Generated Information

Generated Information validity & value

A research strand by Meng-Han Lee (Zaious)

“AI now does information work that used to take many people many hours.
How much of what it produces actually holds up?”

Two Sides of One Question

Value is the labour and time that generation replaces — it is what changes the cost of producing information at all. Validity is whether the output holds up. Judging that depends on verifiability — how far the output can be checked.

The two sides are bound together, and the binding is a strain: the more work a system replaces, the less anyone can still check its output line by line. So validity cannot be left to review after the fact; whatever makes the output checkable has to be designed into the way the content is generated. Designing that in does not settle validity by itself: a claim can carry a real source that does not support it.

Three Kinds of Replaced Work less to check, further down

Code — checked by tests that run

Rules and logic that people used to write by hand. The output can be executed, so it can be held to cases with known answers.

Writing — checked against sources

Content that people used to research and write. There is something to check it against, but only if each claim keeps a trail back to where it came from.

Observation — often nothing to check against

Collecting and observing real people — which is what synthetic data and AI personas stand in for. There is usually no ground truth, only calibration against small real samples, and the added risk that a model describes a population out of its own assumptions.

Synthetic data belongs to the third kind, where there is least to check against.

Where People Come Back In

Generation takes the person out of producing the information, but not out of receiving it. People are still the end that information arrives at: part of what someone knows, and part of what they are taken to be, is now decided by machine-produced content. That is where the main question of this research — how does designed interaction constitute the human? — enters.

An AI persona is the most direct case. It is itself a designed person, so its validity question becomes whether a designed construct is close enough to a real one.

Public Anchors so far

Journal Atlas: Venue Fit Scoring Under Incomplete Evidence, and the Corpus Method It Requires

Writing, checked against sources. The scorer treats fit as a claim with an evidence strength: a dimension with no evidence returns unknown rather than a midpoint, and missing evidence can lower a score but never raise it. The corpus is built with a language model that is not allowed to invent: every fact carries a source URL, or is left blank with the reason recorded. A spot-check of 30 sourced claims (2026-08-15) shows why having a source and holding up are kept apart: twelve sources would not open for an automated client, and of the eighteen that did, fifteen supported their claim, one in part, and two did not.

Status
Published · Proceedings of JCDL 2026, Posters & Demonstrations · 2026-10-13

riftbound-chronicle

Code, checked by tests that run. Card rules and action-legality code for a trading card game, written with LLM assistance and held to case-based evaluations and schema validation.

Status
Open-source repository

Taiwan.md

An open-source, AI-friendly knowledge base about Taiwan, which he takes part in as a collaborator.

Role
Collaborator since 2026-03, not a maintainer

Only public work is listed. This page describes a research direction — it is not a finished methodology, and it gives no figure for hours saved because none has been measured.

Where This Sits in the Programme

A strand alongside the three-axis research programme, not a fourth axis. The programme asks how designed interaction constitutes the human; this strand asks what happens to that question once much of the information a person meets is generated rather than written. See the full programme ↗