When AI-generated evidence becomes tomorrow's data

#Question

What happens when AI-generated scientific or clinical content becomes part of the data used by future public-health systems?

#Answer in brief

The boundary between observation and previous model output can begin to disappear.

Generative AI can now help produce clinical notes, classifications, summaries, extraction labels, research prose, code, synthetic records and analytical interpretations. Some of those outputs are transient. Others are stored.

Once stored, they may later be:

• analysed as routine data;
• retrieved by another model;
• included in a research corpus;
• used for model development;
• summarized into policy evidence;
• treated as if they were direct observations.

At that point, provenance is no longer only a reproducibility concern.

It becomes part of the epidemiological meaning of the data.

Research on recursive training has shown that indiscriminate reuse of model-generated data can progressively distort the learned distribution and lose information from the original data, a phenomenon described as model collapse.^[1] A 2026 medical-informatics viewpoint has extended that concern to recursive clinical AI.^[2]

The public-health problem is broader than model training:

What if AI-mediated representations of reality become tomorrow's evidence about reality?

#Public-health data are already transformed observations

A health-data record is rarely raw reality.

A death becomes a certificate.

A consultation becomes a coded encounter.

A laboratory process becomes a result.

A surveillance event becomes a classified case.

A researcher turns records into an analytical dataset.

Each transformation creates distance between the underlying event and the final variable.

Good information systems preserve enough provenance to reconstruct those transformations.

Generative AI adds another possible layer.

A future record may not simply mean:

person → observation → record

It may mean:

person
→ observation
→ human note
→ AI summary
→ structured field
→ second AI interpretation
→ analytical dataset

The final field can look ordinary while containing several generations of mediation.

#AI-derived is not the same as synthetic

The distinction matters.

An AI system can create several different kinds of information.

#AI-transformed observation

The source event is real, but AI extracts or summarizes it.

Example: an LLM converts a clinical narrative into a structured exposure variable.

#AI-generated interpretation

The system creates a conclusion from observed data.

Example: a generated explanation of why a surveillance signal changed.

#Synthetic data

The model generates records that do not correspond one-to-one with observed individuals.

#AI-mediated human record

A person uses AI to draft or rewrite content that is then entered into an operational system.

These are not equivalent.

A future analyst needs to know which kind of object is being analysed.

#The recursive pathway

Consider a deliberately simplified chain.

Year 1:

A clinician writes a note.

An AI system summarizes it.

The summary enters the electronic record.

Year 2:

A research team trains a classifier on those records without distinguishing AI-assisted from directly authored text.

Year 3:

The classifier helps create new structured labels.

Those labels become part of a surveillance database.

Year 4:

A foundation model retrieves the surveillance database to generate an evidence summary.

Year 5:

The summary becomes part of a policy corpus used to train or evaluate another model.

At each stage, the information may remain useful.

But the relationship between the final object and the original observation has become increasingly indirect.

Without provenance, the system can no longer tell whether multiple apparently independent pieces of evidence descend from the same earlier AI transformation.

#Repetition can masquerade as confirmation

Suppose one AI-generated interpretation enters three downstream documents.

A later model retrieves all three.

The same original claim now appears three times.

The model may treat repetition as corroboration even though the sources are genealogically dependent.

Humans can make the same mistake.

Citation counts, document counts and apparent consensus can become inflated by recursive copying.

The relevant question is no longer only:

How many sources support this claim?

It is:

How many independent evidence lineages support this claim?

That is a provenance problem.

#Model collapse is an analogy, not the whole problem

The Nature study on recursive model training demonstrated that generative models trained indiscriminately on model-generated data can lose information about the original distribution, with low-probability regions affected early.^[1]

Public-health evidence systems are not identical to recursive language-model training.

The analogy is nevertheless useful.

Rare events, unusual presentations, minority populations and uncommon causal pathways often live in the tails of distributions.

If AI-mediated documentation and synthesis systematically compress those tails, later data systems may preserve what is common more faithfully than what is unusual.

That risk requires empirical study rather than assumption.

The important point is that recursive mediation can change what remains visible.

#Rare public-health signals are especially vulnerable

Public health often cares precisely about weak, early or unusual signals:

• a rare adverse event;
• an atypical symptom pattern;
• a small geographic cluster;
• an emerging pathogen;
• an uncommon exposure;
• a new health inequality.

Generative systems are optimized to produce plausible representations based on learned regularities.

If unusual observations are repeatedly summarized into more typical language, the record may become easier to read and less informative for anomaly detection.

