Can a Translation Agency Trust AI? Check the Source

Can a Translation Agency Trust AI? Check the Source

Last Updated: October 2026

Machine translation post-editing is the work of reading a machine’s draft against the source text and fixing what it changed. The job is not to make it read well. It already does, and that is the trap. A line can be smooth, warm and sure of itself in the new language while saying what the source never said, and a reader who checks for style will wave it through.

AI Smart Ventures has guided growing businesses, language teams among them, through AI adoption inside work that clients pay for. One pattern repeats here. Machine output stopped sounding odd a few years back, so the check quietly turned into a read for style, and nobody wrote down the day it changed.

Clumsy output used to guard you. It gave itself away, and even a tired reader caught it. Smooth output gives no warning, so a wrong date, a missing “not”, or a term that got dropped ships looking like your best work. Then the client finds it, or their regulator does, and what you lose is your name for being right.

Key Takeaways

  1. Smooth drafts get edited less: in a June 2026 EAMT study, students who worked from the smoother machine draft made fewer edits and left more of its errors in place.
  2. Read against the source, not the page: ISO 5060:2024 scores a human check by error type and penalty points, with sampling, not by how the text reads.
  3. Error flags do not save the check: trained translators given machine-made error marks and fix tips gained nothing in speed or quality in one 2026 trial.
  4. Safety text is not light-edit work: a 2025 JAMA Network Open study found 92% of Somali discharge sections carried clinically impactful errors from a machine model, against 13% by hand.
  5. The rulebook is being redone now: ISO 18587:2017 is marked for revision, and the draft that takes its place is retitled “Post-editing of non-human translation output”.

Put those together and the check changes shape. You are not grading prose any more. You are holding two files side by side, which is slower, duller, and the one part of the job a client cannot do alone. Brief for that and practical AI stops being a risk you carry in silence.

What does machine translation post editing mean?

Post-editing means a person reads the machine’s draft beside the source file, fixes what the machine got wrong, and signs the result off as fit for use. It is not proofreading, which never looks at the source at all. The machine has already handed you something readable, so the question left is narrower and much harder: does this say what the source says?

Two depths get sold. A light edit fixes only what would mislead a reader and leaves the style alone. A full edit also brings tone and word choice up to the mark for print. Both are promises to the client, so write down which one each job bought, and keep that line in the file. Where you will not offer a depth at all, say so in the quote, because a standard the client cannot see is not a standard.

Why do fluent errors slip past a reviewer?

Because smooth text reads as right, and the eye has nothing to catch on. BlAInded by Fluency, shown at the 26th EAMT conference in Tilburg in June 2026, put 18 British editorials into French with two prompt styles. Students who worked from the smoother version made fewer edits, left more machine errors in place, and landed fewer good fixes than those given a version that stayed closer to the source.

The habit is older than the models. In a 2018 pilot, Martindale and Carpuat showed readers text that read badly but kept the meaning, and text that read well but lost it. People griped loudly about the clumsy lines and shrugged at the wrong ones. Trust tracked how it read, not whether it was right, and that is just the bias a polished draft now walks into your office carrying.

How do you review against the source, not the page?

Open both files and work line by line in source order, checking sense before style. Numbers, names, dates, units, a “not” and an “if” carry most of the risk, so check those first and on their own. ISO 5060:2024 sets out this kind of check: errors sorted by type, weighted with penalty points, scored by a trained person, with sampling when the volume is too big to read in full.

Two checks compared on one translated file, showing which error types a line-by-line read against the source catches and which ones a read for tone alone leaves in place

Order the passes so the quick habits come first. One reader sweeps terms and figures with the source open. A second reads the new text alone for tone. The two never merge into one pass. Then keep a short log of what the machine got wrong on each job, sorted by type, because the same engine fails the same way twice. That log is workflow optimization you can read in a minute.

Which content is safe for a light edit?

Low-stakes text where a slip gets caught further down the line: help pages, support macros, team notes, product listings, user reviews and early drafts for a bilingual contact at the client. The test is not length or topic. It is the catch. Ask who reads this next, and whether that person would spot a sense error. If the answer is nobody, it is not light-edit work, whatever kind of text it is.

The other list is short and firm: contracts, consent and safety wording, drug and care instructions, filings, accounts, and any text a reader will act on alone. Language pair weighs as much as topic. The JAMA Network Open study of 148 sections from 34 children’s discharge instruction sets rated a machine model close to human work on two counts in Spanish, while 92% of the Somali sections carried clinically impactful errors against 13% by hand.

What are the guidelines for post-editing work?

ISO 18587:2017 is the standard in force. It sets out what full human post-editing of machine output must involve, and what a post-editor has to be able to do. It is also on its way out. ISO marks it for revision, and the draft that replaces it went out on 11 September 2026 as “Post-editing of non-human translation output”, with October 2026 listed as the date it is due.

Read the new title rather than waiting on the final text. Dropping “machine translation” for “non-human translation output” widens the standard to whatever made the draft, which is where firms already are. Write your own house rules now: depth by content type, what gets a second reader, and which text never goes near a model. Where a slip could cause harm, your counsel and the client’s own rules set that line, not your workflow.

What should you tell clients about how text was made?

Tell them in writing, before the job starts, in the quote or the statement of work. Four facts cover it: whether a machine made the first draft, how deeply a person edited it, who signed it off, and where their files were handled. None of that needs a tool name, which helps, because the tool will change and the clause should not.

