Do we still need human translators/localizers when real-time AI translation is near-instant and cheap?

The shift

Producing fluent, idiomatic target-language text from a source goes from scarce human hours to near-instant and near-free, across most language pairs and most everyday content. What stays scarce is being answerable when a wrong word carries real cost, adapting meaning across cultures where fluency isn't the hard part, and holding legal liability for what the translation says.

The axioms

• Turning a source text into fluent, accurate target-language prose takes trained human hours — bilingual production is scarce and slow.
• Only someone fluent in both languages can tell whether a translation is right — verification is gated by the same scarce skill as production.
• Adapting a message so it lands in another culture — idiom, tone, taboo, humour, brand voice — requires bicultural judgment, not just word-swapping.
• High-stakes text (legal, medical, safety, regulatory) needs a named human who is accountable if a word is wrong.
• Certified, sworn, and notarized translation carries legal weight because a liable human attests to it.
• The commodity tier — manuals, support tickets, product listings, routine correspondence — is the volume that funds the trade and trains the next generation.
• Ambiguity in a source ("does this clause mean X or Y?") gets resolved by a translator's judgment about intent and consequence.

Invalid axioms

1. Turning a source into fluent target-language prose takes trained human hours. Fluent production was the scarce good the whole trade priced against; a model now returns idiomatic text in seconds for near-zero cost across most major pairs. Habit-trap: orgs still budget, staff, and schedule routine translation as a per-word human line item, and still route "just translate this" internal content through a human queue when the bottleneck has moved to deciding what's worth checking.
2. The commodity tier is the volume that funds the trade and trains the next generation. Manuals, support tickets, product listings, and routine correspondence genuinely flipped — this tier is now largely machine work with light human touch, or none. Habit-trap: agencies and buyers still treat this volume as billable human capacity and as the apprenticeship path, when the entry-level rung it used to provide is quietly disappearing.
3. Language gaps gate who can participate in a real-time exchange. Live speech-to-speech translation is now good enough that a routine cross-language conversation no longer needs a human interpreter in the loop. Habit-trap: still inserting a human broker into low-stakes multilingual meetings and support calls out of habit rather than need. (This is the fastest-moving call here — live interpretation quality is improving quickly, so where "routine" ends and "high-stakes" begins keeps shifting.)

Unchanged axioms

1. High-stakes text needs a named human accountable for a wrong word. In legal, medical, safety, and regulatory contexts, a confidently-wrong translation carries real cost, and a model can't be liable for it. The scarce thing isn't fluency — it's someone who answers for the specific claim and can be held to it.
2. Certified, sworn, and notarized translation rests on a liable human attesting to it. Courts, immigration, and regulators require an accountable signatory, not a plausible output. The legal weight comes from who stands behind the text, which doesn't get cheaper when drafting does.
3. Cultural and creative localization is judgment, not word-swapping. Adapting brand voice, humour, taboo, register, and idiom so a message actually lands — transcreation, marketing, literary work — turns on bicultural taste and knowing what to change and what to leave. Fluent output was never the hard part here.
4. Resolving genuine source ambiguity is a judgment call about intent and consequence. When a clause could mean X or Y and the wrong reading is expensive, deciding which one — often by going back to the author's intent — is exactly the novel, high-stakes ambiguity a model papers over by picking a plausible reading.
5. Verifying a high-stakes translation still needs someone who can actually read both sides. Catching the costly mistranslation requires the same bilingual expertise that used to produce it — and it's now the scarce part of the job, because the cheap fluent output is the thing being checked.

New axioms

1. Verifying machine translation at volume is now the bottleneck. When a plausible translation of everything is free and instant, the scarce act moves from producing it to catching the one costly error buried in a large volume of good-enough output — and confidently-wrong is the default failure mode.
2. Distinguishing "good enough" from "safe" has no clear line. Machine output that reads perfectly fluent gives no signal about whether it's accurate or safe for the stakes at hand; the fluency that used to correlate with competence no longer does, so someone has to decide per-document which tier of scrutiny it needs.
3. Liability for an AI mistranslation in a high-stakes context is unassigned. When a machine-translated medical instruction, contract clause, or safety warning is wrong and someone relies on it, it's unclear who's answerable — the vendor, the deploying org, the human who lightly reviewed it, or nobody.
4. The commodity tier's collapse removes the training ground. If routine translation is where juniors built pair-specific fluency and judgment before graduating to high-stakes work, and that tier is now machine work, the pipeline that produces the experts we still need for the STILL-HOLDS work thins.

Where it breaks

Buyers still route high-stakes text through the same cheap, fast machine pipeline they now use for commodity content (invalid habit) at the exact moment nobody has assigned who verifies it or who's liable when it's wrong (new) — the mistranslation risk didn't go away, it moved into fluent output that looks trustworthy precisely because it reads clean. Separately, the trade is letting the commodity tier collapse to machines (correctly, it flipped) while still assuming a supply of experts will exist for the accountable, high-stakes work (still holds) — but that expertise was trained on the tier that's disappearing, and no one owns the new pipeline.

ai-axioms.vercel.app/axioms/marketing/comms-translation-ai-axioms