AI in Certified Translation: What Is Professionally Defensible
IRCC bans machine translation. USCIS requires an accountable human. Where the line actually falls for an LSP building an AI-assisted certified workflow β and what to tell clients who ask.
Every language service provider selling certified translation has now had the conversation twice: once internally, about whether AI belongs in the workflow at all, and once with a client asking whether the agency uses it.
Both conversations go badly when the answer is vague. The regulatory position is actually clearer than the industry discourse suggests, and an agency that understands it precisely can give a confident answer to procurement rather than a defensive one.
What the authorities actually say
IRCC (Canada) is the only one that names the technology: translations produced by machine translation tools are not acceptable.
USCIS (United States) says nothing about machine translation. Under 8 CFR Β§ 103.2(b)(3) it requires a certification from a translator attesting to completeness, accuracy, and their own competence.
UKVI requires the translator's name, contact details, credentials, and a confirmation of accuracy.
Home Affairs (Australia) requires a NAATI-certified translator, with the practitioner number on the translation.
Three of the four are not making a technology rule at all. They are making an accountability rule: a named, competent, independent person must attest to the translation and be identifiable afterwards. Raw machine output fails not because a machine produced it but because no one is standing behind it.
Even IRCC's rule, read carefully, prohibits machine-generated translations submitted as the translation. It does not prohibit a certified translator from using software while producing work they then certify.
The line
Defensible: a qualified translator uses software to produce a first draft, reviews it against the source, corrects it, and certifies the result in their own name. This is what the profession has done with translation memory, terminology databases, and CAT tools for thirty years. No authority has ever objected to the category.
Not defensible: raw output submitted with no human review. Also not defensible β and this is the one that matters for agency governance β a translator signing a certification for output they did not actually read.
That second case is the real risk, and it is not a technology risk. It is a supervision risk. A certifier who signs an unread draft has committed the same professional violation whether the draft came from an AI system or from an unsupervised junior linguist. The tool changes the volume at which it can happen, not the nature of the breach.
An agency's AI policy should therefore be a review policy, not a tool policy.
What generic MT gets wrong on certified documents
There is a legitimate technical reason certified work resisted automation longer than commercial content, and it is worth being able to articulate to a client.
Generic machine translation is built for prose. Certified civil documents are structured artefacts, and the failure modes are structural:
- Tables flattened into prose, so values no longer sit under their headers
- Right-to-left sources mirrored visually rather than remapped logically β every value lands under the wrong heading
- Seal and stamp text skipped entirely, because it is an image, not text β and seal text names the issuing authority, which officers verify
- Blank form fields dropped silently, with no indication anything is missing
- Marginal annotations missed β on Latin American civil records these carry divorce, legitimation, and correction history
- Identifiers mistranscribed β an 18-digit Chinese ID, a 32-digit Brazilian matrΓcula, a CURP
- Names mangled β diacritics stripped, surname order reversed, transliteration inconsistent with the passport
- Calendar conversions silently wrong β Hijri lunar and Solar Hijri handled as if they were the same system
On a brochure these are quality defects. On a filing they are rejection causes. That distinction is why "we use AI" and "we use MT" should not be the same sentence in a client conversation: purpose-built layout-preserving drafting addresses precisely the failure modes generic engines create.
Answering the procurement question
Corporate and law firm buyers now ask directly. The weak answers are "no, everything is human" β often untrue, and increasingly implausible on price β and "yes, we use AI" with no further structure.
The strong answer has three parts:
- What the tool does. Produces a complete, layout-faithful first draft from the source scan.
- What the human does. A qualified translator reviews it line by line against the source, corrects it, applies client terminology, and certifies it in their own name with their own contact details.
- Who carries the accountability. The named certifier, whose credentials and details appear on the certification statement and who can be contacted by the receiving authority.
Then state the compliance position explicitly: the deliverable meets the destination's requirement β USCIS certification, UKVI credentials, IRCC certified translator or affidavit, NAATI practitioner number β because a qualified human produced and signed it.
Buyers are not usually trying to catch you out. They are trying to establish that the accountability chain is intact.
Governance an agency should actually have
If AI drafting is in your workflow, these controls are what make the position defensible under scrutiny:
A mandatory review step that cannot be skipped. Not a policy statement β a workflow gate. The certifier confirms review before the certification is generated.
Review time proportional to the document. A reviewer processing forty certificates an hour is not reviewing them. Track it.
Certifier competence in the actual source language. A Farsi-competent reviewer cannot certify Pashto. A Hindi reviewer cannot certify Gurmukhi. This is the control most often missing.
An audit trail linking each certification to a named reviewer and a timestamp.
Client disclosure that matches reality. If your contract says human-only and your workflow is AI-assisted, the exposure is contractual, not regulatory β and it is the more likely one to actually bite.
Escalation for documents the tool handles poorly. Severely degraded scans, heavy handwriting, and unusual scripts should route to full human production rather than through a review step that becomes a rewrite.
The independence rule still applies
One constraint no workflow change removes: the certifier must be independent of the applicant. USCIS treats the applicant as an interested party; Home Affairs excludes the applicant, family members, and the migration agent.
For agencies serving law firms, that last exclusion matters commercially. An Australian migration agent cannot certify their own client's documents, which is precisely why they outsource to you. It is worth naming in a pitch.
Frequently asked questions
Can we tell clients we do not use machine translation? Only if true. A more durable claim is that every certified deliverable is reviewed and signed by a named qualified translator who carries the accountability β which is what the authorities require and what buyers are checking.
Does IRCC's ban rule us out of Canadian work? No. It rules out submitting machine output as the translation. A provincially certified translator using drafting software and then certifying the reviewed result is standard practice.
What is our largest actual risk? Certifiers signing work they did not read. Build a workflow gate, not a policy paragraph.
Should reviewers be paid less because drafts are cleaner? Review is judgment and accountability, not typing. The time falls; the responsibility does not. Repricing review as a clerical task is how agencies end up with certifiers who stop reading.
The rule is about accountability, not tooling. Make review a gate rather than a policy, match certifier competence to the actual source language, and give procurement a three-part answer instead of a defensive one.
Want to see how a layout-faithful draft changes review time? Open ZenTrans and run a sample batch β the first 5 pages are free, and volume pricing runs to $0.70 per page.