Life sciences teams are under two pressures at once. Regulated content volume keeps growing, timelines keep compressing, and machine translation helps with both. Content that ends up in a submission, a label, or a validation file also has to be defensible to an auditor, and raw machine output will not survive that. So the working question for a regulated team is how to use machine translation so the result holds up. Machine translation post-editing services, governed by ISO 18587, were written to answer it.
What ISO 18587 actually governs
ISO 18587:2017 is the international standard for machine translation post-editing services, the full human post-editing of machine translation output. It applies only to content that has been processed by a machine translation system. Human translation from scratch is covered by a separate standard, ISO 17100. Within its scope, ISO 18587 sets requirements for two things: the post-editing process itself, and the competences a qualified post-editor must have.
What the standard covers is narrow and deliberate. It does not regulate the machine translation engine or make claims about the AI model. Its scope begins after the machine produces output: a qualified human post-editor reviews and corrects that output against defined requirements, to a quality level comparable to human translation, with documentation at each step. The standard governs the controlled, human, recorded process wrapped around the technology. For a regulated buyer that is the useful thing to standardize, because what a validation file audits is a process.
Why raw machine translation fails in a validation context
Regulated companies operate under the principle that a process affecting product quality must be defined, controlled, and documented, so that it can be shown to work and shown to have worked. A validation file is the evidence that a process meets that bar. Raw machine translation cannot enter that file, for reasons an auditor will name immediately: there is no qualified human accountable for the output, no defined review step, no record of who did what, and no way to demonstrate the result is accurate and complete. An unreviewed machine rendering of a regulatory document is an uncontrolled input that looks like finished work.
That appearance is the difficulty. A machine translation engine can produce fluent, confident output that reads as complete, which makes it easy to treat as done. In regulated content, fluency and accuracy are separate properties, and the space between them is where a mistranslated dosage, a shifted meaning in an instruction, or an altered term in a submission hides. Fluency is what makes unreviewed output dangerous in a context where accuracy is a compliance requirement.
How ISO 18587 documentation fits the validation file
The standard requires records of the post-editing process, so a workflow run under it produces the kind of evidence a validation file needs.
A feasibility assessment. The standard requires deciding, before the project starts, whether the content is suitable for machine translation and post-editing at all. Sensitive, highly nuanced, or high-risk material may call for full human translation. That documented go or no-go decision is itself an audit artifact, showing the workflow was a deliberate choice.
Qualified post-editor records. ISO 18587 defines the competences a post-editor must have and expects those qualifications to be documented. For a validation file, that answers the auditor’s first question about any process step: was the person doing it qualified, and can you prove it.
A defined, recorded process. The standard requires the post-editing to follow a defined process with records at each step. That record is what turns AI-assisted translation from an untraceable shortcut into a documented workflow that can be described and defended.
An accountable human in the loop. Under ISO 18587, a qualified human is responsible for the final output. Human responsibility stays where it was, and the AI component adds a layer that has to be governed. For a validation file, that clear line of human accountability is required, and the standard supplies it.
Taken together, these are the same elements a regulated company documents for any controlled process: a decision to use the method, qualified people, a defined procedure, and records. ISO 18587 maps machine-translation post-editing onto that structure, which is what puts AI-assisted translation inside a validated workflow.
The governance point buyers should take from this
As AI adoption in translation accelerates, almost every vendor uses machine translation or soon will, so the useful question has moved to governance: whether a defined, accountable, documented process sits around the machine output. A vendor can use capable AI tools and still have no post-editing standard behind them, which for a regulated buyer means no accountability framework, and nothing that belongs in a validation file. The question worth asking a vendor is what governs the human review after the machine, more than which engine produced it.
Dynamic Language holds ISO 18587, alongside ISO 17100 (Translation Services) for human translation and ISO 13485 (Medical Devices Quality Management) for the device and life-sciences quality system. Our role on AI-assisted translation is advisory: where machine translation fits a client’s content, where full human translation is the better call, and governed post-editing for the content that qualifies.
Frequently asked questions
What is ISO 18587? ISO 18587:2017 is the international standard for machine translation post-editing services, the full human post-editing of machine translation output. It sets requirements for the post-editing process and for the competences of qualified post-editors. It applies only to machine-translated content. Human translation is covered by ISO 17100.
Does ISO 18587 regulate the AI or machine translation engine? No. It does not regulate the machine translation tool or make claims about the AI model. It governs the human post-editing process that happens after the machine produces output, which is the part a regulated buyer can audit.
Can machine translation be used for regulated life sciences content? It can, when the content is suitable and the output is fully post-edited by a qualified human under a documented process such as the ISO 18587 standard. Highly sensitive, nuanced, or high-risk content may still call for full human translation, and the standard requires assessing suitability before the project starts.
How is ISO 18587 different from ISO 17100? ISO 17100 governs human translation services, while ISO 18587 governs the human post-editing of machine translation output. A provider holding both can apply the right standard to each content type, human translation where that is required and governed post-editing where machine translation is suitable.
What should we ask an AI-translation vendor for regulated content? Ask which standard governs their post-editing, and treat the answer as more informative than the engine name. Certification to the ISO 18587 standard indicates a defined, accountable, documented process around machine output, which is what an audit will look for.
Putting MTPE into a defensible workflow
If your team is deciding where machine translation fits in regulated content, the answer is usually a documented decision about which content is suitable, with a governed post-editing process for the content that qualifies. Dynamic Language provides machine translation post-editing services under ISO 18587, alongside ISO 17100 and ISO 13485, and advises on where the workflow fits your content and where full human translation is the better call. Talk to a Life Sciences Language Specialist.
Tags:
Life Sciences
Comments