A large Greek manufacturing company operates multiple facilities across the country. Head office develops workplace safety policies in Greek, including detailed procedures for end-of-shift protocols. One policy specifies when workers must perform a critical safety step: “Μετά το πέρας της βάρδιας”: after the end of the shift.

At a facility in a different part of the country employing significant numbers of foreign workers, local HR staff receive the new safety policy from head office. They use generative AI to translate it into English. AI renders “Μετά το πέρας της βάρδιας” as “at the end of the shift.” The English version is distributed to foreign workers and sent to the external contractor hired to deliver safety training in English. No one at the facility checks the translation against the Greek source. The compliance team at head office never sees the English version. Translation is treated as an administrative task, not a safety matter requiring technical review.

The foreign workers can’t read Greek. They rely entirely on the English translation to understand their safety obligations. The external trainer uses the translated document to deliver the training. The trainer doesn’t read Greek either. Workers are instructed to perform the safety step “at the end of the shift” while the machinery is still running, while hazards remain active, during the final moments of operational activity. The Greek instruction said to wait until after the shift ends. The English version says to act while the shift is ending. One preposition moved the compliance window into a hazard zone.

The safety risk

The timing of a safety instruction defines when a worker is protected. “After the end of the shift” means the safety step occurs when equipment has been powered down, when moving parts have stopped, when the work environment has transitioned from active to safe. The instruction assumes hazards have been neutralised before the worker acts.

“At the end of the shift” relocates that action to the closing moments of work when machinery is still operational, when the environment remains hazardous. The worker performs the same physical action, but in a different risk context. The mistranslation doesn’t change what the worker does. It changes when the worker is exposed.

This is how language determines the scope of safety obligations. Words specify the conditions under which a duty arises. Legal scholars examining statutory interpretation have documented how temporal markers (words like “after,” “during,” “before”) define when legal obligations apply and when they don’t[1]. When those markers are mistranslated, the obligation shifts to a different point in time, governed by different factual conditions.

In workplace safety, that shift can mean the difference between a controlled environment and an active hazard zone.

The legal problem

The company’s compliance documentation shows workers followed the English procedures. Inspectors review training records, incident logs, and safety audits; all based on the translated policy. Everything appears compliant.

But compliance with the English version is not compliance with the policy the company actually issued. The Greek source required one action. The English translation required another. Workers performed the action the translation instructed them to perform, not the action the company’s safety management system required.

If an injury occurs, the company cannot demonstrate that its workforce followed the procedures designed to mitigate the specific risk. The procedures workers followed were created by the translation error, not by the company’s safety assessment. The company has lost the ability to show that the control measure it designed was actually implemented.

The translated policy became a shadow document; an instruction that looks authoritative, that workers rely on, that trainers use, but that doesn’t match what the organisation requires under its own safety framework.

Why it happened

The output was fluent. The sentence structure was natural. “At the end of the shift” sounds plausible in English. No one checked the translation because they assumed AI handles this correctly. That assumption may appear reasonable. Translation tools have indeed improved. They remain problematic. They remain tools. And critically, most users don’t have the expertise to verify output in languages they don’t speak. So they can’t spot mistakes like this.

The error wasn’t visible to anyone in the chain: not to the HR department that distributed the policy, not to the contractor who delivered the training, not to the workers who implemented it. The only person who could have caught it would need to read both Greek and English, understand workplace safety terminology in both languages, and recognise that the temporal relationship between the safety step and the end of the shift had changed.

That’s not a linguistic skill. It’s a specialised professional competence that sits at the intersection of translation, legal interpretation, and occupational safety.

What this means for compliance

When legal and safety documents cross language boundaries, the translated text determines how obligations are understood and implemented. If the translation changes when a duty applies, it changes what compliance looks like in practice.

This isn’t unique to workplace safety. Contracts, regulatory filings, court judgments, and corporate policies all rely on precise language to define when parties must act, when rights vest, when duties terminate. Mistranslation of temporal language (or indeed any language that defines the conditions under which an obligation arises) creates a gap between what the document requires and what people believe it requires.

That gap invalidates the compliance chain. You can’t prove you followed a procedure if the procedure you followed was created by translation error.

Professional legal translators understand this. We know that translating “μετά” as “at” instead of “after” isn’t a minor linguistic choice. It’s a change in when the legal or safety obligation applies. We verify temporal markers, cross-check procedural language, and flag discrepancies that shift the scope of what’s required.

AI tools don’t do that. They produce fluent output that sounds correct. They don’t understand that one preposition determines whether a worker is safe or exposed. They don’t know when “close enough” isn’t close enough.

For companies relying on AI translation for legal or safety documents, the risk is that the translation becomes the policy. And that nobody notices until something goes wrong.

[1] see for example R. H. Fallon, “The Meaning of Legal ‘Meaning’ and Its Implications for Theories of Legal Interpretation,” University of Chicago Law Review 82:1235