Machine translation and language access for agency communications

Reference article · 5 min read

Most agencies already have machine translation in place, often a widget dropped into the website footer and with no written policy behind it. The technology has improved enough to carry real weight in a language access program, but not enough to carry every kind of content. The line between the two is not drawn by quality scores. It is drawn by what happens when a translation is wrong.

A workable policy sorts content by consequence. Qualified human translators take material where an error changes someone's rights, benefits, or deadlines. Machine translation with human review takes the broad middle. Raw machine output is allowed only where the alternative is no translation at all. The rest is terminology control, honest labeling, and website plumbing.

What the obligation looks like

• Federal civil rights law. Title VI of the Civil Rights Act of 1964 has long been read to require recipients of federal financial assistance to take reasonable steps to provide meaningful access to people with limited English proficiency. Executive direction and agency guidance here have shifted recently, so confirm current expectations with your funding agency and counsel.
• Program and grant conditions. Health, benefits, housing, and election programs often carry their own language requirements through grant terms, state plan obligations, or the language minority provisions of the Voting Rights Act.
• State and local law. Several states and many cities have bilingual services statutes or language access ordinances, often more specific than federal guidance about which documents must be translated and into which languages.

Federal guidance has traditionally used the idea of vital documents, material whose denial or delay could carry serious consequences. It is a useful planning frame even if it is not the operative test in your jurisdiction.

Sort content by consequence

• Qualified human translation. Notices of action, eligibility and benefit determinations, appeal rights and deadlines, consent forms, contracts, complaint forms, court related material, and clinical instructions. If a document could be read back at a hearing, it does not get machine translated.
• Machine translation with human review. Program descriptions, how to apply pages, frequently asked questions, and instructional material, edited by a qualified translator before publication.
• Raw machine output, clearly labeled. Long tail website content, archived material, and pages that would otherwise exist in English only.

The test is not how difficult the text is. It is what a wrong rendering causes.

Glossaries and terminology control

Engine choice matters less than the terminology discipline around it.

• Do not translate list. Program names, system names, form numbers, and statutory titles pass through untouched, so residents can match what they see online to what arrives in the mail.
• Approved term base. Words like hearing, appeal, household, and income have several plausible renderings in most languages, and the wrong one changes the meaning. Settle them once, in writing, with your translators.
• Translation memory. Approved segments get reused, which keeps successive notices consistent and lowers cost.
• Source writing. Plain language, short sentences, no idiom, and no text baked into images improve machine output more than any engine setting.

Treat the glossary and translation memory as agency assets, exportable in standard formats at any time, not only at contract exit.

Disclaimers and labeling

A disclaimer manages expectations. It does not reduce an obligation to provide meaningful access, and it never substitutes for translating what must be translated.

• State plainly, in the target language and next to the content, that the text was machine translated and not reviewed.
• Identify which version is authoritative, where your counsel confirms you may do so.
• Give a path to a human: a number for interpretation, how to request a translated copy, and how to report a bad translation. Route those reports into the glossary.

Website integration

• Published pages beat live widgets for important content. Translated pages can be reviewed, cached, and indexed. Widgets cover the remainder.
• Mark the language. Correct language attributes affect how screen readers pronounce text.
• Test the widget. Auto-translation layers often miss validation messages, dynamic content, and alt text, and some break keyboard navigation.
• Move vital content out of PDFs. Machine translation handles HTML far better, and scanned documents not at all.
• Pre-translate emergency templates. Alert messages should be reviewed in advance, not produced live during an incident.
• Offer a persistent language selector rather than redirecting by browser locale.

Spoken language is a separate problem

Translation tools do not replace interpreters. Live speech translation suits wayfinding and simple counter logistics, not eligibility interviews, clinical encounters, investigations, or any conversation that becomes part of a record. Keep telephonic and video remote interpreting available, assess and formally designate bilingual staff rather than relying on whoever is nearby, and record the interpreter and language used. Deaf and hard of hearing residents fall under separate requirements calling for qualified interpreters and captioning, not translation tools.

What to ask vendors, and how to start

Ask whether submitted content is retained or used to train models, for how long, and where. Ask whether your glossary and do not translate list are applied at translation time. Ask for quality evidence by language pair, since performance varies widely. Ask who performs post-editing and what credentials those translators hold. Ask for accessibility documentation for any widget, then test it. In the contract, secure ownership and export of glossaries and translation memory, limits on retention and training, accessibility conformance, and transition support.

To start, pull the language data you already hold from case records, call center logs, and enrollment systems. Inventory public content and assign each item to a tier. Build the glossary from your most common notices, then write a short policy naming who approves what.

HQT works with state and local agencies on content systems, integration, and the review workflows behind multilingual services.

Frequently asked questions