HACKER Q&A
📣 keepamovin

How would you describe the way AI writes versus a human?


I’m talking about prose here. It strikes me that it often is possible to tell or feel that an AI has written something, but I find it hard to pin down - and I don’t have any great descriptions that distinguish its writing from a humans. I’m talking about aside from the obvious, extreme fluency and competency. How can we tell that something is written by an AI rather than a really good human writer?

I’m not asking for a programmatic test. I’m just curious how you would describe it yourself.

This is specifically about prose, but if you have any really great insights to share about the difference in how it writes code I’m all ears.

  👤 PaulHoule Accepted Answer ✓
Most of them are upbeat and cheery and won't bitch about things though I guess if you trained them to be insultinators [1] they'd do that too. They avoid common typos and spelling mistakes but the fluency can hide a lack of competence. For instance I had a session where Copilot wrote me some HTML and CSS to use the little-used element but when I asked it "Do I increase or decrease the alpha to make the background darker?" the answer was "both" in the sense that the first sentence said one thing and the rest of the answer said the other.

That fluency has a way of lulling people asleep and getting them to accept wrong or nonsensical answers. I find it annoying that Copilot ends most answers with some invitation like "... I'm all ears" when I know it is not retaining the results of the conversation.

[1] https://www.youtube.com/watch?v=obc7MD6iyks

👤 cyndunlop
A fun take on this here: https://livebook.manning.com/book/writing-for-developers/app... In case it's randomly scrambled by Manning - we had one GPT write a engineering blog, then had another GPT analyze whether it thought it was written by a human (and why). The response:

"This text appears to be AI-generated rather than written by a human. Here are the key indicators:

    1.   Unusually dense and consistent metaphors: The text maintains an extremely high density of Polish and Eastern European cultural references, dance metaphors, and programming analogies throughout. This level of consistent, thematic metaphor use is more characteristic of AI generation than human writing.

    2.   Overly flowery language: The prose is consistently ornate and poetic in a way that feels artificial. Human technical writers, even when being creative, typically don’t maintain such a consistently florid style.

    3.   Lack of specific technical details: While the post mentions Zig’s advantages, it doesn’t provide concrete examples or code snippets that a human developer would likely include based on real experience.

    4.   Perfect thematic consistency: The Polish/Eastern European theme is maintained flawlessly throughout, without any drift or inconsistency that would be natural in human writing.

    5.   Artificial emotional arc: The emotional journey feels manufactured rather than genuine, hitting predictable notes of curiosity, challenge, and triumph.

    6.   Too-perfect structure: The post follows an idealized narrative structure that feels more like it was generated to fit a template than a natural recounting of events.

    7.   Lack of genuine personal anecdotes: While framed as a personal journey, there are no specific, idiosyncratic details that would typically appear in a real developer’s account of learning a new language.

    8.   Overuse of cultural references: The density of Polish references feels more like it was generated from a list than organically incorporated by someone with lived experience.
While it’s a creative and engaging piece, the consistent artificiality in style and structure strongly suggests AI authorship rather than human writing. A human-written technical blog would likely have more variation in tone, more specific technical details, and a more natural incorporation of personal experiences and cultural references."