Hacker News Reader

Comment by wdb | Hacker News Reader

How would you recognise AI-generated content? Is it a particular writing style? As a dyslexic person I am having a hard-time recognising it

View on HN

Stating the thesis then (repeatedly) re-stating it with a (very) slightly different wording was a red flag. Professional journalists don't write like that, but AI and schoolkids trying to pad out their essays do.

I'm guessing that a human did edit this AI output, just not very well.

We're all facing the need to build a mental list of which news sources are publishing AI-generated garbage, so that we can disregard them.

I'm still surprised to see that, as far as I can tell, no news outlets have made a public commitment to never use AI in their writing. Seems like it would be an easy way to promote the brand on commitment to quality.

I've griped about this before, but here we still are. We now know that MSN news, for instance, has no credibility due to their publishing AI-generated misinformation. https://news.ycombinator.com/item?id=39043135

I think you’re giving too much credit to “professional” journalists. This is also an excellent way to hit a certain word count if you are a few sentences short. Basically you bloat it up without needing to write anything new. Listicles and other pre LLM blogspam did this all the time.

ChatGPT loves to add "it's important to remember" to every second answer. Also "multifaceted" and a few other adjectives like that.

objection -- ChatGPT does not "love" anything.. a reason to repeat that phrase is that it is written often, for real communication reasons. Declaring that something written often by humans is now an indication of ChatGPT is wrong-direction IMHO and directly objectionable here, a place where reasoning humans think and communicate via writing.

ChatGPT was additionally finetuned after the initial training of the base model, so it's not completely clear if it truly represents the original distribution anymore. I use open weights models and they repeat such phrases less frequently in my experience.

That's completely untrue. Try using any base (non finetuned model) such as llama base, mistral base etc and you'll see that it does not write like that. That response format and writing style was added by OpenAI during the finetuning stage.

the post says "a reason to repeat that phrase"

not THE reason, not THE ONLY reason, so it is not COMPLETELY untrue. agitated?

I can visualize a little cartoon Einstein scribbling away on his chalkboard, but everything I'd actually project on the chalkboard would be nonsense. Relying on my cartoon output or thinking it has value would be madness.

Someone who's studied the subject quite a bit might visualize an actual Einstein chalkboard from memory. Their output could be verified and referenced, but looking for new Einstein-level math on the chalkboard would be madness.

Someone who Einstein himself would consider a peer might use this visualization method as a way to do actual work.

If we're assigning value to our chalkboards we'd be able to explain why we chose the numbers -1, 0, and 1. This would bias any math we'd do towards one of the chalkboards depending on our intent. At this point my chalkboard is useful as a filter.

Putting this all together, we'd see that the overall look of the cartoon responses would be a blend of 0 and 1 styles and would depend on our requests e.g. a reference request would look mostly like 0's art style. My own personal art style will be intentionally absent because it's only ever framing nonsense, by my own admission.

long before CNET was admitting to this it was obvious to me. The writing has the vocabulary and grammar of a accomplished academic but no ability to actually convey the important points concisely.

An easy tell is "level of effort given context"

Reply on news.ycombinator.com