What is pre-AI human data, and why are older records prized?

What does pre-AI human data mean?#

Pre-AI human data means records created by people before AI writing and coding assistants became part of how that work was done. In a business, that includes support replies typed by agents, estimates written by estimators, code reviewed line by line by engineers and inspection reports drafted by quality staff.

The idea is often compared to low-background steel, metal produced before atmospheric nuclear testing that is sought for sensitive instruments because it carries no later contamination. Older records play a similar role for AI: they are known to reflect human judgment, without machine-generated text mixed in.

Why are older human records prized?#

Older human records are prized because AI developers need text and decisions they can trust were produced by people. As more new content on the web and inside companies is drafted or suggested by AI, separating human work from machine output gets harder.

Researchers have studied what happens when models are trained repeatedly on model-generated content, a problem often discussed as model collapse, in which quality and diversity can degrade. Developers respond by valuing sources with clear human origin. Business records add something public text rarely carries: the decision and its outcome, such as the diagnosis that fixed the unit or the review comment that caught the defect.

Evaluation is a second reason. Developers test models against cases with known answers, and a test set that quietly contains model-written examples can flatter the model being tested. Human records with documented outcomes make cleaner test material.

In the SourceX Enterprise Data Value Framework, that shows up mainly as human-generated signal and domain expertise, alongside rights and AI utility.

Which business records count?#

The business records that count most are those where people wrote the content and made the decision inside a system that recorded when, and by whom. The table lists common examples, what makes their human origin clear and what to check.

Paper and scanned records can qualify too, though they need more preparation: text has to be extracted and checked, and the link between a paper job ticket and its invoice may exist only in someone's memory.

Where is the line between pre-AI and AI-assisted records?#

The line between pre-AI and AI-assisted records is specific to each company and often to each team. A support team may have turned on AI-suggested replies well before engineering adopted a coding assistant, while estimators kept writing everything themselves.

Rather than picking a calendar year, gather the evidence that dates each change:

Where the evidence is thin, draw the line conservatively. Labeling a period AI-assisted when it may not have been costs little, while labeling AI-assisted records as human-written can undermine the representations a license relies on.

• 1Purchase and activation dates for AI features in your help desk, CRM, email client or development tools.
• 2The date of your first AI acceptable use policy and each later revision.
• 3Admin logs showing when AI features were enabled, and for which groups.
• 4Fields or tags that mark AI-suggested or AI-generated content, where your systems provide them.
• 5Written recollections from team leads about when working habits changed.

Records after the line still have a place#

Records created after AI tools arrived are not worthless; they simply need labels. A buyer can treat human-written and AI-assisted periods differently only if the supplier marks them, and mislabeled periods undermine trust in the whole package.

Recency is a value driver in its own right. A ticket archive about a product version retired long ago teaches less about current software than recent tickets, and an estimate history built on outdated pricing or codes needs context. The strongest packages often combine a long pre-AI history showing how experts worked with recent records showing current products and processes, each labeled by period.

How is human provenance documented?#

Human provenance in business records is documented mostly through the systems that created them: timestamps, author or agent IDs, edit histories and audit logs. Those are facts a buyer can check without taking anyone's word for it.

Newer standards aim to record AI involvement directly. The C2PA defines a Content Credential, also called a C2PA Manifest, as a cryptographically bound structure that records an asset's provenance, with assertions about origin, modifications and use of AI. Most business records predate any such marker, so system metadata and a written adoption timeline do that work instead.

Illustrative: a consulting firm marks its line#

Illustrative: a fictional operations consulting firm keeps proposals, statements of work, project review notes and internal playbooks in SharePoint and its CRM. Partners began using an AI drafting assistant after the firm adopted an AI policy, and IT can date the change from license activation records.

The firm splits its inventory at that date. Proposals and review notes before it are labeled human-authored, with client names and confidential client details removed. Later documents are labeled AI-assisted and held for a separate decision. Playbooks are excluded because their edit histories mix both periods and cannot be separated cleanly.

How SourceX treats record age#

SourceX records the date coverage of each record family, and the company's AI adoption timeline, during the Supply step of the SourceX five-step transaction. That timeline becomes part of the provenance section of the SourceX Evidence Packet, so a buyer can see which periods are human-authored and how that was established.

Age is weighed alongside the framework's other drivers. Recency and data cleanliness can pull an old archive's rating down even as human-generated signal pushes it up, and preparation cost rises for scanned or loosely structured history. The fit check uses metadata only, so a company can see how its periods compare before sharing any files.