Redefining efficiency in post-editing of LLM-generated translation: why expert translators take more time

Psychology ~3 min read ▶ 3 min 54 s listen
Tom Tom Carol Carol

In short

Experts spend more time post-editing machine-generated translations than initially expected, challenging claims of increased efficiency. This arises because humans must correct errors and ensure accuracy, often rebuilding entire sections when the AI output is nonsensical, rather than simply cleaning up typos.

One idea raised: Cross-check every proper noun against the original source before moving on to ensure accuracy and prevent costly errors.

0:0003:54

Therapy is a place to put down what you have been carrying.

Lines worth sharing

Each one is a ready picture: share the link and it appears as a card in the feed, or save the image. Cards marked ▶ Video also come as a short clip with music: tap to watch, then save or share it.

TomWe were told the machines would speed things up, yet the experts are actually spending longer on every single line of LLM-generated text they have to fix through post-editing. Carol, how does that square with everything they sold us?

CarolThe label “post-editing” hides the reality that we’re doing full translation under false pretenses. Think about a freelance translator getting paid half-rate because the client says the machine did 80% of the work, while she still fixes every error manually. That’s not efficiency, that’s wage theft dressed up in tech jargon.

TomThat word "efficiency" right there in the title changes everything. This Illinois and Kent State study measures effort regulation, proving the time spent proves human judgment is irreplaceable, not just a slow bottleneck. Calling it waste ignores why the expert has to step in.

CarolIrreplaceable isn’t the same thing as efficient, Tom.

TomEveryone reaches for “collaboration,” but that word erases the fact that humans are carrying the whole load. Carol, imagine if we said dentists collaborate with bacteria—sounds nice until it doesn't stop the decay. Where does that line draw when the LLM spits out total nonsense?

CarolNo, the line draws where meaning breaks down completely. A good post-edit keeps the original intent intact, like "the report due Friday," whereas a bad one scrambles dates and names into gibberish. You can spot which category you're in within seconds.

TomSpotting the bad ones in seconds only works when the machine gets close enough to matter. Picture a guy who always skimmed his draft before sending it to his editor. One day he skipped it entirely because the output looked fine. His editor caught three scrambled facts and sent it back without a comment. That silence hit harder than any note ever could.

CarolYour editor catching those errors assumes the machine got close enough to translate at all. If the LLM outputs complete gibberish, you aren't editing anymore—you're starting from scratch, making the extra time pure labor, not skill.

TomStarting from scratch still requires knowing what the source text actually meant, which is the core expertise the model lacks. Carol, calling that raw labor misses that the translator’s judgment is the only thing keeping the final product accurate. Without that check, the client receives garbage regardless of how much time went into the process.

CarolKnowing the source means nothing if the output needs rebuilding anyway.

TomRebuilding from zero kills the value of the original edit. But the expert’s job shifts to verifying the translation against the source, not trusting the AI’s guesswork. "Check every date and name against the original." When that editor caught three scrambled facts in the draft, wasn’t he doing exactly that verification step?

CarolYes, he was verifying, but only after the damage was done. I’d keep the rule simple: “Cross-check every proper noun against the source before moving on.” Verification isn’t skill if it’s just cleanup.

TomCleanup catches typos, but it misses the tone shift that ruins a deal. The label matters here: call it “judgment” and we see why experts stay paid. Call it “cleanup” and the industry stops valuing the mind behind the work. That’s why the extra time proves human judgment is irreplaceable.

CarolYou’re still focused on the label when the question is about efficiency. Expert translators take longer because they calibrate their effort based on how risky each segment is, not just to prove their worth. Try saying exactly this to your client: "The additional time ensures accuracy where meaning is fragile, preventing errors that cheap speed cannot fix." That makes the cost about reliability, not ego.

Read next

Dialogues on the same topic

Other dialogues in Psychology

See all in this world →