Purpose and scope
This guidance sets out requirements and recommended practices for the use of artificial intelligence (AI) and automation in Cochrane’s translation processes. It applies to all individuals and teams involved in the translation of Cochrane content, including volunteer translators. It covers the use of AI within Cochrane’s Translation Management System (TMS), Phrase, as well as the use of external tools.
Key messages
- Accountability. Translation teams remain fully responsible for the accuracy, quality and integrity of their translations, including any decisions to use AI or automation. All work must comply with applicable legal, ethical and data‑protection standards.
- Fitness for purpose. AI and automation may be used where teams can demonstrate that quality, methodological rigour and the integrity of Cochrane content will not be compromised.
- Human oversight. All AI‑assisted outputs must undergo human revision (post‑editing) by a translator with sufficient linguistic and domain expertise. Particular attention must be given to accuracy, terminology, consistency, and a natural, idiomatic flow.
- Preferred environment. Phrase provides integrated automation options, including machine translation and large language models, and supports quality and workflow controls. These should be used as the default environment for AI‑assisted translation.
- External tools. AI or automation tools used outside Phrase may only be used for content that is publicly available and presents no copyright, confidentiality or data‑protection risks. Users must assume that external providers may store submitted content or use it for training unless there is a clear enterprise agreement that states otherwise.
Definitions
- Artificial intelligence (AI). Any technology that enables automated processing or generation of language, including rule‑based systems, trained machine learning, neural machine translation, large language models and generative AI.
- Automation. Tools that support translation workflows without generating new text, for example term extraction, QA checkers, alignment, file preparation and formatting.
- Machine translation (MT). Systems that automatically translate text from one language to another.
- Large language models (LLMs) / Generative AI. Models that generate or transform text, including translation, summarisation and rewriting. These models can produce fluent but incorrect statements, sometimes referred to as hallucinations or fabrications.
- Relationship between MT and LLMs/AI. Machine translation is a specific application of AI focused on translating text. By contrast, large language models and other AI tools can perform a wider range of language tasks, including translation, summarisation and rewriting. These broader capabilities increase flexibility but also introduce additional risks, particularly the possibility of generating content that is not explicitly present in the source text.
- Post‑editing. Human revision of AI‑ or MT‑generated output to ensure accuracy, completeness, consistency and style compliance.
- High‑risk content. Content where inaccuracies could mislead readers or damage credibility, such as abstracts, plain language summaries, conclusions or key clinical terminology.
Principles for responsible use
- Human in the loop. AI assistance must never replace expert human judgement.
- Data protection. No personal data, embargoed material or confidential information may be entered into public or consumer AI tools.
- Bias awareness. Translators should be alert to bias or culturally inappropriate phrasing introduced by AI and correct it.
- Sustainability. Where feasible, use efficient and integrated tools to minimise duplicated processing and unnecessary use of computational resources, thereby reducing environmental impact.
...
- Embargoed or unpublished content in external tools.
- Any content containing personal data or sensitive information in external tools.
- Automated publication of AI outputs without human review.
Roles and responsibilities
- Translation project managers: Decide and discuss with the Multi-language Programme Manager if and how AI will be used for each content type, configure Phrase settings, ensure resources and training, and monitor quality outcomes.
- Translators / post‑editors: Apply AI tools responsibly, perform post‑editing to the required standard, document issues and escalate risks.
- Editors / reviewers: Provide targeted feedback on quality, terminology and style, and flag systematic AI‑related errors.
Approved tools and environments
- Preferred: Phrase, including its integrated MT and LLM capabilities, translation memories, termbases and QA checks.
- External tools: May be used only when content is public, and when doing so does not violate copyright, confidentiality or data‑protection requirements. Users must verify provider settings and opt out of training where possible.
Workflow
Post‑editing and review
- Correct meaning, omissions, additions, contradictions and fabrications.
- Ensure terminology matches Cochrane glossaries, GRADE wording and language‑specific conventions.
- Ensure style is clear, natural and idiomatic, and that reading level is appropriate for the target audience, especially for plain language summaries.
- Resolve inconsistencies across the document and maintain coherence with previously published translations.
...
- Capture recurrent errors to inform engine configuration, termbase updates and translator guidance.
Data protection, confidentiality and copyright
- Do not input personal data, unpublished manuscripts, reviewer comments, restricted datasets or confidential or embargoed communications into external AI tools.
- Use enterprise or Cochrane‑provided accounts where available, not personal consumer accounts.
- Review provider terms of use for data retention, training and intellectual property. Do not use tools that claim ownership over input or output.
- Ensure that copyright and licensing for source and target texts remain with Cochrane or the original rights holder as applicable.
- Where legally required, ensure a valid data processing agreement is in place with any third‑party provider.
When AI should not be used
- When the content requires nuanced clinical interpretation or complex methodological judgement that AI is likely to distort.
- When deadlines or resources do not allow proper post‑editing and QA.
- When legal, ethical or contractual constraints prohibit external processing.
- When a prior assessment shows persistent quality risks for a specific language pair, domain or content type.
Training and quality alignment
- Provide regular training on post‑editing practices, common AI error patterns and the use of Phrase QA features.
- Foster shared quality expectations by discussing and sharing annotated examples of acceptable and unacceptable AI outputs.
- Maintain and update termbases, style guides and reference translations.
Appendix A. Glossary
- Hallucination / fabrication: Fluent but incorrect content produced by an AI model.
- Termbase: A curated multilingual list of approved terms.
- Translation memory (TM): A database of previously approved translations for reuse.
- QA checks: Automated checks for numerical, formatting and terminology consistency.