You Replaced the Interpreter With a Model
Nobody approved that decision. An interpreter line got expensive, a translation widget got installed, a multilingual chatbot went live — and language access stopped being a staffed service and became a software default that no one has tested.
Title VI prohibits national-origin discrimination by recipients of federal financial assistance, and failing to provide meaningful access to people with limited English proficiency is treated as exactly that. The standard is functional — could the person understand and participate — and it scales with the number of people affected, the frequency of contact and the importance of the service. In healthcare, Section 1557 rules go further and condition the use of machine translation on qualified human review where accuracy is essential. None of these frameworks ask whether the organisation intended to exclude anyone.
Six Surfaces, One Question
Language access is not one programme. It is a set of surfaces, each with a different consequence when it fails, and automation has arrived at all of them at once — usually procured separately, by different teams, none of whom owns the language-access plan.
How the Failure Actually Runs
Nobody experiences a "translation error". They experience a missed hearing, a wrong consent, a terminated benefit. The chain from one to the other is short, and the organisation's own records go quiet at exactly the point where the evidence would be.
Detection misreads code-switching, a regional variant or a low-resource language and silently picks the nearest match.
Terms of art — 'appeal', 'coverage', 'consent', 'deductible', 'custody' — are rendered into ordinary words that do not carry the legal meaning.
A form is signed, a deadline passes, a hearing is missed. The organisation records this as a completed interaction.
The person asks a follow-up question, and the escalation path is English-only or routes to a queue with no interpreter.
The file shows an English document and an English encounter note. Nothing records which language was actually used or by what means.
An agency asks for the language-access plan, vendor terms and evidence of who translated the vital document.
Fluency Is Not Accuracy, and It Hides the Difference
The reason machine translation is more dangerous here than an obviously broken system is that its output reads as authoritative. A person receiving a fluent paragraph in their own language has no signal that "appeal" was rendered as a casual request, that a conditional was dropped, or that a deadline was converted into a suggestion. Neither does the staff member who sent it, because they cannot read the output either.
Low-resource languages compound it. Model quality varies enormously across languages, and the populations most likely to depend on language assistance are frequently the ones least represented in training data. A single evaluation of "our translation quality" run on Spanish and French tells an organisation nothing about the languages where its exposure actually sits.
The Cost Argument Is the Record Against You
The federal framework does weigh resources — meaningful access is assessed against the organisation's size and means. But the analysis is comparative, and a procurement decision that replaced a staffed interpreter line with a subscription is documented proof that the service was previously affordable. Budget papers written to justify the switch tend to state the savings and the volume of LEP contacts in the same paragraph, which is the entire disparate-impact case assembled by the organisation itself.
Six Controls Worth Having
Write down which categories are vital documents requiring qualified human translation and which may use machine output with notice. Without the tiering, every decision is made ad hoc by whoever is closest to the deadline.
Notices of rights, appeal instructions, adverse determinations and consents are exactly where a translation defect converts into a lost legal remedy. Exclude these routes from any site-wide translation widget at the technical level, not by policy alone.
A multilingual bot inherits every error in its English source and adds its own. Test non-English paths with the same evaluation set used for English, and have native speakers with domain knowledge score the answers, not just the grammar.
If the system will answer in a language, it must be able to hand off in that language. An English-only human fallback turns a partial accommodation into a documented dead end.
Log which language was used, whether machine or qualified human translation was applied, and who reviewed it. Without that field, the organisation cannot demonstrate what it provided to a specific person on a specific day.
Translation providers routinely disclaim fitness for legal, medical or safety-critical use in their terms. Deploying the tool for exactly that purpose, against the vendor's stated limitation, is the finding that does the most damage in an investigation.
Related Reading
- Section 1557 and AI clinical decision support — the same statute reaching the other half of the automated healthcare stack.
- Speech recognition and accent discrimination — where the input side of the language pipeline fails on the same populations.
- AI in public benefits eligibility — the decisions that the untranslated notice was about.
Find Out Which Pages Are Being Machine-Translated
Most organisations cannot say which of their pages a translation widget currently covers, whether rights notices and forms are inside that scope, or whether a non-English visitor can reach a human at all.
See how your site behaves for the people it is least tested on. Run a free scan and start the language-access plan from what is actually deployed.
This article is general information and not legal advice. Language-access obligations depend on funding sources, sector-specific rules and state and local requirements that differ materially, and federal guidance in this area has changed over time. Consult qualified counsel before changing how your organisation provides language assistance.