Comment by dyauspitr
9 hours ago
This meme of trying to make it sound like LLM text is so obvious is a joke. It’s literally not, you can tell it to write in literally any style and given just a bit of an example of a person’s writing style, frontier models copy it completely and effectively. This argument can probably be leveled at vanilla raw output from an LLM, but even the slightest attempt at obfuscation bears solid fruit.
If someone uses an LLM to write and is able to tailor their writing such that it isn't obviously written by an LLM, then I'm fine with it! But two of my otherwise-favourite news sources -- the Hacker News front page and FT Alphaville -- are inundated by articles where the LLM usage is blindingly obvious.
Well, give it a shot -- you'll likely find that that technique doesn't work nearly as well (at least with Pangram 4) as you think it might. When we had Max on the podcast[0], Adam explicitly asked him about exactly this (after all, you can give an LLM access to Pangram and let it iterate!), and Max reported that someone had attempted to do this -- and ended up burning through $700 in tokens and had a "sad Claude." Another interesting bit: according to Max, newer models are diverging more from human writing not less. I think that that was more anecdotal than quantified, but an interesting comment nonetheless.
[0] https://oxide-and-friends.transistor.fm/episodes/ai-detectio...
99% of college essays and pretty much everything “product” in corporate America is now LLM generated with some marginal oversight. It passes muster for the most part.
Talking about attempts to bypass ML detection:
>This argument can probably be leveled at vanilla raw output from an LLM, but even the slightest attempt at obfuscation bears solid fruit
Whoops, disproven by bcantrill's comment:
https://news.ycombinator.com/item?id=49582629
Let's talk about the detection ability of corporate normies instead:
>pretty much everything “product” in corporate America is now LLM generated with some marginal oversight. It passes muster for the most part.
Goalposts: moved.
You can adapt the output if you care to do so. Almost all do not care to do so.
They just don't care to put the slightest attempt because they have a blindness to the problem. They are not doing it to intentionally mislead people.
This isn't very effective on any models released in recent years. With older ones, you used to be able to influence writing style significantly by just putting examples in the context, but newer models have gone through so much assistant RLHF, they really want to revert back to their default "assistant voice" during their turn.
You can still influence their writing style in a broad manner that might look correct at a glance, but the repetitive little patterns that give it away will always be there - if it was that easy to get rid of them, don't you think the AI labs themselves would've done it before releasing the models?
I think nowadays defeating the detection probably looks like finetuning a smaller LLM and getting it to paraphrase the text from the other one (or just using a more obscure finetune: it'll probably have its own cliches and habits but it will be at least different). As an added bonus this also likely removes the fingerprinting from the output as well. But I think most people are not going to bother with this.