Responsible AI
LEXA translates work that people care about — literature, scripture, contracts, scholarship. That carries obligations. This page sets out how the system behaves, what it will not claim, and where a human stays in charge.
Designed and developed in the United Arab Emirates · Arabian AI Lab, Dubai
Our principles
Human authority over the final text
LEXA produces drafts, not verdicts. Every book translation ships with a parallel review copy — each source paragraph beside its translation — so a human translator can verify the whole work. Where our own checks find a mismatch, the passage is flagged for a person rather than quietly passed through.
Explainable decisions, not a black box
Each translation is returned with the agent's reasoning: which specialist handled the text, which strategy was applied and why, and the specific choices made on idiom, register and terminology. Users can see why the output reads the way it does.
Data minimisation and retention limits
Text submitted for single translations is processed in memory and not retained as a stored document. Book Mode necessarily stores the working text while a long job runs so it can survive interruption, and that material is retained only for the life of the job and its delivery.
Honest limits on legal and certified work
LEXA does not issue certified or sworn translations. Legal-strategy output is a professional draft and is labelled as such; certified legal translation for GCC jurisdictions is referred to qualified human translators. We do not imply an accreditation we do not hold.
Cultural and religious sensitivity by design
Our agents are built for languages with deep literary and religious registers — Arabic, Urdu, Persian and others. They are instructed to preserve theological and cultural meaning faithfully rather than flatten it into generic phrasing, and to avoid substituting one tradition's assumptions for another's.
Resistance to prompt manipulation
User-supplied text is handled as material to be translated, never as instructions to the system. Content that attempts to redirect the agent is translated as text rather than obeyed.
How the system works
Fifteen language agents, each with its own cultural register and expertise, rather than one general-purpose model answering everything.
The system determines whether a text calls for literal, cultural, poetic or legal treatment, and reports that decision to the user.
Every translation is checked by the agent against three tests — accuracy (nothing omitted or invented), fluency (reads as if originally written) and verity (no fabricated content) — and corrected before it is returned.
Book translations run unattended on our servers: the work is divided at paragraph boundaries with a mathematical check that no character is lost, each segment is saved as it completes, failures are retried, and the finished book is delivered by email. The user can close their device without affecting the outcome.
Completed books are delivered with a side-by-side parallel copy and a completeness report so a human reviewer can confirm nothing is missing.
Known limitations
Being clear about what LEXA is not is part of using it responsibly.
- • Output is a draft produced by AI. Publication-grade and legally binding work should be reviewed by a qualified human translator.
- • We do not provide certified, sworn or notarised translation, and do not present our output as such.
- • Machine translation can err on rare terminology, dialect and deliberate ambiguity. The parallel review copy exists so those errors are findable.
- • Scanned images without a text layer are not processed; we tell users instead of silently returning partial work.
- • Underlying language models are supplied by third parties and carry their own limitations and biases.
Data and security
Translations run on secured cloud infrastructure. Credentials and keys are held in protected server-side configuration, never in the browser or the codebase.
Single translations are not retained as stored documents. Book jobs retain their working text only while the job runs and is delivered.
Accounts are authenticated; administrative data is restricted to the operator. Customers' source texts are not used to train models.
Users are told which agent and strategy handled their text, and what the agent decided.
Questions about how we handle your material?
We answer directly — including from institutions with procurement or compliance requirements.
Last reviewed October 2026 · LEXA Translate by Arabian AI · Dubai, United Arab Emirates