Applicability assessment

Assistive Editing or AI-Generated Content? Article 50 Decision Guide

The risky question for document AI is not whether a tool uses a model. It is whether the workflow is only standard editing or whether it substantially alters the input data or semantics enough to require Article 50 marking review.

Last verified: 2026-07-27. This is practical evidence guidance, not legal advice.

Article 50 cover showing the boundary between assistive editing and AI-generated document transformation Comparison of assistive editing and substantive AI transformation, both leading to a documented decision

The decision to document

Article 50(2) includes an exception for systems that perform only an assistive function for standard editing or do not substantially alter input data or its semantics. That exception is useful, but it should not live as a Slack answer or a procurement email. It needs a dated decision record.

Answer first: grammar correction and spelling cleanup are the cleanest assistive-editing examples. Summarizers, drafting assistants, tone rewrites, and semantic translators need closer review because they can change meaning, emphasis, or downstream reader understanding.

Practical decision table

WorkflowArticle 50 assessment postureEvidence to keep
Spell-check and grammar correctionOften a candidate for assistive-editing treatment if no substantive meaning is changed.Examples, product behavior description, reviewer sign-off.
Formatting and style normalizationMay be assistive if it only changes layout or style; review if it rewrites meaning.Before/after samples and rule boundaries.
Contract or policy summarizerNeeds documented review. Summaries can omit conditions, change emphasis, or alter semantics.Input/output examples, risk notes, counsel decision.
Translator with tone or structure rewritingNeeds documented review. Translation plus rewriting can substantially alter semantics.Scope definition, sample outputs, reviewer decision.
Paragraph or article drafting assistantUsually not just editing because it generates new text.Generated-output controls, marking approach, release owner.
RAG chatbot answering from documentsReview 50(1) for interaction disclosure and 50(2) if generated text outputs are in scope.First-interaction disclosure, output behavior, evidence owner.

Minimum decision record

  1. Describe the user workflow in plain language.
  2. Attach representative input and output examples.
  3. State whether the output creates new text, condenses text, rewrites text, translates text, or only edits presentation.
  4. State whether the output substantially alters input data or semantics, with reasoning.
  5. Record the applicable Article 50 paragraph and any exception relied on.
  6. Name the product owner, engineering owner, legal/compliance reviewer, and review date.
Procurement risk: "we think it is just editing" is weak evidence. A short decision record with examples is stronger and reusable across customer security reviews, counsel reviews, and release gates.

Where Compliance Glossary helps

The decision guide depends on consistent terms: assistive function, standard editing, input data, semantics, provider, deployer, AI-generated text, and public-interest publication. Compliance Glossary lets legal or compliance approve those definitions once, scan Confluence pages for drift, and export the record.

Use it for evidence, not legal conclusions: keep the approved vocabulary, decision fields, review owner, deprecated synonyms, and page scan results together. Counsel still owns the Article 50 interpretation.

Sources

Compliance for Confluence

See how Compliance for Confluence turns approved terminology into audit evidence inside Confluence.