[neya]

I can tell you this much - most people who are opposed to AI writing blog articles are usually from the editorial team. They somehow believe they're immune to being replaced by AI. And this stems from the misconception that AI content will always sound AI, soul-less, dry, boring, easy to spot and all that. This was true with ChatGPT-3xx. It's not anymore. In fact, the models have advanced so much so that you will have a really hard time distinguishing between a real writer and an AI. We actually tried this with a large Hollywood publisher in-house as a thought experiment. We asked some of the naysayers from the editorial + CXO team to sit in a room with us, while we presented on a large white screen - a comparison of two random articles - one written by AI, which btw wasn't trained, but just fed a couple of articles of the said writer on the slide into the AI's context window, and another which was actually written by the writer themselves. Nobody in the room could tell which was AI and which wasn't. This is where we stand today. Many websites you read daily actually have so much AI in them, just that you can't tell anymore.

[A_D_E_P_T]

Counterpoint: GPT-4 and later variants, such as o3 and 4.5, have such a characteristic style that it's hard not to spot them.

Em dashes, "it's not just (x), it's (y)," "underscoring (z)," the limited number of ways it structures sentences and paragraphs and likes to end things with an emphasized conclusion, and I could go on all day.

DeepSeek is a little bit better at writing in a generic and uncharacteristic tone, but still... it's not good.

[bpodgursky]

If you ask them to speak in a different voice, they will. It's only characteristic if the user has made no effort at all to mask that it is AI generated content.

[A_D_E_P_T]

Sure, but 99% of them don't bother, even when they ought to know better.

And even when the most obvious "tells" are removed, articles can sometimes nevertheless seem AI-written. Just check this one out:

> https://searchengineland.com/ai-visibility-aexecution-proble...

[vonunov]

ChatGPT voice: https://news.ycombinator.com/item?id=44243641

[whywhywhywhy]

> We asked some of the naysayers from the editorial + CXO team to sit in a room with us, while we presented on a large white screen - a comparison of two random articles - one written by AI, which btw wasn't trained

Needlessly close to bullying way to try and prove your point.

[neya]

> We asked some

Which part of this looks like bullying? It was opt-in. They attended the presentation because they were interested.

[koakuma-chan]

Have you tried gptzero?

[vonunov]

Well, at least it passes the basic sanity test I've been using. ZeroGPT, on the other hand, gives a "100% AI" rating to the introduction and preamble of the US Declaration of Independence (from "When in the Course of human events" to "provide new Guards for their future security").

[neya]

Yep, it is not able to recognize. To be fair, it's not just dump it into ChatGPT and copy paste kind of AI. We feed it into the model in stages, we use 2-3 different models for content generation, and another 2 later on to smoothen the tone. But, all of these are just OTS models, not trained. For example, we do use Gemini in one of the flows.