A PM sent me a marketing tagline last week with the note "just post-edit the MT, should be quick." The MT output was grammatically perfect. It was also useless. Because there was no way to check it against the English — the English was a wordplay on a brand name that means nothing in Turkmen. The machine did its job flawlessly and produced garbage, and the whole "post-edit it" instruction fell apart on contact.
That's the thing nobody wants to hear: transcreation isn't a harder version of translation. It's a different job with a different deliverable, and the entire economics of MT post-editing assume a deliverable that transcreation doesn't have.
Post-editing needs a source. Transcreation burns it.
Here's how MTPE actually works and why agencies love it. The machine gives you a draft, you compare it against the source, you fix drift, you ship. The vendor reports say it cuts cost 30 to 50 percent while keeping human-level accuracy, and that's roughly true for the right content — a user manual, a product spec, a support article. Over 70 percent of freelancers in Europe are on MT tools now. The market's sitting around 700 million dollars this year and climbing toward a billion by 2032. That model runs on one assumption: there's a correct answer sitting in the source, and the human's job is to enforce fidelity to it.
Transcreation deletes that assumption on purpose. The source stops being the thing you're faithful to. The brief becomes the thing you're faithful to — the tone, the audience, what the client wants the reader to feel and do. The English tagline is a reference, not a target. Sometimes I throw it out entirely and write something that lands in Turkmen the way the original landed in English, and the two texts share nothing but intent.
So when you ask me to "post-edit" that, you're asking me to check a translation against a source I was explicitly told to abandon. There's nothing to compare. No fidelity to score. COMET can't touch it, my QA panel can't flag it, and the MT draft is worse than a blank page because now I have to un-see a wrong answer before I can find a right one.
"More creative" AI is the wrong pitch
The AI dubbing crowd has figured out that "cheaper" doesn't sell to people who care about quality, so the new line is "more creative." One localization vendor calls its freelancers Cultural Ambassadors, and its CEO argues — with a straight face — that AI dubbing can be more creative than the traditional kind, because now you get to decide whether to clone the original voice, how to handle accents, all these new knobs.
Those are real choices. They're just not where the creativity lives. Deciding whether to clone a voice is a production decision, like picking a font. The creative act in localization is deciding what the line should mean in the new language, and that decision happens before any voice speaks a word. You can give me infinite voice-cloning knobs and I still have to solve the same problem I've solved for 15 years: this joke doesn't exist in Turkmen, so what do I put there instead.
And the market keeps proving the audio isn't the hard part. Amazon started testing AI dubbing on twelve licensed titles back in March, explicitly hybrid, human experts in the loop for QC — and early viewers still called it robotic. YouTube's auto-dubs, out since last December, got described as "text-to-speech level stuff" that you can't even turn off. The tech isn't failing at pronunciation. It's failing at judgment, which is the only part that was ever worth paying for.
Where this leaves subtitling, and me
Subtitles are still the default, and part of that is just law — the CVAA in the US, the AVMSD in the EU, caption rules in the UK, Canada, Australia. Broadcasters and VOD platforms have to caption, so subtitles get made whether or not anyone's excited about AI dubbing. And plenty of markets prefer reading anyway; the US, UK, much of Asia lean subs, while Germany, Brazil and Mexico lean dubbed. Subtitling isn't going anywhere, and it's the format where transcreation and hard constraint collide hardest — you're transcreating a joke and fitting it in 42 characters at a reading speed a human can actually follow.
What I'd tell a PM: stop putting transcreation and MTPE on the same rate sheet with the same words. They are not two points on one scale. When you send me creative copy, don't send me the MT draft and don't ask me to "post-edit" — send me the brief, the brand voice, the reference, and pay for the hours, not the words. I'm not correcting a machine's homework. I'm writing something the machine had no way to attempt, because there was never a right answer to converge on.
The irony is that the more MT eats the fidelity work, the more the creative work stands alone and undisguised. There's no source to hide behind. Just a brief, a blank line, and someone who has to decide.