AI translation governance is the set of policies, standards, and human review processes an organization uses to control the quality and risk of machine-translated content. Tools like Google Translate have made automated translation widely accessible, but accessibility is not the same as fitness for purpose. Understanding how these systems work, where they perform well, and where they require human oversight is the foundation of any responsible AI translation governance program.
Google Translate is one of the most widely used machine translation (MT) systems in the world. It does not translate the way a human linguist does. Instead, it learns from enormous collections of already-translated text, identifying statistical patterns between languages and using those patterns to predict translations for new content.
Early training data came from high-quality sources including United Nations documents, which are officially published in the organization’s six languages. That gave Google’s system a large pool of verified translated text to learn from. Beyond formal sources, the system also draws on translated content across the web, though the quality of that material varies considerably. Handling variation in source quality is itself a significant research challenge in MT development.
Over time, Google Translate moved from phrase-by-phrase statistical translation to neural machine translation (NMT), which models entire sentences and handles context more naturally. That shift improved fluency significantly for common language pairs. It did not eliminate the quality variation that makes governance necessary.
MT performs best on language pairs with large, high-quality training datasets: Spanish, French, German, Mandarin, and other widely documented languages. For lower-resource languages, where less translated text exists, quality drops noticeably.
Domain-specific content presents a separate problem regardless of language pair. Medical records, legal contracts, government procurement documents, and clinical trial materials use terminology where small differences in phrasing carry significant consequences. A system trained primarily on general web content may not handle those distinctions reliably. The output may be fluent-sounding without being accurate, which is a harder problem to catch than obvious errors.
This is the core challenge that AI translation governance exists to address: MT output can appear polished while containing errors that only a subject-matter expert would recognize.
ISO 18587 (Machine Translation Post-Editing) is the international standard that defines how organizations should apply human review to MT output. It is part of Dynamic Language’s Global Management System, which is certified to five ISO standards: ISO 9001 (Quality Management), ISO 17100 (Translation Services), ISO 27001 (Information Security), ISO 13485 (Medical Devices Quality Management), and ISO 18587 (Machine Translation Post-Editing).
ISO 18587 governs two things in particular: the qualifications post-editors must hold, and the levels of post-editing that apply depending on how the content will be used. Full post-editing brings MT output up to the quality of human translation. Light post-editing corrects only errors that affect comprehension, for use cases where stylistic polish is not required. The standard specifies when each level is appropriate and what quality checks apply before content is considered ready.
Holding this certification means the post-editing process is documented, independently audited, and tied to a quality management system, not simply claimed on a website.
Human-in-the-loop translation (HITL) refers to any workflow where a qualified human reviewer evaluates and corrects MT output before it is used. The term covers a range of intervention levels, from light editing of low-risk content to full re-review of regulated or safety-critical material.
The appropriate level of human-in-the-loop involvement depends on the content type, the consequences of error, and the regulatory requirements that apply. For internal operational content with no compliance implications, light review may be sufficient. For patient-facing healthcare materials, government records, legal documents, or clinical trial submissions, full human review against the source text is typically required.
MTPE quality assurance, the process of verifying that post-edited MT output meets the required standard before delivery, is a distinct step from post-editing itself. A mature AI translation governance framework treats these as separate, documented stages, not a single informal review.
Some content categories require certified translation regardless of how good the underlying MT output appears. In these cases, AI translation governance means having clear policies about which content goes through full human translation or certified post-editing, and documenting that process for audit purposes.
Categories that typically require certified human review include:
Clinical and medical documents. Patient consent forms, clinical trial protocols, investigational brochures, and similar materials carry direct safety implications. Errors in these documents can affect patient outcomes and regulatory standing.
Government and legal records. Documents submitted to government agencies, used in legal proceedings, or required under procurement contracts often specify certification requirements. MT-only output does not meet those requirements.
Regulated product labeling. Product instructions, safety notices, and labeling for regulated industries require accuracy that can be demonstrated and audited.
For organizations in healthcare, public sector, life sciences, or pharma, the question is not whether to include human review, but how to document and govern it. That is what an ISO 18587-certified workflow provides.
If you are evaluating language service providers or building an internal policy, a sound AI translation governance framework typically covers six areas:
Content classification. A clear policy defining which content types are eligible for MT-only, MT with light post-editing, MT with full post-editing, and human translation from scratch.
Post-editor qualification. Verification that reviewers hold the subject matter expertise required for the content domain, not just general language proficiency.
