AI data market
Why records created before generative AI may be worth more
By SourceX Editorial · Updated
Short answer
Records created before generative AI tools entered everyday work may be worth more because they are very likely written entirely by people, which model developers value as AI-written text spreads. The advantage depends on proof: keep original creation dates through migrations, and record when each of your systems first offered AI drafting.
Key takeaways
- Records from before AI drafting tools carry a cleaner human signal because no writing assistant shaped them.
- Age alone adds nothing; old records about retired products or obsolete processes may teach little.
- Original timestamps, audit logs and full commit history are the evidence, and migrations that reset dates destroy it.
- Record the date each system enabled AI drafting so archives can be split into before and after segments.
Why might records from before generative AI be worth more?#
Records from before generative AI may be worth more because they are almost certainly human-written. Support replies, code reviews, sales emails and engineering notes from that period show how people reasoned and wrote without a drafting assistant, which is the signal model developers want to learn from.
Researchers have studied what happens when models are trained repeatedly on text produced by other models, a problem often called model collapse, in which rare patterns fade and output narrows. Later work suggests the effect is much weaker when original human data is kept alongside synthetic data rather than replaced by it, which is itself an argument for preserving human records. As more workplace text is drafted or polished by AI, verified human records get harder to find, and an older archive is the easiest place to find them.
When people say pre-2023 data, this is usually what they mean. There is no single cutoff date, because companies and tools adopted AI drafting at different times. The useful line is the date your own systems changed.
How far does the low-background steel comparison go?#
The low-background steel comparison goes only part of the way. Steel made before atmospheric nuclear testing carries less radioactive contamination, so it is prized for sensitive instruments. Text written before AI drafting tools is uncontaminated in a similar sense: no model drafted it. Even then the line is not perfectly clean, because spelling, grammar and autocomplete tools suggested words and phrases long before generative AI, though they did not write whole replies or reviews.
The comparison breaks down in a second place. Old steel is useful regardless of what it was originally made for, while old records are useful only if their content still matters. A support archive about a product line retired long ago, or procedures that no longer apply, may be entirely human-written and still of limited use.
Which older records gain the most?#
Older records gain the most when the human reasoning in them is the point. Free text written by skilled people benefits far more than structured fields, which no model would have written anyway.
Linked records gain twice. A pre-AI support thread that connects to the engineering fix and the release that shipped it shows human reasoning and a verifiable outcome, which is more useful than either piece alone.
| Record | Why human authorship matters | Evidence of date |
|---|---|---|
| Support replies and escalation notes | Show how agents explain fixes and handle frustration | Ticket created and updated timestamps, audit trail |
| Code reviews and pull request discussions | Show how engineers critique design and catch defects | Commit and review history in GitHub or GitLab |
| Engineering and product design documents | Capture tradeoffs and rejected options | Confluence or Notion version history |
| Sales and account emails | Show negotiation and objection handling | Mail headers and archive metadata |
| Estimates, inspection and maintenance notes | Record field judgment in the technician's own words | Job system created dates and change logs |
| Order, invoice and status fields | Little gain from age; few structured fields were ever drafted by a model | System timestamps, plus the date any AI auto-classification of categories or priorities began |
When is older not better?#
Older is not better when recency or relevance matters more than authorship. The SourceX Enterprise Data Value Framework treats recency and human-generated signal as separate value drivers, and the two often pull in opposite directions.
Many buyers want both: a long human-written history plus recent records that reflect current practice. Labeling the boundary between them is more useful than choosing one.
| Factor | Favors older records | Favors recent records |
|---|---|---|
| Authorship | Written before AI drafting tools | Mixed authorship unless AI use is tracked |
| Subject matter | Durable skills such as troubleshooting or negotiation | Current software versions, regulations and equipment |
| Process | Long histories that show change over time | Current workflows a buyer's product must match |
| Linkage | Complete if systems were never migrated | Usually better linked in current systems |
How to prove when a record was created#
Proving when a record was created depends on keeping system evidence intact. A buyer cannot verify authorship directly, but it can check that dates come from the system of record and were not reset along the way.
