# Writing provenance versus AI-text detection Writing provenance and AI-text detection answer different questions. A process record describes observed actions while a document was produced. An AI-text classifier estimates a label or score from the text supplied to it. Humanly's central evidence is the recorded writing process and its configured context; its source percentages are not AI-text classification probabilities. ## Compare the evidence, not just the score | Question | Recorded writing provenance | AI-text classification | | --- | --- | --- | | What input is needed? | Activity recorded during a supported writing session | A passage or finished document supplied for classification | | What can the result describe? | Observed edits, input sources, in-platform assistance and task settings | A model's assessment of the supplied text | | Can it reconstruct an earlier unrecorded draft? | No; importing a final document does not recover its missing history | A classification does not reconstruct the editing history | | What remains uncertain? | Unobserved assistance, idea origin, intent and other outside activity | The actual sequence of actions that produced the text | | What must a reviewer add? | Task rules, context, the writing itself and interpretation of coverage | The classifier's validation conditions, uncertainty and independent context | This comparison describes the approaches, not a ranking of vendors. A specific product may combine multiple forms of evidence. Check what a product actually observes and what its report claims before treating similarly named scores as equivalent. ## What does Humanly add? Humanly connects writing controls, recorded activity, source attribution and available replay in a certificate. A reader can ask a concrete question, such as whether an AI suggestion was applied to a paragraph or how a passage developed. The certificate also preserves context about the writing environment. A shared certificate's full text and history depend on its display settings. A reader should identify which evidence is actually available before interpreting a summary. See [what certificates can show](/reference/evidence) for the difference between final-text composition, process totals and the integrity seal. For example, a writer may type an initial argument, paste a quotation, accept permitted polishing and delete several earlier paragraphs. A process view can distinguish those recorded actions. That does not establish that the argument's ideas were formed without outside help or that the quoted material was cited correctly. ## Is Google Docs version history enough, or do I need a certificate? Version history can be useful when the question is how a document changed. A Humanly certificate is relevant when a review also needs the configured writing rules and recorded input sources. Choose the evidence that answers the actual question, and arrange any Humanly recording before drafting begins. | Review need | Google Docs version history | Humanly certificate | | --- | --- | --- | | Inspect revisions | Earlier versions and the people who made changes | Available recorded activity and replay, alongside the writing setup | | Give someone access to history | Browsing earlier versions requires permission to edit the file | A shared certificate exposes history only when enabled, with an optional access code | | Interpret assistance | Revision changes provide drafting context; they do not establish the origin of ideas | Recorded typed, pasted and in-platform AI-assisted input provides additional context, with outside assistance still unobserved | Google's [version-history documentation](https://support.google.com/docs/answer/190843?hl=en) explains viewing, restoring and naming earlier versions, and notes that revisions may occasionally be merged. These are useful collaboration features; the comparison does not assume that a missing revision means improper writing. Neither an editing history nor a certificate establishes independent authorship by itself. Importing a finished Google Doc into Humanly does not turn its earlier revisions into a Humanly process record. See [certificate coverage and sharing](/reference/evidence) before choosing what to collect or share. ## Is Humanly's typing detector an AI-text detector? No. Auto Typing Detector evaluates recorded writing behavior for signs of automated typing. Its object is the interaction pattern, rather than whether the final prose resembles AI output. Abnormal Behavior uses separate event rules. Either kind of signal can require further review, and a detector may be disabled or lack enough usable evidence. Research performance for a typing detector does not establish accuracy for detecting AI-written prose. The [research guide](/reference/research) provides the Humanly paper's study context and limitations without treating benchmark results as guarantees for an individual writer. ## Which approach fits the decision? If you can arrange the writing workflow in advance and need to inspect how work developed under stated permissions, a configured process-recording environment is relevant. If only a finished document exists, its earlier Humanly activity cannot be created retrospectively. Keep any later classification separate from claims about an observed history. Neither approach should substitute for reading the work and considering the applicable rules. A course that allows AI assistance needs a different interpretation from a task that disallows it. A typed-input record is not proof of intellectual independence, and a valid certificate signature verifies protected records rather than authorship truth. The [Humanly product discussion](https://writehumanly.net/research/beyond-post-hoc-detection) explains the motivation for recording process evidence. Start with the [product guide](/reference/product) for available workflows or [AI policy](/reference/ai-policy) for their boundaries. Consult the [privacy policy](https://writehumanly.net/privacy) when deciding what process evidence is appropriate to collect or share. --- Reviewed: 2026-09-24. Applies to: Humanly writing workflows reviewed on 2026-09-24. Available controls and evidence depend on the writing setup. Research findings refer to arXiv:2607.21758v1. Canonical: https://writehumanly.net/reference/provenance-vs-ai-detection [Markdown version](/reference/provenance-vs-ai-detection/index.md) ยท [All reference topics](/reference)