Hacker News Reader

Comment by CharlieDigital | Hacker News Reader

Startup I'm at has generated a LOT of content using LLMs and once you've reviewed enough of the output, you can easily see specific patterns in the output.

Some words/phrases that, by default, it overuses: "dive into", "delve into", "the world of", and others.

You correct it with instructions, but it will then find synonyms so there is also a structural pattern to the output that it favors by default. For example, if we tell it "Don't start your writing with 'dive into'", it will just switch to "delve into" or another synonym.

Yes, all of this can be corrected if you put enough effort into the prompt and enough iterations to fix all of these tells.

View on HN

> if we tell it "Don't start your writing with 'dive into'", it will just switch to "delve into" or another synonym.

LLMs can radically change their style, you just have to specify what style you want. I mean, if you prompt it to "write in the style of an angry Charles Bukowski" you'll stop seeing those patterns you're used to.

In my team for a while we had a bot generating meeting notes "in the style of a bored teenager", and (besides being hilarious) the results were very unlike typical AI "delvish".

Of course the "delve into" and "dive into" is just its default to be corrected with additional instruction. But once you do something like "write in the style of...", then it has its own tells because as I noted below, it is, in the end, biased towards frequency.

Of course there will be a set of tells for any given style, but the space of possibilities is much larger than what a person could recognize. So as with most LLM tasks, the issue is figuring out how to describe specifically what you want.

Aside: not about you specifically, but I feel like complaints on HN about using LLMs often boil down to somebody saying "it doesn't do X", where X is a thing they didn't ask the the model to do. E.g. a thread about "I asked for a Sherlock Holmes story but the output wasn't narrated by Watson" was one that stuck in my mind. You wouldn't think engineers would make mistakes like that, but I guess people haven't really sussed out how to think about LLMs yet.

Anyway for problems like what you described, one has to be wary about expecting the LLM to follow unstated requirements. I mean, if you just tell it not to say "dive into" and it doesn't, then it's done everything it was asked, after all.

I mean, we get it. It's a UX problem. But the thing is you have to tell it exactly what to do every time. Very often, it'll do what you said but not what you meant, and you have to wrestle with it.

You'd have to come up with a pretty exhaustive list of tells. Even sentence structure and mood is sometimes enough, not just the obvious words.

This is the way. Blending two or more styles also works well, especially if they're on opposite poles, e.g. "write like the imaginary lovechild of Cormac McCarthy and Ernest Hemingway."

Also, wouldn't angry Charles Bukowski just be ... Charles Bukowski?

> ...once you've reviewed enough of the output, you can easily see specific patterns in the output

That is true, but more importantly, are those patterns sufficient to distinguish between AI-generated content from human-generated content? Humans express themselves very differently by region and country ( e.g. "do the needful" in not common in the midwest, "orthogonal" and "order of magnitude" are used more on HN than most other places). Outside of watermaking, detecting AI-generated text is with an acceptably small false-positive error rate is nearly impossible.

All of what you described can change wildly from model to model. Even across different versions of the same model.

Maybe a database could be built with “tells” organized by model.

Exactly. Fixing the old tells just means there are new ones.

> Maybe a database could be built with “tells” organized by model.

Automated by the LLMs themselves.

No thanks, I’d like it to be accurate ;)

Regular ol tests would do

I should have been more precise. I meant the LLMs would output their tells for you, naturally. But that's obvious.

They can’t know their own tells… that’s not how any of this works.

Thinking about it a bit more, the tells that work might depend on the usage of other specific prompts.

Not sure why you default to an uncharitable mode in understanding what I am trying to say.

I didn't say they know their own tells. I said they naturally output them for you. Maybe the obvious is so obvious I don't need to comment on it. Meaning this whole "tells analysis" would necessarily rely on synthetic data sets.

I always assumed that they were snake oil because the training objective is to get a model that writes like a human. AI detectors by definition are showing what does not sound like a human, so presumably people will train the models against the detectors until they no longer provide any signal.

The thing is, the LLM has a flaw: it is still fundamentally biased towards frequency.

AI detectors generally can take advantage of this and look for abnormal patterns in frequencies of specific words, phrases, or even specific grammatical constructs because the LLM -- by default -- is biased that way.

I'm not saying this is easy and certainly, LLMs can be tuned in many ways via instructions, context, and fine-tuning to mask this.

Couldn't the LLM though just randomly replace/reword things to cover up its frequency in "post"?

Reply on news.ycombinator.com