Said early, this reads as a service note. Said after a complaint, it reads as an admission, and that gap is the whole point. Vendors selling automation as AI will tell you the clause is not needed, since quality is handled; a client who later learns a machine drafted their contract wording will not agree. Keep one short paragraph in every bid, name the depth you sold, and let your client terms carry the rest.

How should you brief and judge your reviewers?

Brief them on error types, not on speed, and judge them the same way. Give each one the content type, the depth bought, the word list, and the handful of error types that bite most on this account. Then check their work by sampling against the source, not by counting lines cleared per hour. A reader judged on throughput will quietly reward themselves for trusting the draft.

Do not expect a tool to carry the brief for you. In a 2026 trial, trained English to Dutch translators worked with machine-made error marks and fix tips; neither lifted quality or speed above a plain edit, though they liked the tips. Over-editing costs you too. In a study of 20 students’ edits, 21.2% of the 132 flagged problems came from changing text that was already fine, against 50.8% that came from the machine.

AI Smart Ventures offers AI Consulting for firms that want this check written down before machine output reaches a client file.

Frequently Asked Questions

Is post-editing a machine draft worth the effort?

Yes for bulk, low-stakes text, and not always for the rest. It buys time, not freedom from checking, and the saving shrinks as the stakes rise. On safety or regulated work the check takes about as long as doing the job by hand, because the draft has to be read twice against the source. Decide by content type, then quote the work on the depth you really sell.

What are the three main kinds of machine translation?

Rule-based, statistical and neural. Rule-based tools apply grammar rules and word lists written by hand. Statistical tools pick the likeliest wording from large piles of past translation. Neural tools learn from the same stock with deep networks, and they read the most smoothly of the three. Most work today, general-purpose models included, sits in that third group, which is why current errors read so well.

Does a fluency check count as a quality check?

No. One asks whether the new text reads well. The other asks whether it says what the source said. They come apart just where it matters, because a sure-footed line can carry a lost “not” or a wrong figure and still read beautifully. Keep them as two passes with two briefs, and never let the style read stand in for the read against the source.

What if nobody in-house reads the target language?

Then buy that reading rather than skip it. A check on the new text alone cannot find a sense error, and a back-translation can hide one, since a second machine pass may quietly repair the damage on the way home. Keep a vetted reader for every pair you sell. If you cannot staff a pair to the depth a job needs, turn that pair down and say why.

Can automatic quality scores replace a human reviewer?

No, and the 2026 evidence on tool-assisted checking is sobering. ISO 5060:2024 covers human review of output only, and a 2026 trial of machine-made error marks with trained translators found no gain in quality or speed over a plain edit. Use scores to sort which files get a full read, then have a person read those against the source anyway.

Do clients have to be told a draft came from a machine?

Your client terms and your own counsel settle that, not a vendor’s sales page. The business answer is simpler: tell them, in the quote, before the work starts. Say whether a machine made the first draft, how deeply it was edited, who approved it, and where their files were handled. Buyers who find out later treat it as something you hid.

Does post-editing need a different skill than translating?

Yes, and the standard treats it that way. ISO 18587 sets out what a post-editor must be able to do on its own terms, because the work is a hunt: read the source, guess where this kind of draft fails, and leave alone what is already right. Strong translators are not always strong post-editors, and the training that closes the gap is error-spotting drill, not general AI literacy.

How should an agency start reviewing against the source?

Pick one account this week, list the kinds of text it sends you, and mark each one light edit, full edit or human only. Then run next week’s jobs through two passes: terms and figures against the source first, tone second. Most firms can test that in two weeks with no new tools. Schedule a consultation to design the check and brief your readers.

Executive Summary

A firm can trust machine output as a draft and never as a result. The failure that matters is not clumsy text, which gives itself away, but smooth text that says what the source did not, and 2026 work shows the smoothest drafts get edited least. The fix is a matter of order, not of tools. Read against the source. Check figures, names and a stray “not” apart from tone. Set the depth by content type, keep legal, care and safety wording out of light editing, and tell clients in writing how their text was made.

What Should You Do Next?

This week, take one live account and mark every kind of text it sends you as light edit, full edit or human only. Then rebuild your brief around error types instead of turnaround, and add one short paragraph to your next bid saying how a draft gets made and how deeply it gets checked.

AI Smart Ventures offers AI Consulting for growing businesses putting AI inside client work without losing the accuracy clients buy, with capability building for the people who check. Schedule a consultation to design the check for one account.

People Also Read

About the Author

Nicole A. Donnelly is the Founder of AI Smart Ventures and an AI Adoption Specialist with 20 years of experience as a founder and CEO and over a decade leading AI adoption initiatives. She helps businesses integrate artificial intelligence with clarity and confidence, driving innovation and sustainable growth. Nicole has trained over 20,217 professionals in Applied AI, delivered 624 workshops, and worked with close to 1,000 organizations across diverse industries.

Expertise: AI Transformation, AI Strategy, AI Implementation, AI Adoption, Applied AI, Marketing, Business Operations

Connect: LinkedIn | Website

Disclaimer: This content is for informational purposes only and does not constitute professional business or technology advice. Results vary based on industry, existing systems and implementation commitment. Contact AI Smart Ventures for a consultation regarding your specific situation.