Has Neural Machine Translation Reached the End of Its Useful Life
- 1 day ago
- 5 min read
For any new regulated translation workflow you build in 2026, there's little reason to start it on neural machine translation. Segment-level NMT still runs fast and cheap, and it keeps doing useful work inside engines companies already paid to tune. For anything you're starting fresh, large language models constrained by your translation memory and checked by certified linguists produce output NMT can't match on context, consistency, and audit readiness.
What actually changed
Classic neural machine translation reads one segment at a time. It translates a sentence without seeing the paragraph around it, so a term you defined on page 3 can drift by page 40. That's fine for a support ticket. It's a problem for an Instructions for Use booklet or a defense bid, where a single inconsistent term draws a reviewer's query.
LLM translation reads the whole document. It holds context across sections, keeps terminology steady, and produces text that reads like a person wrote it. The trade-off is real. LLMs cost more per word and run slower than a tuned NMT engine, and for regulated content that cost buys coherence a notified body or a procurement officer won't flag.
Where NMT still earns its keep
Neural MT isn't finished. It stays the right tool in a narrow set of cases:
You've already sunk heavy investment into a domain-tuned engine, and the switching cost outweighs the quality gain.
Your content is high-volume, low-stakes, and repetitive, where throughput matters more than document-level polish.
You need deterministic, reproducible output for a pipeline that can't absorb the variation an LLM can introduce.
Your latency budget is tight enough that per-document LLM processing won't fit.
Outside those cases, the argument for starting new work on NMT is thin. The quality gap now runs the other way.
Why LLM output still needs a certified human
An LLM that reads your whole document still makes mistakes a fluent reader would catch. It can invent a plausible term, flip a negation, or paper over an ambiguity the source left on purpose. In regulated content a confident wrong answer is worse than an obvious one, because it slips past a skim.
So post-editing matters, and the standard behind it matters more. ISO 18587 sets the requirements for full human post-editing of machine translation output, including the competences the post-editor holds and the responsibility they take for the final text. Today it addresses MT output. The revision underway broadens it to post-editing of non-human translation output, which pulls LLM results squarely into scope, with the revised version expected in 2026.
Human translation itself runs under ISO 17100, which requires qualified linguists and an independent revision step. Pairing the two, generation under 18587-grade post-editing and review under 17100, is what makes AI output defensible in an audit.
How we run it at AD VERBUM
AD VERBUM built our production workflow around this shift rather than against it. We ingest your translation memory and term base first, so the model starts from approved terminology instead of a generic guess. Our LangOps System then generates output on client-tuned open-weight models we host ourselves, constrained by those term bases, and a certified subject-matter linguist reviews every regulated project for technical accuracy and compliance.
The pipeline runs on ISO 27001 and ISO 42001 certified, EU-hosted infrastructure, so the audit trail exists before anyone asks for it. That ISO 42001 AI-management governance is what the EU AI Act now expects from any AI-assisted regulated workflow, and it's still rare among translation companies that hold ISO 42001.
How to decide for your own content
Before you keep or retire an NMT engine, ask five questions:
Is the content regulated, and would an inconsistent term draw a reviewer's query?
Have you already paid to tune an engine, and what's the real switching cost today?
Does your provider constrain generation with your TM and term base, or ship generic output?
Is there a certified human post-editing step under ISO 18587, with a named person responsible?
Where does the data sit, and does the workflow produce ISO 42001 and ISO 27001 evidence?
If the content is regulated and the answers point to generic output with no certified review, the engine isn't your problem. The workflow around it is.
Our AI translation services
Our translation services for regulated sectors run on ISO 27001 and ISO 42001 certified, EU-hosted infrastructure, with no reliance on public cloud tooling for core processing. Every project runs through our AI+HUMAN hybrid workflow: we ingest client Translation Memories and Term Bases first, our proprietary LLM-based LangOps System generates output constrained by client terminology on client-tuned open-weight models, and our certified subject-matter experts review for technical accuracy and regulatory compliance. Our QA is aligned to ISO 17100 and ISO 18587, with sector-specific requirements such as ISO 18587 post-editing of machine and LLM output and ISO 42001 AI-management governance applied where relevant. We serve Life Sciences, Legal, Finance, Defense, and Manufacturing clients across 150+ languages with 3,500+ subject-matter linguists. For teams managing audit-sensitive content, contact us to discuss your security and compliance requirements directly.
FAQ
Is neural machine translation obsolete?
No, but its role has narrowed. For new regulated workflows, LLMs constrained by your translation memory and reviewed under ISO 18587 outperform segment-level NMT on context and consistency. NMT stays defensible mainly where a tuned engine already exists and the switching cost is high.
What's the difference between NMT and LLM translation?
Classic NMT translates one segment at a time without document context, which lets terminology drift across a long file. An LLM reads the whole document, holds terms steady, and produces more fluent text, at higher cost and latency. For regulated content, the document-level consistency is worth the trade-off.
Does LLM translation still need human review?
Yes. LLMs can invent a plausible term or flip a negation in ways that pass a quick read. A certified post-editor under ISO 18587 takes responsibility for the final text, which is what makes the output audit-ready for a notified body or procurement review.
What does ISO 18587 cover, and is it changing?
ISO 18587 sets the requirements for full human post-editing of machine translation output and the competences of the post-editor. A revision is underway to broaden it to post-editing of non-human translation output, which brings LLM output into scope, with the revised version expected in 2026.
Is LLM translation compliant with the EU AI Act?
The EU AI Act (Regulation 2024/1689) sets the legal rules, and ISO 42001 gives you the management system to meet them repeatably. Running LLM translation under ISO 42001 and ISO 27001 produces the risk management, data governance, and human oversight evidence the Act expects. Certification is not automatic legal compliance, so the audit trail still matters.
When does NMT still make sense?
When you've already invested in a domain-tuned engine, when content is high-volume and low-stakes, or when you need deterministic output or very low latency. Outside those cases, starting new regulated work on NMT is hard to justify against an LLM plus certified review under ISO 17100 and ISO 18587.