A client sent me a demo last month of an AI-dubbed training video. The tool had relipped the presenter so his mouth formed Turkmen-ish shapes. Clever. Unsettling. And wrong in a way nobody on their side could see, because the man was smiling while saying a sentence I'd rewritten to be an apology. The smile was timed to the English line. The meaning had moved. The face hadn't.
That gap is the whole story of where creative localization is right now.
The mouth is not the message
Everyone's excited about the synchronization — ElevenLabs, HeyGen, the dubbing pilots that let a voice "speak" twelve languages while the lips follow along. It's genuinely impressive engineering. And it solves the least interesting problem.
Lip-sync was never the hard part of dubbing. The hard part is that a joke in English runs seven words and the Turkmen equivalent runs thirteen, or none, because the pun doesn't exist. The hard part is register — whether this presenter is addressing a colleague or a superior, which in Turkmen changes the verb, the pronoun, the whole posture of the sentence. An AI dubbing tool will happily cram my thirteen words into the seven-word mouth slot by speeding up the audio or chopping a clause. The lips match. The sentence is now gibberish delivered at auctioneer speed.
So when an agency asks me to "just adapt the script for the dubbing tool," I ask the question that annoys them: what wins when they conflict — the timing, or the meaning? Because they will conflict. On every emotional line, they will conflict. And if nobody decides in advance, the tool decides by default, and it always picks timing. It has no other opinion.
Subtitles lie about how much room you have
I do a lot of subtitling into Turkmen, and the constraint that breaks people is reading speed, not translation quality. You can write a perfect line that nobody can read in the 2.3 seconds it's on screen. I've gone on about this before, but here's the part that connects to the creative work: transcreation and subtitling pull in opposite directions, and the same job often needs both.
Transcreation says: rebuild the idea so it lands. Add the cultural scaffolding. Explain the thing the source assumed everyone knew. Subtitling says: cut it to 42 characters a line, two lines, and get out. One wants to expand, the other wants to compress. When a marketing video comes in with a slogan that's a wordplay, and it needs subtitles, I'm being asked to be creative and brief in the same breath. Those are not the same skill. They're barely the same job.
The honest answer, which I give maybe once a quarter, is: pick one. If this slogan matters, give it a full creative pass and accept the subtitle will paraphrase the spoken line — viewers can live with a caption that captures the sense rather than the words. If timing is sacred, don't send me a pun. Send me the brief behind the pun, the thing it's trying to make the viewer feel, and let me build something that fits the box.
What I actually want in the brief
Here's the usable part. When creative AV work lands on my desk, the files are never the problem. SRT, VTT, whatever — fine. The missing thing is always intent.
Tell me who's speaking to whom. A CEO to shareholders and a technician to a trainee are different languages in Turkmen before I translate a single word. Tell me whether this is funny on purpose. You'd be amazed how often I can't tell, and getting that wrong turns a wry aside into a flat statement or, worse, an accidental insult. Tell me what cannot move — a product name, a legal disclaimer, a number on screen. And tell me whether a human is checking the output or whether this ships the moment the tool spits it out.
That last one changes everything. If a reviewer who reads Turkmen sees it before it goes live, I can take risks, leave notes, flag the two lines where I chose meaning over sync and let them decide. If nothing stands between my file and the published video, I translate defensively. I shave off the clever bits. I protect against the worst case because I can't trust the pipeline to catch it. The output is safer and duller, and the client paid for the dull because they skipped the review.
The demo with the smiling apology never shipped, by the way. Someone on the client side eventually watched it with the sound on and felt that something was off, even without a word of Turkmen. They couldn't name it. They just knew the man's face was arguing with his voice.
That instinct is the whole job. The tools can move the lips now. They still can't tell when the face is lying.