How can translators work with AI?

Machine translation is not new to professional translators. Quebec's order of certified translators, terminologists and interpreters, OTTIAQ, notes that its members have built neural machine translation into their work since 2016, with steps in place to catch the machine's errors. What has changed is the arrival of general-purpose AI chatbots that translate on request. They are often fluent and fast, and they bring a new risk: OTTIAQ points out that they can produce inaccurate translations and even invent content.

The International Federation of Translators (FIT), which represents translator associations worldwide, updated its position on machine translation in 2025 to cover these tools. Its message is balanced. AI can produce acceptable output for many kinds of text, but it can also introduce bias, drift in terminology and miss cultural nuance, so it should complement translators rather than replace them. FIT prefers the phrase "human at the core" to "human in the loop": the translator leads the process and takes responsibility for the decisions.

Step 1: Decide whether AI belongs on this job

FIT suggests judging each use case by five factors: the field, the kind of text, its topic, who will read it and why it is needed. A text where an error could hurt someone, such as medical information or certain user manuals, carries high risk. Marketing and public policy carry reputation risk. Lower risk applies when experts use a rough translation only to sort or screen information, not to publish it.

Step 2: Protect confidentiality

Both OTTIAQ and FIT warn that free online translation and chat tools give no guarantee that your text stays confidential. OTTIAQ adds that texts sent to free tools may be reused or used to train AI, and that such services may not meet Quebec or Canadian data protection laws. Before pasting a client's document anywhere, check your contract, use a tool whose data terms you have read and your client accepts, and remove names and identifiers when you can.

Step 3: Give the tool your terminology

Post-editing means correcting and improving a machine's draft until it meets the standard a client expects. It goes faster, and needs fewer fixes, when the draft follows your terminology from the start. Here is a worked prompt for an English-to-French job:

Translate the text below from English into Canadian French for a general audience of homeowners.
Rules:
- Use "vous", a clear and friendly register.
- Use this glossary exactly:
  heat pump = thermopompe
  rebate = remise
  energy audit = évaluation énergétique
- Keep all numbers, dates and product names exactly as written.
- Do not add or remove information. If a sentence is ambiguous, translate it literally and mark it with [CHECK].
Text:
[paste the text]

Step 4: Post-edit and check

Read the draft side by side with the source, sentence by sentence. Then run a short checklist:

1. Meaning: no omissions, additions or reversed meanings, especially in negatives and conditions.
2. Numbers and names: every figure, date, unit and proper noun matches the source.
3. Terminology: glossary terms used everywhere, without synonyms creeping in.
4. Register and audience: one form of address throughout and vocabulary right for the reader.
5. French conventions: quotation marks, spacing, capitals and number formats follow the client's style guide.
6. [CHECK] marks: each one resolved and removed.

A log of what you fixed helps you price the next job honestly. For example, an illustrative 500-word page might need 3 terminology fixes, 2 changes of meaning and 6 style edits. If meaning errors are frequent, translating from scratch may be faster. FIT cites research showing that gains vary by language pair and text type, that post-editing is sometimes slower than translating, and that some translators find it more draining.

Step 5: Say what the client is getting

FIT describes two labels for readers: Professionally Verified Translation, for text a qualified professional has checked, and Unverified Translation, for raw machine output. Whatever words you use, tell clients which one they are receiving, and keep the decision to deliver unchecked output with them, in writing.

Which jobs still need a full human translation?

FIT lists areas where machine output reaches its limits and human judgment is indispensable: patient consent forms that need cultural adaptation, legal texts that depend on legal requirements or precedent, and literary work that must keep its style. It also names institutional, medical, marketing and audiovisual translation, and highly technical manuals, as fields that still need professional expertise. Subtitles are a known weak spot, because machines struggle with timing, space limits, visuals and spoken language.

OTTIAQ's guidance points the same way. It recommends a human expert in every translation process, and especially for complex or sensitive documents where an error can have serious consequences, and it asks the public to hire a certified translator whenever a translation must be reliable.

In practice, "full human translation" does not mean no software. A translator may still use translation memories, terminology databases or an AI tool to double-check consistency. It means a professional writes the text and answers for every word, which is what a court, a hospital or a publisher needs.