Most health systems now run some machine translation through their patient-facing content. Section 1557 permits that, with one condition: when material is critical to a person’s rights, benefits, or meaningful access, a qualified human translator must review the output. The harder question for a compliance team is evidentiary. Asked which translated content received that human review, most organizations cannot answer from records.
Language access compliance in healthcare has become a documentation problem as much as a linguistic one. What follows is what the guardrail requires, why the gap tends to open in the workflow rather than in the translation itself, and what a control that survives review looks like.
What the guardrail says
The final rule, effective July 5, 2024, allows covered entities to use machine translation, and attaches a human-review requirement in three circumstances: when the underlying text is critical to an individual’s rights, benefits, or meaningful access, when accuracy is essential, and when the content is complex or technical. In those cases a qualified human translator must review the machine output. Raw output alone does not satisfy the standard for that material.
The rule also defines what qualified means for the person performing that review. A qualified translator works to a defined written standard: accuracy and impartiality without omissions or additions, command of specialized vocabulary, and preservation of the tone and meaning of the source. Competence is measured against stated criteria rather than assumed from fluency.
Why the gap opens in the workflow
The rule sorts patient content by consequence. Most translation workflows sort it by file. A batch of documents goes out to a vendor or a tool, comes back translated, and gets published, with nothing in the process recording which items carried rights-or-benefits weight and therefore needed a qualified human in the loop.
That mismatch is what turns a reasonable practice into audit exposure. An organization can be applying human review to exactly the right content and still be unable to demonstrate it, because the decision lived in an individual’s judgment rather than in a documented routing step. In a compliance file, an undocumented control and an absent control read the same way.
A second regulator is asking the same question
Health care is not the only sector where this demand is surfacing. Regulators in other domains are converging on a similar expectation, which is that an organization applying AI to consequential content be able to show where human oversight sat and on what basis. The European Union’s AI Act is the current example most often cited. For an organization that operates across borders, or that reads the direction of travel as durable, building the record once answers more than one regulator.
What a defensible control looks like
A control that holds up has three parts, and none of them require abandoning machine translation.
Triage at intake. Content gets classified before translation rather than after. The test is whether the item bears on a person’s rights, benefits, coverage, or access to care. Notices of coverage, appeals and grievance materials, consent forms, and eligibility correspondence sit on one side of that line. Wayfinding, event notices, and general information sit on the other.
A tiered workflow tied to the classification. High-consequence content routes to full human translation with an independent revision step. Lower-consequence content can run through machine translation with qualified human post-editing. Because the classification determines the route, the decision is made by policy rather than case by case.
Records that link each file to the route it took. This is the part that answers the auditor’s question. For every translated item, the file shows the classification applied, the workflow it triggered, and the qualified linguist who performed or reviewed the work.
Two standards give a compliance team documented evidence for this. ISO 17100 (Translation Services) defines the qualified-linguist requirements and the mandatory independent revision step for human translation. ISO 18587 (Machine Translation Post-Editing) defines the process and competence requirements for human review of machine output, which is the activity the rule requires on critical content. Dynamic Language holds both, so the review a compliance team needs to evidence is performed under an independently audited process rather than an internal practice.
One health plan engagement shows the pattern in practice. Benefit plan documents were translated across several markets on a fixed budget. Rather than send whole files through a single process, we engineered the documents and classified the content inside each one: the passages carrying real liability for a member, and the passages that were informational, such as lists of products covered under a given plan. Liability-bearing content went through full human translation with two linguists and a separate quality review. Informational content went through machine translation with human post-editing. Because the file engineering is a one-time cost, the saving carried across every language in the program and funded two languages that had not fit the original budget, Vietnamese and Korean. The compliance value sits alongside the cost outcome. The classification is recorded per file, so the plan can show which content received human translation, which received post-editing, and on what basis.
What else the rule requires
The guardrail sits alongside several other language access obligations a compliance review will test.
Notice of Availability. Covered entities must actively tell people that free language assistance is available, in English and in at least the 15 most common languages spoken by individuals with limited English proficiency in the relevant state, posted in physical locations and on the website and included in significant communications.
Defined standards for qualified interpreters and translators. A qualified interpreter must be proficient in English and at least one other language, interpret accurately and impartially, handle specialized vocabulary, preserve tone and sentiment, and follow interpreter ethics. A parallel written standard applies to translators.
Limits on family members, minors, and staff. Covered entities are discouraged from relying on a patient’s family or friends and restricted from using minors except in narrow emergencies. Bilingual staff may interpret only where they meet the qualified-interpreter standard for that work.
Coverage of patient companions. Language assistance obligations extend to companions of the patient, such as a parent or spouse who is the person a provider actually communicates with about care.
The core duty is long-standing: take reasonable steps to provide meaningful access to individuals with limited English proficiency, and provide language assistance free of charge. Patients cannot be required to bring or pay for their own interpreters, confidentiality obligations apply, and remote interpreting by video and audio is available. For most programs the foundation is already in place, and the gaps tend to be the notices, the qualification documentation, and the machine-translation guardrail.
A note on the current regulatory picture
Section 1557 has been the subject of litigation and administrative revision across multiple administrations, and specific provisions have been challenged or stayed at various points. The requirements described here are those established by the final rule effective July 5, 2024. Because enforcement details and the status of particular provisions can shift, a covered entity building or updating its language access program should confirm the current posture with counsel rather than assume any single version is settled. The underlying statutory duty to provide meaningful access has been consistent.
Frequently asked questions
Can we use machine translation for patient materials under Section 1557? Yes, with a limit. For content critical to an individual’s rights, benefits, or meaningful access, where accuracy is essential, or where the material is complex or technical, machine translation output must be reviewed by a qualified human translator. Lower-stakes content has more latitude.
How do we show which content received human review? Through records tied to each translated file: the risk classification applied at intake, the workflow that classification triggered, and the qualified linguist who performed or reviewed the work. A practice that exists without documentation is difficult to evidence in a compliance file.
What counts as critical content under the machine translation guardrail? Material bearing on a person’s rights, benefits, or meaningful access to services. In practice that includes notices of coverage, appeals and grievance materials, consent documents, and eligibility correspondence. Content where accuracy is essential or the subject matter is complex or technical also falls inside the requirement.
Who counts as a qualified translator for that review? Someone working to a defined written standard: accurate and impartial rendering without omissions or additions, command of specialized vocabulary, and preservation of the tone and meaning of the source. Competence is measured against criteria rather than inferred from bilingualism.
What is Section 1557? Section 1557 is the nondiscrimination provision of the Affordable Care Act (42 USC 18116). It prohibits discrimination based on race, color, national origin, sex, age, or disability in health programs and activities receiving federal financial assistance, and it is the basis for language access obligations toward people with limited English proficiency.
Which version of the Section 1557 rule applies? The requirements described here come from the final rule effective July 5, 2024. The statutory duty to provide meaningful access has been consistent, but specific regulatory provisions have been revised across administrations and some have faced litigation, so confirm the current status of any particular requirement with counsel.
Reviewing your language access program
If your team is checking a language access program against Section 1557, the machine-translation guardrail is usually where the documentation is thinnest, followed by the Notice of Availability and the evidence behind interpreter and translator qualification. Dynamic Language can review how your language services map to those requirements and provide the documentation a compliance file needs. Talk to a Language Access Specialist.
Tags:
Healthcare
Comments