Keep two records, with a link between them
When you ask AI to summarize a note, keep the note too. Treat the summary as a separate interpretation that points back to the words it was made from.
This is a practical recommendation, not a finding from a new experiment. It follows Deep Well Labs' distinction between human expression and machine interpretation, described in the Boundary Charter.
A useful summary can help you navigate a long record. It does not become the original simply because it is shorter or easier to read. The distinction matters when you want to check a qualification, revisit an unresolved question, or understand how a later decision developed.
Notice what changed in the retelling
Consider this invented example:
Table 1. Scroll horizontally to see all columns.
| Record | Text | What it represents |
|---|---|---|
| Original note | “I might postpone the launch if testing uncovers a serious problem.” | A conditional possibility, expressed by a person. |
| AI summary | “The launch is postponed because testing found problems.” | A machine interpretation that changes both the decision and the evidence. |
| Human review | “I have not postponed the launch. The condition is still open.” | A later human clarification, kept alongside the earlier records. |
The error is not a spelling mistake. A possibility has become a decision, and a hypothetical problem has become an asserted finding. Keeping the original visible makes that difference inspectable. This example illustrates a failure mode; it does not estimate how often any model makes it.
A workflow you can use today
- Keep the original. Save the words you actually wrote or the recording you intentionally made before asking for a summary. Use a copy for transformations.
- Label the interpretation. Give the output a clear label such as “AI summary.” Record which source it uses. Where available, include the tool, model, and generation date.
- Review consequential wording. Compare decisions, promises, dates, uncertainty, and attributed statements against the original. A fluent sentence is not evidence that it matches.
- Keep your response distinct. If you adopt an idea or correct a summary, save that as your later response. Do not silently rewrite the earlier record to make the sequence look cleaner.
- Keep the connection when exporting. If you move to another tool, retain the original, the interpretation, and the reference between them. Open the export and check that those parts are actually present.
You can do this with ordinary files and links. It does not require a particular AI provider or a Deep Well Labs product.
Provenance helps you inspect, not certify truth
The W3C PROV Data Model provides a technical vocabulary for tracking how records were produced, derived, and attributed. That supports a useful distinction: knowing where a statement came from is different from knowing whether it is correct.
A linked source can itself contain an error. A person can change their mind. A model can misread accurate source material. Preserving the sequence gives you something to examine; it does not settle those questions for you.
The workflow above is Deep Well Labs' practical application of the source-versus-interpretation boundary. It is not a claim that W3C prescribes this workflow or that a provenance standard proves a summary accurate.
What this does not promise
This approach does not reconstruct an unexpressed thought, prove that an original statement was true, or guarantee a correct AI summary. It also does not establish that a particular product currently offers every step described here.
The point is more modest: when you return to your record, you should be able to tell what you expressed, what a machine contributed, and what you decided afterward.
Sources and further reading
- Deep Well Labs and Origo Boundary Charter: the published company distinction between research and product claims, including human source and interpretation.
- W3C PROV-DM: a provenance data model covering records, activities, derivations, and responsibility. Its technical definitions inform the provenance discussion; the example and practical recommendations are this explainer's own.