Let’s not pretend this is a feel-good story. If you translate for a living, the last two years have been brutal, and the numbers say so. In a 2024 Society of Authors survey, 36% of literary translators had already lost work to generative AI and 43% said their income had dropped because of it. More than three-quarters expect it to keep eating their earnings. On the forums, the mood is rawer than that — people quitting, people refusing jobs, people watching agencies offer 30% of the old rate for “just cleaning up the AI.”
So this isn’t a piece telling you AI is your exciting new co-pilot. It’s the more useful version: how the translators who are still standing actually use ChatGPT — to become the person who fixes the machine instead of the person the machine replaced. There’s a real difference, and it mostly comes down to a few habits.
What’s actually happening to the work
Two things are true at once, and you have to hold both.
The first: the bottom has fallen out of commodity translation. Standard human rates used to sit around $0.15–0.30 per word. Post-editing machine output — MTPE, the job clients increasingly want — pays roughly $0.05–0.15, and some agencies push it lower. Adoption isn’t a trend anymore, it’s the baseline: per industry tracker Nimdzi, MTPE jumped from about 26% of language-service projects in 2022 to nearly 46% by 2024. If you do general business or marketing translation through agencies, you’ve felt every bit of this.
The second thing, the one the doom posts miss: the machine isn’t actually good enough to be left alone. Working translators who use it daily put ChatGPT’s accuracy at maybe 85–90% — fine until you remember that the last 10% is where the lawsuits, the brand disasters, and the “that’s not what the doctor meant” live. And here’s the quiet plot twist. Clients are now generating their own AI translations, shipping them, getting complaints, and coming back to humans to fix the mess. One translator on Reddit put it flatly: agencies sometimes “pay you more to fix AI slop than you’d have charged them for the translation.” There’s now a giant pool of bad machine translation out there. Someone has to clean it. That someone has leverage.
The four-step loop the adapters actually use
The translators who’ve made peace with this don’t translate from scratch and they don’t blindly accept machine output. They run a loop. Here it is.
1. Draft. Feed the source text to ChatGPT with a real brief, not “translate this.” Tell it the domain, the audience, the tone, and the target variant (“Latin American Spanish, marketing copy, informal tú”). You get a fast first pass.
2. Compare, don’t accept. Put source and target side by side — in a CAT tool like memoQ or Trados if you have one, or just two columns. Read for meaning first, polish second. This is the actual job now.
3. Post-edit with your judgment. Fix the terminology, the register, the cultural misses, the confident-sounding nonsense. This is where your years go to work. The machine gives you a runway; you do the landing.
4. Reprice around the new shape. This is the step people skip, and it’s the one that saves your income. You’re not selling keystrokes anymore. You’re selling accuracy, accountability, and the fact that your name stands behind it.
The adapters who are surviving don’t accept post-editing at any price. They quote 60–80% of their full rate for genuine MTPE and walk away from the 30% “just fix the AI” jobs — because those jobs are a trap where you do most of the work for a third of the money and carry the blame when the AI was wrong. Saying no to bad rates is part of the workflow, not a separate fight.
The five things to never auto-trust
ChatGPT is a confident liar. It will hand you a fluent, professional-looking translation that’s quietly wrong, and it never flags its own uncertainty. So these are the zones where you slow down every single time:
- Legal and contracts. A mistranslated clause can void an agreement or contradict a statute. Jurisdiction-specific terms are where machines fail hardest.
- Medical and regulated content. “Take this twice daily” vs. “twice weekly” isn’t an edit — it’s harm. Domain expertise isn’t optional here.
- Anything with voice or culture in it. Literary, marketing, humor, idiom. AI flattens these into bland, “correct” prose that loses the thing the client is actually paying for.
- Confidential material. Don’t paste a client’s unreleased or sensitive document into a consumer AI tool. Check your confidentiality terms first; anonymize when you can.
- Names, numbers, and dates. AI transposes and “corrects” these silently. Check every one.
What this means for you
If you’re a generalist doing agency work. This is the squeezed middle, and pretending otherwise helps no one. The move is to specialize into something the machine can’t touch, and to get firm about which rates you’ll accept. Volume at 5 cents a word is a road to burnout, not a business.
If you already specialize (legal, medical, literary, sworn, game localization). You’re the most insulated, and AI can genuinely speed your routine passes. Use it for the boilerplate; charge your full judgment rate for the parts that matter. Your niche is the moat.
If you’re in-house. You’re becoming the person who reviews and owns the AI output for your whole org. Lean into it — the “human accountable for the translation” role is more secure than the “person who types it” role ever was.
If you interpret. Live, high-stakes interpreting is far less substitutable. But AI is excellent for prep — building terminology lists, briefing yourself on a topic, running mock Q&A before a tough assignment. That’s leverage you control.
If you’re just starting out. Don’t skip the fundamentals because the machine seems to do them. The whole value of a post-editor is knowing when the fluent answer is wrong — and you can’t know that without the craft underneath.
What ChatGPT can’t fix for you
It can’t carry liability. It can’t read a room or a culture. It can’t keep your client’s secrets, and it can’t tell you when it’s wrong — which is exactly when you need to be told. It also can’t set your rates or refuse exploitative ones; that’s on you. And it won’t rebuild the commodity end of the market that’s gone. That part isn’t coming back. The work that’s left is the work that needs a human who knows things, and that’s a different, smaller, better-paid job than the one a lot of us trained for.
The bottom line
The translators losing this fight are the ones competing with the machine on speed and price. The ones winning are competing on the thing the machine doesn’t have: judgment you can put a name behind. Run the four-step loop, guard the five risk zones, and quote like a post-editor who’s worth it — because the demand for cleaning up bad AI is real and growing.
If you want to build the post-editing and glossary workflow properly, end to end, the AI for Translators course is built for exactly this moment. And the companion piece on building a translation glossary in five minutes covers the one habit that keeps a big project consistent.
Sources
- SoA survey: a third of translators losing work to AI — Society of Authors (2024)
- AI and the threat to translation jobs — CNN (January 23, 2026)
- ITI Spring 2024 Pulse Survey results — Institute of Translation & Interpreting
- Will AI and ChatGPT replace translators? — PoliLingua
- How AI translation is changing pricing: $0.01/word vs human translators — ChatsControl
