Post by @onekind@beige.party

2026-11-04 15:17 UTC

@thepurpleoctopus@lgbtqia.space AI has access to all the same cues and context a human listener would have when they are both working with audio. I am testing AI models for translating audio into text where I have known-good human translations to compare the output to. It isn't perfect but it surprises me now and then by picking semantically appropriate alternatives to the literal wording. GPTs refer backwards up the sentence tree to find words and phrases that determine the contextual likelihood of the next word in the sentence. When you use a large model on audio, it can pick up phatics, hesitation, intonation, cadence — all sorts of subtle signals as to inflection and contextual meaning. I recognise that AI is gutting the translation industry. That's taking bread out of my friend @AnthoDerv@famichiki.jp's mouth and I'm furious about the business models that are causing that. I do the testing work because in my field (public health) every second WASP practitioner is going 'finally I can communicate with those [ethnicity] people without having to run a focus group and pay for translation.' My goal is to identify the safety risks if they all charge down that path. But the bad news is the tech is actually pretty sophisticated. AI models lack two things by definition though: judgment, and experience. So an AI-translated or generated guidebook is going to give tourists a lousy experience because (a) it has no experience of travel or the locale and (b) it has no judgement of what's a good tourist experience. That's the place to focus the critique.

Replies (1)

@thepurpleoctopus@lgbtqia.space 2026-11-04 15:32

@onekind@beige.party @AnthoDerv@famichiki.jp I think you're making some genuinely interesting points here, and I agree that the technology is more sophisticated than many of its critics acknowledge. Audio models can pick up intonation, hesitation, cadence, and context, and sometimes produce surprisingly nuanced translations. We don't need to deny that to have a serious discussion about its limitations. One thing worth distinguishing, though, is that interpreting and translating are different professions. Interpreters translate live speech into live speech. Translators mostly work with written texts. Audio translation is a relatively specific use case, whereas my own work is primarily literature and academic articles. Those are rather different beasts. In literature, you're not merely transferring information. You're translating voice, rhythm, ambiguity, irony, cultural references, and the author's deliberate choices. In academic work, you're navigating terminology, argumentation, disciplinary conventions, and sometimes the difference between a defensible claim and an intellectual catastrophe. A model can produce a plausible sentence without understanding why that sentence matters, whether it preserves the author's meaning, or what gets lost in the process. And I think your public-health example highlights an important concern. "Finally, I can communicate with those people without paying for translation or consulting them" is not necessarily a triumph of accessibility. It risks becoming a colonial attitude with a shiny new interface. The question isn't simply whether the technology can translate a sentence. It's whether anyone has the judgment, experience, and accountability to decide that the translation is fit for purpose. A model has never navigated a foreign city, misunderstood a local custom, or experienced the consequences of bad advice. It has no lived experience to draw upon and no moral responsibility for the harm its output might cause. And while I appreciate your willingness to test these systems critically, I remain furious about the economic model that uses their capabilities to make human translators redundant while pocketing the savings. The technology may be impressive. The business model still deserves scrutiny. Human translators aren't obsolete because a machine can produce plausible sentences. We're doing something considerably more complicated than replacing words. We're helping people understand one another, with all the judgment, responsibility, and humanity that entails.

Open ##5480332