Neural Translation Grows Up: A Field Test Against Human Translators in Three Business Domains
Photo: Staff Sgt. Tracci Dorgan, Public domain, via Wikimedia Commons
The Question Every Buyer Is Asking
Since neural machine translation (NMT) replaced phrase-based systems around 2016, the pitch has been relentless: translation quality approaching human parity, at a fraction of the cost. For businesses allocating localization budgets, the question is no longer whether NMT is impressive — it plainly is — but where, specifically, it can be trusted. To find out, we reviewed engine output against professional human translators across three document types with very different quality demands: marketing copy, legal contracts, and technical documentation.
Marketing Copy: Fluency Without Judgment
On raw fluency, NMT is startlingly good. Headlines, product descriptions, and social copy come out of the engines readable, natural, and free of the stilted phrasing that characterized older systems. For internal-facing content or first-draft exploration, the productivity gain is real.
Where machines fail is persuasion. Marketing translation is transcreation: it requires deciding what the sentence should accomplish, not merely what it says. One test slogan about reliability translated accurately but flattened a deliberate wordplay into blandness — grammatically correct, commercially dead. Human translators routinely made different, better choices, restructuring sentences around target-audience idiom. NMT also showed inconsistent register, drifting between formal and casual within a single landing page.
Verdict: usable for drafts and low-stakes content; a liability for brand voice.
Legal Text: Precisely Wrong Is Worse Than Roughly Right
Legal documents exposed the sharpest divide. NMT handled boilerplate — confidentiality clauses, standard recitals — with high fidelity. But in provisions where meaning turns on a single word, the engines produced confident, plausible errors: obligation-level verbs softened, defined terms rendered inconsistently across the same document, and jurisdiction-specific concepts translated literally into terms that do not exist in the target legal system.
The deeper problem is that machine errors in legal text are invisible. A human reader cannot distinguish a faithful translation from a confidently wrong one without consulting the source — which defeats the purpose. In a domain where ambiguity costs six figures, the economics are unambiguous: human legal translation, ideally with subject-matter specialization, remains non-negotiable.
Verdict: acceptable for internal comprehension and due-diligence triage; unacceptable for execution documents.
Technical Documentation: The Sweet Spot
Technical content proved to be NMT's strongest showing. Manuals, specifications, and knowledge-base articles feature controlled vocabulary, repetitive structures, and limited syntactic variety — exactly the conditions neural engines handle best. With a termbase enforcing consistent product terminology, engine output for some segments was indistinguishable from human translation, and post-editing time dropped dramatically.
The residual weaknesses: numbers and units occasionally reformatted to the wrong locale convention, and cross-references that broke when sentence structure changed. Both are caught cheaply by a light post-editing pass.
Verdict: production-ready with post-editing; the clearest cost case in the test.
The Hybrid Workflow We Recommend
The data does not support choosing between machines and humans. It supports matching the tool to the risk profile:
- Tier 1 — High-stakes content (contracts, regulatory filings, financial disclosures): professional human translation with legal or domain review. No exceptions.
- Tier 2 — Brand-facing content (campaigns, web copy, PR): human transcreation, with NMT as a research and drafting aid only.
- Tier 3 — High-volume, low-risk content (technical docs, support articles, internal communications): NMT with human post-editing, backed by a maintained termbase.
The Bottom Line
Neural translation in 2018 is a genuinely transformative tool — and genuinely unsuited to everything. Organizations that treat it as a substitute for professional translation will discover the difference in the most expensive way possible: in a marketplace, a courtroom, or a customer's inbox. Those that build tiered workflows around it will capture the cost savings without inheriting the risk. The technology decides speed and scale; humans still decide meaning.
