The editing exemption needs an input boundary

Editing systems operate on a continuum. They correct spelling, rewrite tone, translate, synthesize missing frames, replace voices, generate illustrations, and restructure arguments. A product label such as editor does not establish where assistance ends and synthetic generation begins. The boundary depends on the input, the transformation, and the resulting change in content or meaning.

Article 50(2) includes limits related to assistive standard editing and cases that do not substantially alter supplied input or its semantics. The Commission's transparency guidelines provide current guidance as the obligation approaches application. An operational system needs evidence that supports its scope decision without pretending to resolve every interpretive question automatically.

The input must remain part of the record. Without it, a later reviewer sees only the output and a product mode. They cannot tell whether the system corrected a sentence, invented a paragraph, translated faithfully, or replaced the original claim. Storing everything indefinitely is unnecessary; a proportionate receipt can preserve hashes, transformation class, bounded diffs, and higher-detail evidence for disputed or high-reach cases.

Classify the transformation, not the feature name

Define transformation classes with examples and counterexamples: surface correction, format conversion, translation, semantic rewrite, generative completion, substitution, and mixed-origin composition. Product teams can map operations to the classes, while legal owners determine which classifications support a particular obligation or exception. Keep the mapping versioned because model behavior and product affordances change.

A percentage-of-bytes metric is not enough. One generated sentence can reverse the meaning of a long document. A small facial change can alter identity. A brief synthetic audio segment can change the apparent speaker. Evaluate semantic and contextual consequence, and route uncertain cases to a more conservative marking or review path.

The name of the tool does not decide whether the meaning changed; the input-to-output transformation does.

Mixed workflows need compositional receipts. A human may draft text, use a model for structure, manually revise it, generate an image, and export through another service. The final artifact should not inherit the least demanding classification from any one step. Preserve component origin and the transformations that matter to the final transparency claim.

User interfaces should avoid asking creators to make legal classifications they cannot evaluate. Ask observable questions where human input is necessary: Did you supply the original recording? Was a real person's appearance changed? Did generated text introduce new factual claims? Combine those answers with system telemetry and present the resulting disclosure before publication.

Disputes are valuable test cases. When a creator, subject, platform, or audience challenges a classification, preserve the input, output, rule version, and decision. Add representative disagreements to the regression corpus. The objective is not to eliminate judgment but to prevent the same boundary from being improvised differently each time.

Make the obligation operational

Begin with the observable input, transformation class, semantic-change decision, and mixed-origin composition behind an editing-scope claim. Express it as a control object rather than a policy summary: scope, triggering condition, applicable system or model version, permitted exception, effective time, evidence source, and the consequence when the control cannot establish compliance. This lets engineering, product, legal, and operations examine the same boundary without pretending their responsibilities are interchangeable.

The minimum receipt should retain input and output identifiers, operation sequence, bounded diff or component manifest, classification rule version, user declarations, uncertainty, and reviewer decision. Keep the record proportionate and protect confidential information, but make it possible to determine which rule, artifact, system version, and accountable decision governed the event. A folder of undated screenshots may show that work occurred; it rarely proves that the operative control held for the affected release.

Test the implementation by applying surface edits, translations, semantic reversals, generative completions, identity changes, and mixed workflows while comparing classifications and disclosures. Include ordinary cases, boundary cases, degraded dependencies, and known exceptions. Preserve the starting state, observed output, machine-readable evidence, user-visible result, and any human intervention. Re-run the test after changing a model, content pipeline, interface, standard, provider, or policy interpretation.

The product-scope owner with the transparency policy owner should decide whether the evidence supports continued operation, a narrower scope, a compensating control, or a hold. The owner needs authority over the affected release and access to the evidence. Record unresolved interpretation separately from a technical defect so an engineering patch does not masquerade as a legal conclusion.

Monitor both presence and effectiveness. A marker can exist but be stripped downstream. A disclosure can render but arrive after exposure. A document can be submitted but refer to an obsolete model. Pair a control-presence measure with a consequence or comprehension test, give the claim a review date, and reopen it when a dependency changes.

Maintain a dependency register for the control. Model endpoints, editing pipelines, content formats, user interfaces, identity services, submission portals, vendors, and external standards can change the evidence without changing the policy text. Name which changes invalidate the last test and which monitoring signal proves that the dependency remains inside the reviewed state.

Exercise the exception path as carefully as the ordinary path. Record who can invoke it, which facts they must supply, how long it lasts, what capability or distribution is reduced, and which compensating evidence remains. An exception without expiry and re-entry criteria becomes a second operating model that can silently outlive the reason it was approved.

Keep public and executive claims no broader than the tested boundary. Say which systems, releases, formats, routes, and dates the evidence covers, and identify material exclusions. When a control fails or a dependency moves, update the claim and the remediation record together. A transparent limitation protects more credibility than a universal statement built from a narrow passing test.

— Dispatches · Summit Cognitive