Neural Translation: When the Machine Oversteps the Mark
Neural translation has become the obvious choice. Fast, fluent, and cost-effective, it is now widely used by businesses to translate contracts, reports, technical documents, and strategic content.
However, behind this promise of performance, one essential question is all too often overlooked: what actually happens to your data once it has been translated?
And above all, who has access to it?
This is precisely where neural translation can become transgressive.
Neural Translation and AI: A Dangerous Confusion
Before going any further, it is essential to clear up a common misunderstanding.
Neural translation is not, by definition, synonymous with global artificial intelligence or data sharing. However, mainstream neural translation tools almost always rely on connected, shared, and scalable architectures.
In other words, in many cases, the translated texts are:
- sent to remote servers,
- stored temporarily or permanently,
- analysed to improve the models,
- sometimes reused to train the AI.
According to a study published by Gartner in 2024, 72% of companies do not know exactly how their data is processed by the AI tools they use.
This lack of awareness creates a major risk.
When the Machine Oversteps: The Notion of Transgression
In a professional context, a machine becomes transgressive when it goes beyond the scope for which it was authorised.
This is exactly what happens when internal, confidential, or sensitive documents:
- leave the company’s perimeter,
- are integrated into global models,
- escape the control of the client.
The CNIL regularly reminds us that the use of uncontrolled AI tools can lead to breaches of the GDPR, particularly regarding purpose, retention, and transfer of data.
In certain sectors — legal, defence, industry, healthcare, finance — this type of data leak can have serious, even irreversible, consequences.
Local Neural Translation: An Alternative Approach
In response to these risks, an alternative exists: local neural translation, reserved for a single client.
At ALPIS, this approach is based on several non-negotiable principles.
First, the neural engines are deployed locally, with no connection to global AIs. Next, each client has their own dedicated translation memory, strictly private and not shared. Finally, no data is ever used to train an external model, now or in the future.
In this way, translated documents:
- remain within the client’s environment,
- are never shared,
- never leave the secure environment contractually agreed.
In this specific context, it is no longer AI, but a controlled linguistic tool, serving the client and the client alone.
Why ALPIS Does Not Use AI in This Context
It is important to state this clearly.
At ALPIS, we do not use mainstream AI, and this is a deliberate choice.
Not out of a rejection of technology, but because confidentiality, data sovereignty, and legal responsibility take precedence over speed or apparent cost.
According to IBM’s Cost of a Data Breach 2023 report, over 9 billion sensitive data records have been exposed or reused via poorly controlled automated systems, often without the knowledge of the companies concerned.
In this context, using a global AI to translate sensitive documents is to accept a risk that many management teams still underestimate.
Professional Translation and Security: A Strategic Choice
Choosing a translation provider is no longer just about linguistic quality. It is now a strategic decision, at the crossroads of cyber security, regulatory compliance, and the protection of information assets.
This is why ALPIS prioritises:
- local neural translation,
- dedicated translation memories,
- total absence of data sharing,
- and strict control of data flows.
In other words, technology is placed at the service of trust, not the other way round.
Conclusion: The Real Question Is Not Performance, But Control
Neural translation is not dangerous in itself.
What is dangerous, however, is not knowing where your data goes, who uses it, and for what purpose.
At ALPIS, the answer is simple:\
your data stays with you.\
It is never shared.
It is never fed into any global AI.
And it is precisely for this reason that, in this context, we do not use AI.