The model does not know which ambiguity was deliberate

Legal drafting hedges, qualifies and leaves things open on purpose. Language models tidy. That instinct — to clarify, shorten and resolve — is the single most dangerous thing about AI output in legal content.

--:--:-- [PST]
S01

What we work on

  • Contracts and commercial agreements
  • Court filings, pleadings and judgments
  • Evidence and discovery material
  • Corporate governance and compliance documentation
  • Regulatory correspondence
  • Notarised and certified translation
  • Terms, policies and consumer-facing legal text
S02

Why AI output is a specific risk in legal content

Unrequested improvement. A language model asked to translate a clause will often produce a cleaner clause: shorter, clearer, with the hedging removed and the ambiguity resolved. In most content that is an improvement. In a contract it is a substantive change to the parties' position, made silently, by a system that had no idea it was making it.

Negation and conditionality. "Shall not", "unless", "save that", "provided that", "to the extent that" — the structural machinery of legal drafting, and the places where a fluent system occasionally drops or inverts a term without leaving any trace on the surface of the sentence.

Terms of art do not translate as words. Legal concepts are jurisdiction-bound. Some have no equivalent in the target legal system, and the correct treatment is a decision — retain, gloss, adapt — that requires knowing both systems. An engine produces the nearest lexical match and moves on.

Certified translation carries a signature. Where a translation must be certified accurate, someone is personally attesting to it. That attestation cannot rest on an unreviewed machine draft, and no serious translator will give it on one.

S03

What we deliver

  • Reviewers with legal qualifications or documented legal specialisation, matched to jurisdiction and area of law.
  • A separate pass for conditionality, negation and modal verbs — the categories where a silent error changes an obligation.
  • Findings reported, not just corrected. Your counsel sees what was wrong in the machine output, which matters when the same pipeline is producing your next hundred documents.
  • Certified translation where required, with the attestation resting on a documented human process.
  • Terminology held across a matter, so the same defined term reads the same way across every document in a transaction or a case.
S04

Confidentiality

Most legal content is privileged, confidential or both. We will sign your NDA rather than push ours. If content cannot pass through public cloud infrastructure — which is common in this sector and increasingly a procurement requirement — tell us at the start and the job is configured accordingly.

S05

Languages where this matters most

Arabic, for Gulf and North African legal and government work, where the register and variety decision is significant and where machine output is weakest on the best available evidence. German, for DACH commercial and regulatory documentation. Japanese, Korean and Chinese for cross-border corporate and IP matters. And the European set for multi-jurisdiction transactions.

S06

Talk to us about legal content

Send a clause set or a filing. We will return it reviewed, with anything that changed an obligation flagged as a critical finding.

Talk to us about legal content