ChatGPT Is Apologetic, Claude Is Self-Righteous, Both Need a Human Translator

A freelance translator's accidental experiment with ChatGPT and Claude found that both AI chatbots degrade on longer documents, require constant coaching, invent plausible-sounding but false terminology, and exhibit cultural bias, concluding that human translators are still needed for accuracy and nuance.

It all happened very fast, too fast to just pivot, idiot. One minute I was a decent human with languages in my head https://correresmidestino.com/tag/work/ . The next I was a monkey, proofing a machine spitting out sentences faster than I can type—and arguing with it, to boot, because language is more than words strung together and we don’t live in a world where 0 or 1 is always the answer. Let’s call it my accidental experiment https://correresmidestino.com/sorry-i-was-busy-unfucking-my-ai-fucked-life/ with ChatGPT, Claude and the rest of the AI gang as translation tools. Over the past few months, several clients have gotten into the habit of sending me AI-translated documents and paying me a 50% rate because half the work is done, right? I also experimented on my own to see how obsolete I was on a scale of 10. The human isn’t winning the existential battle so far, but if you read this, you might understand why translation—and many other skills—isn’t as simple as “ just upload it to ChatGPT https://correresmidestino.com/dont-you-just-upload-it-to-chatgpt/ ”. ChatGPT and Claude struggle with longer documents And by longer documents, I mean several paragraphs. Let’s say I upload a standard 1,000-word press release to ChatGPT or Claude, with a comprehensive prompt and locked terminology. The first paragraph is so amazing that you immediately plan to start saving for your own data center. The second paragraph is so-so, with grammatically plausible French but endless English-shaped French: “dans les mains de chaque client,” “raconter l’histoire de la façon dont…,” “faire une impression festive”… It goes downhill from there, with the actual meaning wandering off and Google Translate-era lexical calques creeping in, as the model apparently starts processing each word individually and forgetting about the meaning altogether: “coupe fraîche” for “cool fit”, “gobelets froids” for “cold cups”, and “porte-cartes-clés” for “key card holder”. I can’t explain why from a technical perspective, but if you’re translating with AI, keep it on a short leash, one paragraph at a time. It’s like a schoolkid told to write an essay who starts playing a game on his phone the minute you stop checking on him. Yeah, at this stage it’s often faster to just translate it yourself. I can focus for longer than Claude or ChatGPT. ChatGPT and Claude require constant coaching You can feed ChatGPT or Claude a solid prompt and still have it completely disregard it at any point, for mysterious AI reasons. Or the model follows terminology without understanding why the terminology exists. See the example below. The prompt says that “green” should be translated as “écoresponsable” environmentally friendly , so whenever “green” shows up, that’s what it uses, never mind that “green” is also a colour. In another project, the brand guidelines called for a lowercase c in Mastercard—Claude complied for the first five instances and then casually broke the rule on the sixth. ChatGPT is often apologetic when I correct it. Claude, on the other hand, is a self-righteous and patronizing ass who insists that I’m wrong. When AI doesn’t know, AI invents ChatGPT and Claude sound plausible and knowledgeable. That’s the worst part. No, wait, the worst part is that they automatically fill in the gaps instead of flagging uncertainty. Both Claude and ChatGPT confidently invent French titles for reports that were never published in French, come up with acronyms that sound authoritative but were created two seconds ago, make up supposedly standard terminology and vehemently defend their choices when challenged, often adding another layer of hallucination or pointing to a dubious source. No, an anonymous user on Reddit isn’t proof you’re right, Claude. You can’t trust the output. It could be right. It could be terribly wrong. Flip a coin. Cultural bias is showing AI can translate, but AI can’t localize. Put plainly, these are machines, so they can’t talk to humans the way we—yes, I’m human—do. They use semantic equivalences, but they lack the millions of tiny judgment calls involved in figuring out what a normal human would actually say in this exact context. Marketing copy telling French people to snack on PB&J sandwiches is tone deaf because it’s just not a thing on this side of the Atlantic, no matter what the source document says. A professional translator will suggest a croissant and a coffee, which makes a lot more sense in context. One of my clients sells lovely “winter” jackets. Yeah, European winters… so Canadian fall. Without a friendly human shivering in fall and freezing in winter to adapt the message, the marketing copy falls flat. AI isn’t magic. Large language models were trained on a huge amount of English-language internet text, and that content is disproportionately shaped by US and broader North American culture. They choose terms based on what appears statistically common in this corner of the world—too bad if you’re living another life elsewhere. AI writing is soulless and gimmicky I know, everybody says that and it’s subjective. Hell, apparently loving em dashes is enough to get you accused of being AI now https://correresmidestino.com/my-blog-hit-the-front-page-of-hacker-news-and-it-was-weird/ . It’s hard to define AI writing. Text generated by models like ChatGPT or Claude can be fluent, coherent, persuasive, funny, or stylistically sophisticated. It can fool you for a paragraph or two and, granted, a set of instructions written or translated by yours truly can sound just like AI. But once you see enough generated text, you can’t unsee it. It’s very generic, for instance, and it reuses the same patterns over and over again—perfectly symmetrical sentences, structures with “including XYZ” repeated throughout, unnecessary conclusions wrapping everything up. It doesn’t create emotions but names them. It doesn’t give human examples and falls back on one-experience-fits-all adjectives: “bustling streets, vibrant cafés, and charming architecture”. It says all the right things, and you forget them instantly. Let me put it this way. I remember and can still quote reports I translated or edited three years ago. I can’t remember what I post-edited an hour ago. I “fixed” the machine output, but I didn’t leave my human print on it. I’m not paid for that. But when I have to check terminology, detect hallucinations, restore consistency, spot cultural nonsense, undo literal translation, fix tone, and sometimes compare every sentence against the source because the model tends to skip what it doesn’t know how to translate—was half the work actually done? Right. You don’t care about my economics. Fair enough. Let’s put it differently. If a human isn’t there to parent the machine, do you think your document looks professional, creative and publishable? Trusting AI feels like the right 2026 move. Everybody is talking about it. You’ll be more efficient, you’ll save time, energy and money But I keep coming back to the same uncomfortable feeling: I’m losing control, and so are you. It’s producing the words. I’m checking them, correcting them, sometimes rewriting half of them—but they didn’t come from me, or from you, in the first place. At some point, “saving time” starts looking a lot like handing over your voice and hoping the machine got it right. So how obsolete am I on a scale of 10 after twelve months of this? More than I’d like, less than my clients seem to think.