A California court ran the numbers and told on itself. Orange County Superior Court customized a machine-translation engine for legal terminology, put certified translators on quality control, and reported that its Spanish output was 80% usable as-is. That figure got passed around as a success. Read it again. Twenty percent of a court document was not usable, and nobody in the story could tell you which twenty percent without a human reading every line.
That's the number I keep coming back to. Not because 80% is bad — for a first pass on legal text it's genuinely good. But because the missing 20% doesn't announce itself. It hides in a defined term rendered two ways, a negation dropped, a party name that drifted from "the Lessor" to something softer. The 80% is the part you don't get paid for. The 20% is the job.
What the certificate actually is
Here's the thing agencies and clients keep half-understanding. When I certify a legal translation, I'm not certifying that the words are pretty. I'm putting my name on a statement that the target says what the source says, and I'm making myself the person you come after if it doesn't. That's a legal act. It has an address, a signature, a human who can be deposed.
AI can't do that. Not as a limitation of the current models — as a category problem. Lionbridge put it plainly earlier this year: for now, courts require a human translator or a language company to formally certify the work, because AI can't be held accountable. A document produced through a fully AI-driven workflow, no human involvement, cannot be certified for legal proceedings. Full stop. You can run the fanciest multi-engine setup, compare outputs, pick the version with strongest agreement — and you still have nothing you can file, because there's no one standing behind it.
So the workflow that's winning isn't AI-instead-of-me. It's AI-draft, human-certifies. The LLM generates a draft, prepped with the client's glossary, told to preserve party names and defined terms. Then a certified human reads it, fixes the 20%, and signs. The signature is the product. Always was. The draft is just how fast I get to the point where I can honestly sign.
And I like this arrangement, mostly. It moves my time off transcription and onto judgment, which is what I'm actually worth. What I don't like is when a client pays me to certify and quietly expects me to rubber-stamp a machine draft I never fully read. That's not a discount. That's fraud with a nicer margin.
The courts are already drawing the line — at verification, not tools
People in this business love to argue about whether using AI is "allowed." The courts have mostly stopped caring about that question. What they're punishing is not checking.
The Ninth Circuit made this explicit in a June disciplinary order this year — the LNU v. Blanche mess. The court said, in so many words, it wasn't sanctioning the lawyers for using generative AI. It sanctioned them for signing and filing work they hadn't independently verified. That principle runs straight through to translated evidence. Nobody's going to hang you for opening a chat window. They'll hang you for certifying something you didn't read.
The National Center for State Courts landed in the same place with its 2025 guidance: MT is fine for administrative purposes or a preliminary look, but testimony and high-stakes documents get full human translation and review. Their operational checklist reads like something I'd have written after a bad week — train the engine on court-specific glossaries and certified forms, roll it out on narrow use cases first, keep secure internal AI behind a firewall, get confidentiality agreements from any outside reviewer, and disclose to all parties when AI translation was used.
That disclosure point matters more than it looks. If you're pasting a client's contract into a public AI tool, you may be handing away privilege. Courts have started treating documents uploaded to public generative platforms as potentially outside attorney-client protection. So the "free" workflow — grab the file, paste it into whatever's open in the browser, clean it up — isn't a shortcut. It's a way to blow up your client's case before you've translated the first clause.
Where I actually earn it
A 2026 study on Arabic–English legal translation found the pattern anyone who does this work already lives: humans sometimes trip on specialized terminology, and AI reliably chokes on culturally specific concepts with no clean equivalent in the other language. Now put that in Turkmen. Half my legal work sits on top of Russian-era legal structures and Soviet-inherited terminology, mapped onto contracts drafted to English or Spanish templates. There is no engine trained on that overlap. There's barely a glossary. The concept without a direct equivalent isn't the exception in my pairs — it's Tuesday.
So when I certify a Turkmen legal translation, I'm not competing with the machine on speed. I'm the accountability the machine structurally can't provide, sitting exactly where the 20% lives. That's not a defensive crouch. It's the most durable thing I sell.
Give me the AI draft. Genuinely, I'll take it — it saves me the boring 80%. Just don't confuse the part that got faster with the part you're paying for. You're paying for a name at the bottom that a court can find.