A client asked me last month to certify a translation I hadn't done. The draft came from their in-house tool — some fine-tuned model, they said, trained on their own contracts. All I had to do was stamp it. Put my name on it. Half a day's work, easy money.
I read it. Clause 4 referred to a definition in clause 2 that the machine had quietly translated two different ways, three pages apart. Same defined term. Two renderings. In a supply agreement that's not a typo — it's an argument waiting to happen. I told them I'd have to retranslate half of it before I'd sign anything. They paused. Because they didn't want the translation. They wanted the signature.
That's the whole business right now, if you look at it straight.
The signature is the thing you're actually selling
MTPE has swallowed the middle of the market. The GTS survey this year put frequent-plus-occasional post-editing at roughly 88% of freelancers, and Nimdzi has adoption jumping from 26% in 2022 to 46% by late 2024. Commodity words are collapsing toward $0.05–0.15 a word while full human translation sits at $0.15–0.30. If your work is interchangeable, the price already told you so.
But certified legal work and high-stakes technical work didn't move the same way. And it's not because the machine can't produce a plausible sentence. It's because somebody has to be liable for the sentence, and a model can't be.
When I certify a Turkmen translation of a drilling contract, I'm not just claiming the words are accurate. I'm putting a name that can be sued behind them. That's a legal act, not a linguistic one. The Swiss legal-translation study that made the rounds this year found frontier LLMs still hallucinating non-existent terms and botching the harder terminology — which matters — but honestly, even if they nailed every word, you'd still need a human who's accountable when clause 4 contradicts clause 2. The signature is the deliverable. The translation is just what you check before you give it.
The AI Act made this a compliance problem, not a preference
For a while, "a human should sign off" was something we said because we cared about quality. Now there's regulation lining up behind it.
The EU AI Act came into force in August 2024 and is phasing in on a schedule that matters to anyone selling into Europe. Prohibitions on certain systems kicked in February 2025. General-purpose AI obligations landed August 2025. High-risk system obligations arrive August 2026. And the tool doesn't get a pass because it was built in San Francisco or Shenzhen — if you offer the service in the EU, you're in scope.
Meanwhile the liability question sits there unresolved. If a mistranslation causes real damage, you don't get to shrug and point at the model vendor. Try holding an AI maker harmless in court and watch how far that gets you. The practical answer the industry keeps landing on is boring and correct: professional post-editing or certification, a named human on the hook. ISO 17100 processes exist precisely so there's a documented chain of who did what. That chain has a person at the end of it. It has to.
Even the Court of Justice of the EU is publicly wrestling with how much AI belongs in its own legal translation, workload and budget pressure notwithstanding. When the institution that writes the rules can't figure out where the line goes, don't expect your agency's risk department to be relaxed about it.
Oil and gas hasn't gotten the memo about cutting corners, thankfully
Energy work is where I see the human-gate held hardest, and for good reason. A safety procedure that's wrong doesn't produce a bad review. It produces an incident. The providers I work with in this space still run legal linguists and native subject experts through certified processes, and nobody's in a hurry to change that.
What is changing is the shape of the demand. Renewables terminology is flooding in — hydrogen, carbon capture, offshore wind — and a lot of it doesn't have settled Turkmen equivalents yet. I'm coining, checking, arguing with myself. There's more VR and eLearning safety training to localize. More real-time interpretation for field briefings where three nationalities stand around one wellhead. The machine helps me move faster through the draft. It does not tell me whether "managed pressure drilling" has a term of art the client's engineers already use on site. That I still have to know, or find out by asking a human who's been on the rig.
So here's what I'd tell a PM staring at a fine-tuned model and a tight budget: run the draft through it, sure. But be honest about what you're then buying from me. You're not buying words per hour. You're buying a name that will stand behind those words when a lawyer or a regulator comes asking who's responsible.
That name has always cost more than the typing. It just took a machine that types for free to make everybody notice.