We've Been Asking the Wrong Question About AI Translation
Where is the line for humans in AI translation? We ask it at every conference and rarely answer it.
But last week on a MultiLingual Magazine panel, the question finally moved... just not in the direction I expected.
We set out to discuss language technology in the AI era, but instead, we ended up talking about agents going rogue in testing environments, like the Hugging Face attack. The audience's questions changed:
How can we beat the speed of an AI's decision in high-risk situations?
AI leaders are calling for slower development for the good of humankind. Is that realistic?
Should we move from human in the loop to human in control?
Is there a real risk it could k*ll us?
Luana Y. Ferreira, Eddie Arrieta, Joachim Lepine and I approached these questions from an unexpected angle: humanity. Not accuracy scores, but what happens to the public when we use AI in high-risk situations, translation included. Translation wasn’t the focus, anymore. Translation was a natural risk area for all of society.
That one shift in framing changed everything.
We stopped measuring AI translation risk in lost revenue, lost jobs, or the vague "harm" many of us cling to to justify our positions. We measured it against the greater good. That's the larger conversation translation belongs in.
From there, the discussion split in two: should we even still be debating human in the loop anymore in the age of agents, and where can we realistically put up a fence around different types of content.
Human in control vs. human in the loop
Human in the loop is the argument everyone uses, usually means a person checks or approves a step before the AI continues. Human in control means we decide what the system may do, and under what conditions.
Human in control is the better goal for 2026, because... well, a person in the loop doesn't automatically have real power.
The distinction matters as agents start running their own loops: planning, using tools, checking results and acting without waiting for us. We shouldn't need to intervene in every loop, but we should define their permissions and which actions need our approval. Many large companies already see AI implementation this way. They want to see what agents are doing, adjust boundaries and revoke access when needed.
The headlines make it look like AI is going rogue. But a rogue agent is a control failure: humans didn't set proper boundaries. They skipped that step. That's where our resources should go now.
However, setting boundaries starts with knowing which content can hurt people. That's why I believe translation workflows should be designed around risk, not a single quality standard.
The Three Content Tiers For the AI Era
Bridget Hylak and I 1.5 years ago started to view content (and globalization initiatives as a whole) in risk tiers. We wrote an article about it in Multilingual. I feel when we released it, no one was ready to accept viewing content in this way. Last week, I feel everyone saw how important it was.
Tier 1: Raw AI output. Fine-tuned models, QA-supported pipelines, high-speed delivery. Built for scale. Accuracy does not cause harm.
Tier 2: Creative. Marketing, media, localization, and anything where tone, rhythm and cultural nuance matter. Accuracy CAN cause harm.
Tier 3: Highly technical. Legal, medical, scientific and other high-precision content where errors could mean life or death. Accuracy WILL cause harm.
I walked into that panel expecting the usual debate about quality and jobs. I walked out realizing we'd been asking the wrong question. The risk of AI translation isn't only what it costs our industry. It's what it could cost the people relying on it.
Frame it that way, and the line becomes clearer: we draw it by risk, and we stay in control.
Localization family, your work starts now.
