The weekly paper
for curious people
See the connections.
Understand the consequences.
Swipe to see more sections →
Sick of fake news?
Same evidence. Different perspective.
Viewing through
Start with the shortest supported answer, then follow the evidence and the limits together.
What comes into focusIt keeps the claim, its source and the point where certainty ends in one view.
What this view may missA specialist Lens can make one practical consequence easier to see.
Changing the Lens changes what comes into focus. It never changes the evidence underneath.
Issue 003 Analysis 01 · Institutions and AI
A report can survive while the meaning around it disappears. AI can make knowledge easier to find—but it can also repeat one old source until it sounds like a chorus.
What people heard
Once every report is searchable by AI, institutional memory is solved.
Supported distinction
Does retrieval preserve understanding?
Short answer. No. Search can help people find a document, but they still need to know where it came from, whether it is current, what it meant, who acted on it and what happened next.
Choose how to read this
The facts do not change. Each view uses the same published sources and leaves the same questions open.
See through another Lens
Each view notices something useful. None is allowed to stand in for the complete evidence.
Institution
An organisation needs knowledge that survives changes in staff, systems and structure.
What this view explainsWhy a permanent archive is necessary but not enough.
What it may missLosing records does not prove that all useful knowledge was lost.
Current understanding
Government research finds both lost knowledge and examples of lessons being kept and reused.
AI can improve search, learning and some tasks. It can also multiply weak or outdated information.
Five answers drawn from one source are still only one source.
Lens argues that people need the document, the reason behind it and what happened next.
How the pieces connect
The diagrams show which is which. Two events appearing close together does not mean one caused the other, and a risk is not the same as a finding.
Five AI answers agreed—but every one repeated the same outdated policy. They sounded independent. They weren’t.
The report survived
Research across Westminster governments finds that forgetting is not inevitable. Staff turnover, the ability to absorb knowledge, deliberate choices and the stories institutions tell about themselves all shape what survives.
The lost Fitzgerald records show the problem clearly. The institution continued valuable work, yet some original records were not safely preserved. Even a perfect archive would not automatically explain which recommendation changed what, why it mattered or what later evidence challenged it.
The document can survive while the meaning around it fades.
Then AI arrives
Evidence from real workplaces shows that generative AI can improve some tasks and pass experienced workers' practices to less-experienced colleagues. A 19-month health-system study found useful new ways to search for, share and retain knowledge.
The same study also found risks. A model still needs to know which source is current, reliable, relevant, independent or replaced.
Finding information gets easier. Keeping its meaning does not.
A simple test
Lens built a test in which Policy v2 replaces Policy v1, but search returns only the old version. Five agents repeat, confirm, summarise, act on and record that same outdated source.
The result is five answers that agree because they all started from the same mistake. None has an independent, current source.
The test does not show how often this happens in real systems. It shows why five outputs are not the same as five sources.
The central idea
For each important decision, we should be able to trace the evidence, the interpretation, who had authority, what was done and what followed.
Each link needs support. Events appearing close together do not prove causation. Knowing where a document came from does not make it true. Keeping a policy does not prove anyone followed it.
A trustworthy system lets people reconstruct how the answer changed without asking them to trust one person or one machine completely.
The case against us
It can search buried records, preserve why a decision was made, share useful practices, expose contradictions and bring past results into later decisions. Evidence already supports parts of that promise in limited settings.
The outcome depends on the technology, the people using it and the rules around it. AI can multiply good knowledge or bad knowledge.
The real question is not whether AI remembers. It is whether people preserve what the machine needs to understand the record properly.
What we still do not know
Show me the evidence
We separate established facts, reported claims, our own analysis and questions the evidence cannot yet answer.
Government and university research
Queensland Parliamentary Crime and Misconduct Committee
Strategic Organization
Lens test using one replaced policy
Lens interpretation informed by research on organisational memory and learning
Lens comparative assessment
The question
Find out whether saving documents is enough to preserve what an organisation knows, and whether AI helps or makes the problem worse.
