If you are building LLMs into production infrastructure, take notes from Canada’s federal government. The GCTranslate tool, an AI-powered translation service deployed by the Canadian government, has been live for approximately six months. According to documents obtained by CBC/Radio-Canada via a Freedom of Information Act request, the system is generating significant hallucinations and mistranslations in official government documents, particularly when converting English to French.

Hallucinations in Production

The error rate in GCTranslate is not just cosmetic; it represents a failure in context awareness and entity resolution. The tool has been documented translating the abbreviation 'TB' for the transportation bureau into 'Tuberculosis.' More critically, it has altered proper nouns entirely, changing a person’s name from 'Mike' to 'Stephanie.' These are not minor syntax errors; they are fundamental failures in named entity recognition that would break most enterprise data pipelines. The semantic drift extends to policy language and technical manuals. The 'Buy Canadian Policy' was rendered as 'Achetez une politique canadienne,' which literally translates to 'Buy a Canadian policy,' completely missing the protectionist economic intent. In a more visceral example, the shorthand for 'cumulative growth' was translated to 'sperme croissance,' a phrase that is both medically incorrect and contextually absurd in an economic report.

Safety and Accessibility Risks

Experts quoted by Radio-Canada warn that these errors have real-world consequences for infrastructure and public safety. An inaccurate translation of a tax form could exclude French-speaking citizens from specific credits or benefits, creating an accessibility gap in digital government services. More alarmingly, poor translations of instruction manuals for weapons or heavy equipment could lead to physical injury or worse, as safety warnings lose their precision.

The Human-in-the-Loop Mitigation

It is worth noting that GCTranslate is currently a pilot project and is receiving regular updates. The government is employing human translators to assess output quality and retrain the models, a classic human-in-the-loop feedback mechanism. However, for developers, this serves as a stark reminder that even with active retraining, LLM-based tools in critical infrastructure require rigorous guardrails and verification layers before full autonomy is granted.

Key Takeaways

  • GCTranslate has been live for six months with persistent hallucination issues in English-to-French translation.
  • Errors include entity mismatches (Mike to Stephanie) and abbreviation failures (TB to Tuberculosis).
  • Semantic drift poses risks to public safety in equipment manuals and financial equity in tax forms.
  • The project relies on human-in-the-loop retraining to mitigate current model deficiencies.

The Bottom Line

Shipping an LLM into government infrastructure without robust entity validation is a recipe for disaster. If you wouldn't let a junior dev push unreviewed code to prod, don't let your AI hallucinate safety warnings.