodiverse people accused of being AI), which is quite dangerous to say the least.

Maybe vibe detection? “My guts are feeling like this thing is AI” is as “accurate” as AI detection tools.

Yeah, there are things that everybody can feel and realize it’s AI-generated, such as those thick-line yellowish cartoons, sycophancy plus emoji-prefixed section titles plus excess of em-lines plus follow-up suggestion at the end (typical patterns from ChatGPT), content which happens to have a visible watermark (e.g. Sora videos), among other unambiguous scenarios.

Outside that, it’s no different of a game of guessing, a dangerous game when it comes for human relations, because it may lead to irreparable professional damage to a dev’s reputation (for code and PR being accused/labeled as AI) or societal damage to an artist’s reputation, it’s serious accusation to do.

I say this because I’m a ND myself and I was accused of being AI many times due to the way I express and communicate, and since I openly have a nuanced take on AI (I’m neither anti-AI nor pro-AI, I’m aware of both the pros and cons and this is enough for radical anti-AI people to point fingers at me while dismissing all the points I share with them against AI), and I’m certainly not the only one facing those kinds of accusation.

People often have no idea how this situation can shatter and consume a person from within, especially when the said person put their own blood on producing something (be it code, art, or whatever) just for the thing to be dismissed as AI-generated and/or ghosted to Internet oblivion, with no proper means of recurring or defending themselves from such accusations.

I like these guidelines, maybe I’ll share them with my team. It’s nuanced and reasonable enough to be a minimal common ground