Evidence checked
We checked research on how organisations remember and learn, the loss of some Fitzgerald Inquiry records, and a small Lens test using one outdated policy. The test shows how repeated answers can share one bad source. It does not show how often this happens in the real world.
Best-supported answer
AI can help people find information and do some work better. But finding a document is not the same as understanding it. People still need to know where it came from, whether it is current, what decision followed and what happened next. AI can spread useful knowledge. It can also repeat the same mistake at scale.
Why we told it this way
The evidence shows both sides: AI can help people find and share knowledge, but it can also repeat old or weak information. It does not support saying AI will always help or always harm organisations.
The evidence shows both benefits and risks. It does not support saying AI is always good or always bad for organisational memory.
The test shows how one outdated source can produce several agreeing answers. It does not show how often this happens or how much harm it causes.
Checked 28 August 2026 Sources, disagreements and unanswered questions remain visible.
Possible effects · We cannot say how likely
Organisations could demand a visible chain from source to meaning, decision, action and outcome before treating retrieval or repeated AI agreement as understanding. That may make weak shared sources easier to detect, but the approach has not been proven across organisations.
Research and records show variable organisational memory, loss of original inquiry material and both helpful and risky changes when generative AI enters knowledge work.
Making source lineage and decision ownership explicit could prevent several people or systems from mistaking one repeated source for independent support.
Where the connection stopsA bounded demonstration and case studies do not establish how often machine folklore occurs or that one design will prevent it.
How people may respond
Repeated agreement can be interpreted as corroboration even when several answers share one source; visible lineage could change that response.
What the evidence does not showThe demonstration shows the mechanism, not how often it occurs or which continuity design works best.
Outputs retain source versions, decision context and ownership.
Users can distinguish independent support from repeated dependence on one record.
Attention moves from the number of agreeing answers to the quality and independence of their basis.
Teams, reviewers and system owners: Correct the source and propagate the revision through dependent decisions.
A correction can become a committed change across the knowledge chain.
What we know has changedWe have not established that this possible change has happened.
What this does not showVisible lineage does not guarantee that people act on it.
This is the first time Lens has mapped this path. We have no later evidence showing whether it is happening more, less or about the same.
Several fluent answers present the same conclusion without visible shared provenance.
Agreement may be interpreted as independent confirmation.
Confidence and reuse rise while the underlying source receives less scrutiny.
Teams and automated systems: Repeat or operationalise the answer in further work.
One stale source can become embedded across multiple decisions.
What we know has changedWe have not established that this possible change has happened.
What this does not showThe demonstrated risk is not a measured prevalence claim.
This is the first time Lens has mapped this path. We have no later evidence showing whether it is happening more, less or about the same.
If teams preserve source versions, decision context, ownership and observed results
Then people could challenge stale consensus and update the underlying decision rather than merely retrieving more copies of it.
Would weaken this: The record exists but is not used. Ownership and update responsibility remain ambiguous.
Scope: Organisational knowledge continuity. Horizon: Across policy and system changes.
If several tools or teams reuse the same outdated record without visible provenance
Then repetition could make an unsupported answer appear independently corroborated and spread the error more widely.
Would weaken this: Independent current sources support the same conclusion. Version checks reject the outdated dependency.
Scope: Shared-source errors, not all AI-assisted work. Horizon: During repeated retrieval and reuse.
Research across Westminster governments found that organisations sometimes lose important knowledge, but not always. Staff turnover, workplace choices and the stories institutions tell about their past can affect what survives.
Open evidence ↗A Queensland parliamentary inquiry found that original Fitzgerald Inquiry documents were moved into CJC intelligence files and that evidence suggested more than 4,000 original documents were later destroyed, although the exact count was uncertain.
Open evidence ↗A 19-month study in an academic health system found that generative AI changed how staff found, created, shared, kept and forgot knowledge. Some changes helped. Others multiplied weak or risky information.
Open evidence ↗In a Lens test, five AI answers agreed because all five used the same outdated policy. No other up-to-date source supported them.
Open evidence ↗What could change this assessment?