Enough with the LLM BS, Already

I know that, in doing this, I contribute to the problem in its own way, but I simply can’t bear it anymore.

The AI frenzy I have seen among my fellow translators has to stop. It feels like I’m watching otherwise intelligent, literate folks suddenly spouting flat earth theory.

LLM/Generative AI/Whatever you want to call it doesn’t actually do what it says it does. It isn’t a useful tool. It doesn’t think, it doesn’t translate, it doesn’t explain. LLMs are sophisticated statistical algorithms that spit out words they got off the internet in “likely arrangements.” Any meaning that those words seem to have is provided solely by the reader. Any time you ignore the fact that nothing of what LLMs generate is rooted in factual reality, you are deceiving yourself. The idea that some part of the communication might be “wrong” is, in itself, a mistake. It’s all generated the same way, from statistical analysis and algorithmic generation. The good and the bad are both equally likely. Seemingly accurate BS and clearly inaccurate BS are both equally BS.

The industry calls LLM mistakes “hallucinations” but they’re really just expressions of the nature of the beast.

https://www.lakera.ai/blog/guide-to-hallucinations-in-large-language-models

The great success of LLMs is only in that they *feel* right, so no one really takes the time to check if they are, actually, so. It certainly seems amazing that the words string together in clear sentences that we can interpret as more or less connected to our prompts. But we are professionals at working with words. We should be holding our work to a higher standard that “Eh, it feels right.”

Because folks, there’s nothing under there. It’s empty. Let’s take a recent Facebook post froma a translator I saw. It was a query about the pros and cons of different translations of a tricky term. The results included sources, for example, and a lot of double talk that boiled down to “it all depends on the context.” If you actually took the time to read those sources, though, none of them actually supported any of the points of the response. They were just linked by including the terms in question. Sometimes. One of them was actually Vietnamese, which is odd given that the query was in English and talking about Japanese. But it had a (poor) English gloss of a phrase including the term, so… Source?

That is what LLMs *do*. They were trained not to be right, but to sound right. To be really convincing liars. (Which is actually a pretty easy thing to get away with when discussing elevated academic ideas, because they’ve been dominated by empty bloviating for decades, anyway. I’m sure LLM generated philosophy and linguistics papers are already filling journals, because no one ever read or understood that shit anyway.)

The trick of LLMs is, quite literally, a scam. As in, they seem to have evolved parallel to techniques designed specifically to fool people. The datasets were trained to generate sentences that felt right, rather than be right, by having them rated by non-expert users in developing nations. It is, as the kids say, all vibes. And the result is that the patterns are now set to be convincing, but nothing more.

You can see what I’m getting at here:

How much of AI’s recent success is due to the Forer Effect?

and here:

https://softwarecrisis.dev/letters/llmentalist/

The upshot is, if you adopt LLMs into your work flow, you are intentionally adopting a bullshit engine. Nothing that comes from an LLM should ever be trusted. There is no accountability for the inevitable errors and fabrications that they bring to your work. If ChatGPT were an employee, you’d fire them almost immediately for fraud. Which you would know if you applied critical thinking to the results you’re getting, rather than allowing the facade of legibility convince you that “The Machine Understands!”

What we as translators should be doing is not wondering how to use AI, we should be educating our clients about the dangers of trusting. Any work that is done by AI is without value. Literally. If the results of AI are good enough for a given task, it was never worth doing in the first place. Which is not to say that everything we do is equally valuable. Indeed, corporate boilerplate has always been BS, so a BS engine might, in fact, be exactly what you need to do that. Whether you feel comfortable engaging in that cycle is another matter.

And that’s not even getting into the heavy environmental cost of each query, or the deeply immoral and exploitative labor practices that have lead to the current status, or the terrible people running these shows.

I know that it seems like this is the way the world is going, but we don’t have to embrace the idiocy.

Anyway, all this is to say, enough already. Damn.