Quality checkpoints. Documented steps for MTPE quality assurance before delivery, separate from the post-editing step itself.
Certification audit trail. Evidence that the process is independently verified, not self-certified. ISO 18587 certification requires annual audit by an accredited registrar.
Error categorization. A defined system for classifying post-editing errors by type and severity, which feeds continuous improvement.
Client-facing transparency. Clear documentation of which portions of a project used MT and at what level of post-editing, so clients can make informed decisions.
Dynamic Language has provided translation, interpreting, and localization services since 1985. We apply machine translation in workflows where it adds speed without reducing quality, and we pair every MT project with human post-editing governed by our ISO 18587-certified process. Every project is assessed for MT suitability before automation is introduced, and post-editors with domain expertise handle the review.
Our Global Management System is certified to ISO 9001 (Quality Management), ISO 17100 (Translation Services), ISO 27001 (Information Security), ISO 13485 (Medical Devices Quality Management), and ISO 18587 (Machine Translation Post-Editing). As an NMSDC-certified Minority Business Enterprise (MBE), we also bring supplier diversity credentials to regulated-industry procurement. We provide services in 200+ languages and dialects.
If your organization is building or reviewing an AI translation governance policy, [contact our team] to talk through your requirements.
AI translation governance is the framework of policies, standards, and human review processes an organization uses to manage the quality and risk of machine-translated content. It defines which content types are eligible for machine translation, what level of human post-editing applies to each, and how the process is documented and audited. A mature governance framework is tied to a quality management system and independently verified, not self-certified.
ISO 18587 (Machine Translation Post-Editing) is the international standard governing how qualified human reviewers should correct and verify machine translation output. It specifies post-editor qualifications, levels of post-editing (full and light), and the quality requirements that apply before MT output is considered fit for use. Organizations that hold this certification have their post-editing process audited annually by an independent registrar.
Google Translate learns translation patterns from large collections of previously translated text, including formal sources such as United Nations documents and a broad range of web content. It uses neural machine translation (NMT), which models full sentences rather than phrase-by-phrase, to produce more natural-sounding output. Quality varies by language pair and content domain. Technical and regulated content requires human review before use.
Machine translation post-editing (MTPE) is the process by which a qualified human reviewer corrects and refines MT output to meet a defined quality standard. Full post-editing brings MT output up to the quality of human translation. Light post-editing addresses only errors that affect comprehension. ISO 18587 defines the qualifications and processes that apply to each level.
Human-in-the-loop translation refers to any workflow where a qualified human reviewer evaluates machine translation output before it is used. The level of human involvement varies: it may mean full re-review against the source text, or lighter correction of comprehension errors only. For regulated content, human-in-the-loop review is a compliance requirement, not an optional quality step.
For regulated industries such as healthcare, life sciences, and government, Google Translate alone is not sufficient for documents with legal, clinical, or compliance implications. These documents require certified human review. MT may be part of a compliant workflow when it is paired with appropriate post-editing governed by a documented quality process such as ISO 18587.
MTPE quality assurance is the verification step that confirms post-edited machine translation output meets the required quality standard before delivery. It is a separate stage from post-editing itself. A documented MTPE quality assurance process, reviewed by a qualified second reviewer, is a component of a responsible AI translation governance framework.
Key questions include: Do you hold ISO 18587 certification, and from which registrar? How do you classify content for MT eligibility? What qualifications do your post-editors hold for my content domain? How is MTPE quality assurance documented? What is disclosed to clients about MT use on a project? Can you provide audit trail documentation if required?
Machine translation is generally not appropriate for patient-facing clinical documents, legal filings, government submissions with certification requirements, regulated product labeling, or any content where an undetected error carries safety or legal consequences. For these content types, a certified human translation or fully post-edited MT workflow governed by ISO 18587 is the appropriate choice.
Certified translation is produced or reviewed by a qualified human translator and attests in writing that the content is accurate and complete. Machine translation is automated and provides no such assurance. In regulated industries, certified translation is required for documents that carry legal, clinical, or compliance weight. MTPE can produce certified-quality output when governed by a documented process such as ISO 18587.
A company preparing to localize for the EU typically identifies…
Read More →
15215 52nd Avenue S., Suite 100
Seattle, WA 98188-2354
inbound@dynamiclanguage.com
206.244.6709
Toll-free: 800.682.8242
Copyright 2026 © Dynamic Language. All rights reserved.
