Analysis of legal texts

Official documents — contracts, non-disclosure agreements, public offers, and regulations — are one of the most difficult tasks for verification algorithms. They are written in a specific language that is as far removed from everyday human speech as possible. It is dry, full of repetitions, strict formulations, and passive voice.

But there is a significant difference between a real legal text, verified by generations of lawyers, and an imitation generated by an algorithm in a few seconds.

How the algorithm works on complex documents

Legal language has been formed over decades. It has heavy, but strict and established constructs. Popular neural networks are trained on more modern, universal Western standards. Therefore, when an algorithm is tasked with writing a "contract," it often makes stylistic errors.

• 1. Corporate cliché syndrome.
  A real service agreement will use constructs like "The Contractor undertakes to perform works of appropriate quality". An algorithm, trying to maintain an official style, often slips into marketing pathos: "The Parties undertake to ensure seamless integration of innovative solutions to maximize efficiency". For a professional lawyer, such formulations are ordinary semantic water with no force.
• 2. Structural symmetry.
  A real law or offer is built unevenly depending on the importance of the clauses. The section on liability can take up three pages of small text, and the section on deadlines — one line. The machine strives for symmetry: it will generate five paragraphs of three lines each, trying to visually balance the document.
• 3. Over-explanation of the obvious.
  In real contracts, all terms are fixed once at the very beginning (in the "Terms and Definitions" section). A neural network, on the other hand, may start additionally explaining the meaning of the word "Force majeure" in parentheses right in the middle of a paragraph so that the reader can more easily understand the meaning. This is not done in business correspondence.

Confirmation through primary source search

As in the case of directories, the main tool for confirming the authenticity of a document is checking open databases. The Orhuman analyzer always runs the text through search engines before issuing a final verdict.

If a complex and dry text is part of a real normative act published on official legal portals or government websites, the system does not mark it as artificially generated. The text is assigned the status "Editorial text".

Thus, we distinguish a real verified law from a plastic imitation created by a neural network in a couple of seconds.