- Alibaba researchers in 2025 found LLM translation hallucination rates between 33% and nearly 60% across 17 models and 11 language pairs, with errors that pass undetected because the output reads fluently.
- A 2024 Midwestern hospital found critical terminology errors in 47 patient-facing documents after a vendor switched to consumer-grade AI translation tools without disclosure.
- A 2026 Intento report found LLM translation carries a median of 31 errors per 1,000 words; adding human post-editing and quality assurance raises costs to around $590 per million source characters, making the headline per-character AI saving largely irrelevant for high-stakes content.
LLMs hallucinate in translation at rates between 33% and nearly 60%, according to a 2025 Alibaba study covering 17 models and 11 language pairs. The errors are hard to catch because they read well. That dynamic played out in concrete terms in 2024, when a Midwestern hospital discovered critical terminology errors in 47 patient-facing medical documents, traced to a vendor that had quietly switched to consumer-grade AI translation tools without telling anyone. Reviewers had flagged nothing, because the output sounded fine.
The Fluent Hallucination Trap
A model optimised for fluency will generate a confident, well-formed sentence even when the underlying meaning is wrong. That is not a bug in the usual sense, it is the design goal working against the use case. A May 2024 arXiv paper by Saadany et al. found that standard quality metrics do not penalise these errors, which means automated quality gates routinely pass translations containing what the paper calls “catastrophic meaning distortions.”
The English word “sanction” is a useful example. It can mean either to approve or to penalise depending on context, and a model trained on general-purpose text will pick one reading and translate it fluently. The output looks valid. The meaning may be inverted. That gap between surface fluency and semantic accuracy is largely invisible to non-expert reviewers working without a source-text comparison.
Where Errors Get Expensive
Legal translation carries significant exposure. A misconstrued contract clause can shift liability, and something as mundane as decimal conventions creates genuine ambiguity across languages: “1.000” means one thousand in some countries and a number between one and three decimal places in others. Law Surface, a legal technology provider, uses AI for an initial translation and analysis pass, followed by review from a legal expert, on the basis that the organisation, not the tool, is accountable for the output. The 2024 Air Canada tribunal, while not a translation case, confirmed that principle: organisations are responsible for AI-generated content errors regardless of which tool produced them.
Low-Resource Languages Bear the Most Risk
AI translation models perform worse than professional interpreters in less common languages while performing comparably in widely spoken ones such as Spanish. That performance gap tracks training data volume almost directly. The communities most dependent on translated public health and legal information are also the ones least well served by current models, a structural problem that better fluency scores do not address.
Idioms and culturally specific expressions compound the issue. Machine translation trained on formal text tends to flatten or mistranslate informal registers, and studies have flagged this as a recurring challenge for users in lower-resource language communities, though the research does not yet quantify how often those failures carry material consequences.
The Post-Editing Economy
Machine Translation Post-Editing, where human linguists review and correct AI output, has become the standard industry response. A May 2025 Acolad survey found that most translators expect demand for pure human translation to fall while demand for post-editing grows. The economics are not straightforward. SnapIntel reported in April 2026 that MTPE rates per word run 30% to 50% below full human translation rates, compressing translator earnings. For dense technical content with poor AI output quality, a post-editor still needs to verify terminology, check for hallucinations and confirm omissions, work that per-word rates do not always reflect.
A May 2026 Dynamic Language report identified a governance gap running alongside the quality problem: enterprise buyers often do not scrutinise how language service vendors use AI internally, which creates liability exposure when AI tools are deployed without the buyer’s knowledge or adequate human review. The hospital incident is one example of how that gap materialises.
What the Cost Numbers Actually Show
A March 2026 Lokalise report put traditional human translation at roughly $150,000 per million words, against about $5,000 per million words for an orchestrated AI workflow with contextual grounding. The 2026 Intento report adds necessary context: the median LLM solution costs about $4.59 per million characters versus $24.99 for traditional MT engines, but LLM output carries a median of 31 errors per 1,000 words weighted by severity, compared to 58 from MT engines.
That error rate is what closes the cost gap. A two-step solution combining AI translation with post-editing costs roughly 2.5 times as much as unreviewed AI output alone, and a three-step solution adding quality assurance runs around $590 per million source characters. For high-stakes content, the headline per-character AI cost is largely irrelevant, the actual spend is dominated by the human review layer on top of it.
Metrics That Miss the Point
Standard evaluation metrics optimise for fluency, not fidelity. The Saadany et al. paper is explicit: commonly used metrics do not penalise critical errors, and the gap between a high quality score and a safe translation can be substantial. In response, researchers have begun developing word-level critical error detection approaches that identify the precise location of an error within a translated sentence rather than flagging the sentence as a whole.
For teams deploying AI in production workflows, existing CI pipelines built around standard metrics are measuring surface attributes while semantic errors pass through undetected. Without retraining error detection models on domain-specific datasets, the quality gate is not measuring what it needs to measure.
Building a Reliable Hybrid
Anglia Translations discloses AI use to clients and assigns a specialist to select appropriate linguists and tools for each project, requiring client consent before any AI involvement and declining to offer raw machine translation without that understanding. The operational logic for high-stakes content is less ambiguous than the cost tables suggest: AI without review may be acceptable for internal, low-stakes material where a misreading carries minimal consequence. For contracts, clinical materials and patient safety information, expert post-editing is the floor. For court filings and regulatory submissions, full human translation remains the benchmark.
The 2025 Alibaba hallucination data, read alongside the broader pattern of AI safety gaps going undetected until an incident surfaces, points to a governance question that sits upstream of model quality: who is responsible for knowing when AI translation was used, and what reviewed it?
