Post-Editing Machine Translation with AI

Machine translation post-editing (MTPE) is now most working translators' biggest revenue stream. The question is no longer whether to do it but how to do it well. Generative AI changes the game by turning post-editing from a manual sentence-by-sentence chore into a partially automated workflow with you as quality gatekeeper.

What You'll Learn

• The two flavors of MTPE: light and full
• How LLMs improve raw MT output before you ever see it
• Practical prompts for handling typical MT failures
• How to price MTPE work fairly given AI assistance

Light vs Full Post-Editing

Industry standards (ISO 18587) define two MTPE service levels:

• Light post-editing (LPE) — Make the MT output understandable and accurate. Errors of style or fluency are acceptable if meaning is intact. Typical use: internal communications, support content, large-volume product descriptions.
• Full post-editing (FPE) — Produce a target text indistinguishable from a quality human translation. Style, fluency, brand voice, and accuracy all matter.

The pricing gap between LPE and FPE is large, often 2x. The pricing gap between FPE and from-scratch translation is shrinking — sometimes to 20% — which is where the squeeze on freelancers comes from. Knowing the difference and pricing accurately is now a survival skill.

Where Raw MT Fails

Despite huge advances, neural MT engines still systematically fail at:

• Dropped negations — "ne pas" rendered as if "pas" wasn't there
• Number errors — especially when source uses words and target expects digits, or vice versa
• Polysemy — wrong sense of a word ("bank", "pitch", "set", "spring")
• Idioms — translated literally instead of idiomatically
• Long sentences — agreement, anaphora, and structure break apart
• Untranslatable cultural references — left untranslated or absurdly literal
• Inconsistent terminology — across the same document
• Tag and formatting damage — placeholders, HTML, code fences misplaced
• Low-resource language pairs — quality collapses outside major European/Asian pairs

Your post-editing pass is what fixes these. AI as a polishing layer between MT and you can fix some, but not all.

LLM-Polished MT: A New Two-Step

The newer workflow is:

1. MT engine produces a raw draft.
2. LLM polishes the draft, applying your glossary and style guide.
3. You post-edit the polished draft.

The LLM polish prompt:

"Below is a raw MT draft (DE → EN). Apply the following project rules and produce a polished draft:

1. Use UK English spelling and punctuation
2. Apply this glossary: [paste glossary]
3. Use 'we' for company self-reference (not 'the company')
4. Active voice preferred over passive
5. Sentences over 30 words must be split
6. Preserve every number, date, and proper noun exactly as in the source
7. Do not add or remove information from the MT draft

Produce only the polished draft, no commentary. Then in a separate block, list any segments where you were uncertain."

The "uncertain segments" output is where you focus your human attention.

Specialized Post-Editing Prompts

Negation check

"Compare source and target. Find any segment where a negation (not, no, never, without, neither, nor, n't) appears in the source and is missing or altered in the target. Or vice versa. List with segment numbers."

Number and date verification

"List every number (including ordinals, percentages, decimals) and every date in the source. For each, confirm whether it appears identically in the target. Flag mismatches."

Polysemy resolution

"The English source contains the word 'bank' in segments 5, 14, and 22. Determine from context which sense is intended in each segment (financial institution, riverside, slope, etc.) and confirm the German target reflects the correct sense. If not, propose a correction."

Idiom flagging

"Identify any English idiom in the source that has been translated literally into the Japanese target. Suggest a more idiomatic Japanese rendering for each."

Tag and placeholder integrity

"Each segment in the source contains XML tags and {placeholder} variables. Confirm that every source tag and placeholder appears in the corresponding target segment, in a sensible position. Flag any segment where tags are missing, duplicated, or misplaced."

Pricing MTPE Honestly

Three pricing models you'll encounter:

1. Per-word with MT discount — Standard rate × percentage. Industry-typical 60–70% of full rate for FPE, 35–50% for LPE.
2. Hourly with MTPE-specific rate — Common for messy or low-quality MT.
3. Bundled per-word rate — Single rate that includes MT plus polish plus post-edit, often the lowest. Avoid this unless the MT is genuinely good for your pair.

To price a job fairly, sample-test the MT first.

"I am about to quote a 12,000-word EN → ES post-editing project. Below is a 400-word sample of the source plus the agency's MT draft. Assess: (1) overall MT quality (1–5), (2) the percentage of segments that are usable as-is, lightly editable, or need full retranslation, (3) the recurring error patterns, (4) my realistic hourly throughput on this. Then suggest a fair per-word rate."

You now negotiate from data, not gut feel.

The Ethics of Disclosure

You owe your client honesty about your process. Specifically:

• If the client asked for "human translation", you should not be running their text through MT or LLM at all unless they've consented.
• If the client asked for "MT post-editing", you can use any post-editing tools they have not specifically forbidden — but tell them which.
• Confidentiality clauses matter. Some clients' contracts forbid sending content to third-party AI providers, period.

In 2025 several agencies began requiring translators to disclose all AI tools used in production. Expect this to spread. Document your workflow now so disclosure is straightforward later.

When NOT to Post-Edit MT

Some content should still be translated from scratch, no MT involvement:

• Sworn/certified translations (legal requirement in many jurisdictions)
• Literary work, poetry, lyrics
• Brand-defining marketing copy
• Highly sensitive medical or legal documents where MT error is unacceptable
• Languages with weak MT support
• When the client explicitly forbids MT

For these, MT is not just suboptimal — it's an ethical and sometimes legal problem.

A Defensible MTPE Workflow

For a 10,000-word FPE project:

1. Sample-test the MT. 400 words. Assess quality. Confirm or renegotiate price.
2. Pre-translation brief. As in lesson 7: domain, terms, structure.
3. Build glossary. As in lesson 5: AI-extracted, human-verified.
4. LLM-polish the full MT. Apply glossary + style + locale rules.
5. Run automated QA prompts. Negations, numbers, idioms, tags.
6. Human post-edit. This is where you actually earn your fee. Skim every segment, fix every flagged issue, rewrite anything the AI got fluently-wrong.
7. Final consistency pass. Term-by-term scan as in lesson 4.
8. Deliver with a clean handover note documenting your process.

Total time: typically 40–60% of from-scratch translation time for content where MT is moderate-quality. For poor MT or specialized domains, MTPE can take longer than translating from scratch. Know your limits and price accordingly.

Key Takeaways

• Light vs full MTPE matter for both pricing and quality. Don't blur them.
• LLM polishing of raw MT — applying glossary and style — is the new productivity layer.
• Specific prompts (negations, numbers, polysemy, tags) catch what neural MT systematically misses.
• Sample-test before pricing. Document your AI use. Refuse MTPE for content where it isn't appropriate.