Provenance standards are starting to give AI involvement its own label: the C2PA specification, a provenance standard for media content, added an AI disclosure assertion in version 2.4, released in April 2026. Business records are not media files, so for an operating company the practical version is a short register of which systems introduced AI features, when and for which teams.
- Keep the original created and updated timestamps when exporting or migrating records.
- Preserve audit logs and version history in helpdesks, wikis and project tools.
- Keep full Git history rather than squashed or rewritten repositories.
- Record the date each system enabled AI reply suggestions, summaries or drafting, by team.
- Note migrations, imports and bulk edits that may have changed dates or text.
- Store this evidence with the data inventory so it travels with any package.
What can erase the pre-AI advantage?#
The pre-AI advantage is erased by anything that changes old text or resets its dates after AI tools arrived. Most of these events are routine IT work, which is why they go unnoticed until a buyer asks how the dates were established.
| Event | What it does to the evidence | Safeguard |
|---|---|---|
| Migration to a new helpdesk, CRM or ERP | Created dates may be reset to the import date and edit history dropped | Map original timestamps into the new system and keep the old export |
| AI summaries or rewrites applied to old records | Model text is attached to, or replaces, human-written content | Store AI summaries in separate fields and keep the original text |
| Knowledge base articles revised with AI | An old article now carries newer, mixed authorship | Use version history to keep the pre-rollout revision |
| Squashed or rewritten Git history | Commit dates and review threads lose their original order | Archive the full repository before any history rewrite |
| Bulk edits and re-imports | Updated dates cover the whole archive and hide what changed | Log bulk operations with dates and scope |
Illustrative: splitting a helpdesk archive at the AI rollout date#
Illustrative: a fictional B2B software company has run its support desk in Zendesk for many years and enabled AI-suggested replies for its agents partway through. Its engineering team uses GitHub and Jira, and adopted an AI coding assistant later still.
Before licensing, the support lead and the CTO record the rollout dates from admin settings and change logs. Tickets are split into a before segment and an after segment, and after-segment replies that used suggestions are flagged where the helpdesk records it. Pull requests are split at the coding assistant's rollout date.
The package carries both segments, clearly labeled, so a buyer can choose the human-only history, the recent records or both. Nothing about the older segment had to be rewritten; the work was documenting what already existed.
How does SourceX document record dates?#
SourceX records date ranges, system sources and known AI-tool rollout dates as part of provenance in the SourceX Evidence Packet, alongside licensing rights, permitted use, the privacy record and release authorization.
Because human-generated signal and recency are separate drivers in the SourceX Enterprise Data Value Framework, older and newer segments are assessed on their own merits rather than averaged together. Nothing is shared during the initial assessment, which needs only system names and approximate years of history.
Frequently asked questions
Is there an exact cutoff date for human-only records?
No single date fits every company. AI drafting arrived in different tools and teams at different times, and some employees used outside tools before any official rollout. The defensible cutoff is your own: the date each system enabled AI features, backed by admin settings or change logs, with a margin for informal use.
Do migrated records keep their pre-AI value?
They can, if the migration kept original timestamps, author roles and text unchanged. Some imports reset created dates to the import date or strip history, which makes authorship harder to show. Check a sample of migrated records against the old system or export logs before relying on them.
Are records created after AI tools arrived worthless?
No. Recent records reflect current products, processes and regulations, and many are still largely human-written. Their value improves when you can show which ones involved AI drafting and which did not. Labeling both segments clearly is more useful to a buyer than discarding either, and recent records often link more completely across systems.
Should we stop using AI writing tools to protect future data value?
That decision should rest on how the tools help your teams, not on a possible license. A better step is to log where AI drafting is used, so future records can be labeled. Human edits, decisions and outcomes keep their value even when a draft started with a tool.
How do buyers check whether text was written by AI?
Detection tools exist, but their results on short business text are hard to rely on, so buyers lean on provenance: system timestamps, rollout dates and supplier attestations. Some also review samples by hand, looking for templated phrasing that appears after a rollout date. Clear documentation from the supplier is usually more convincing than any detector score.
Sources
- C2PA specification version 2.4 (April 2026) added an AI Disclosure Assertion (c2pa.ai-disclosure), among other additions. Source
Related resources
See if your company qualifies
A short company assessment. No data uploads are needed.