A polished representation is not always a lossless representation.

For surveillance, preserving atypicality can be more important than improving prose.

#AI disclosure is necessary but not sufficient

CDC guidance on generative AI in scientific work recommends disclosure of substantive AI use, including tool, purpose and human oversight, and calls for enough methodological detail to support reproducibility when AI is used analytically.^[3]

That is important for publications.

Operational data systems need an analogous but more granular approach.

A paper can contain one AI-use disclosure.

A health dataset may contain millions of records with different origin states.

The provenance therefore needs to exist at the appropriate level:

• document;
• field;
• transformation;
• analytical artifact;
• model output.

The right granularity depends on consequence.

#A proposed evidence-origin state

This article proposes that AI-mediated public-health information carry a machine-readable evidence-origin state where feasible.

A simple vocabulary could distinguish:

OBSERVED
HUMAN_DERIVED
AI_DERIVED
AI_GENERATED
SYNTHETIC
MIXED
UNKNOWN

That alone is insufficient.

It should be accompanied, where appropriate, by:

• source identity;
• transformation type;
• AI model/tool identity or family where available;
• date;
• human verification state;
• downstream modifications;
• confidence or uncertainty;
• lineage to the prior artifact.

This is a proposed provenance pattern, not an established public-health standard.

#Unknown should be an allowed state

A dangerous governance instinct is to force every record into a clean category.

Legacy systems will often lack provenance.

Vendor systems may not expose model details.

Human and AI editing may be inseparable.

The correct state may be:

origin unknown.

That uncertainty should remain visible.

Invented provenance is worse than missing provenance.

#Recursive evidence can affect causal inference

Suppose an AI-generated variable derived from an outcome also influences a later exposure classification.

The analyst may treat the variables as independent measurements when they partly share a generative pathway.

Or an AI summary may incorporate information recorded after the nominal index date, creating temporal leakage.

Recursive evidence is therefore not only a research-integrity issue.

It can create:

• measurement dependence;
• temporal leakage;
• circularity;
• duplicated information;
• hidden conditioning;
• spurious apparent replication.

The data lineage belongs in the causal model.

#Training data are only one downstream use

Much discussion focuses on whether AI-generated data contaminate future training corpora.

That is only one pathway.

Public-health systems can recursively reuse AI-generated information through:

• surveillance;
• evidence synthesis;
• guideline development;
• operational dashboards;
• automated coding;
• quality indicators;
• policy memoranda;
• public communication;
• evaluation reports.

A system can therefore become recursively self-referential even if no foundation model is ever retrained.

The knowledge system itself can recycle its own interpretations.

#Independent anchors become valuable

As AI mediation increases, data closest to direct observation become more valuable as reference points.

Examples include:

• original laboratory measurements;
• source images;
• raw sensor data;
• validated registry events;
• human-adjudicated samples;
• archived pre-AI historical data;
• independently collected sentinel data.

The 2026 clinical-autophagy viewpoint proposes protected human-verified reference data as a safeguard against recursive degradation.^[2]

The exact architecture remains debatable.

The underlying principle is strong:

a recursively mediated system needs anchors whose relationship to observation is known.

#Provenance should travel with the artifact

A common failure is to keep AI-use information in a separate policy document.

The downstream analyst then receives the data without the provenance.

A stronger architecture binds provenance to the artifact itself.

If a structured variable was AI-derived, that state should survive:

• export;
• linkage;
• transformation;
• aggregation;
• publication.

WHO Europe has emphasized that trustworthy AI depends on strong health-data governance, quality, representativeness and responsible data use.^[4]

Recursive AI makes lineage another dimension of that governance.

#A public-health recursion audit

Before reusing a dataset or evidence corpus, ask:

1. Which elements are direct observations?
2. Which were generated or transformed by AI?
3. Can those elements be traced to source observations?
4. Have AI-derived objects been reused as inputs to later AI processes?
5. Are apparently independent documents descendants of the same generated source?
6. Could rare information have been compressed during summarization?
7. Is there a human-verified or observation-proximal reference set?
8. Can downstream users filter or stratify by evidence origin?
9. What proportion of origin is unknown?

The purpose is not to purify data of all AI influence.

It is to know what kind of evidence the system contains.

#Synthetic data need explicit boundaries

Synthetic data can be valuable for privacy-preserving development, testing and simulation.

The problem is not that synthetic data exist.

The problem is loss of boundary.

A synthetic record should never become indistinguishable from an observed record merely because both share the same schema.

Likewise, an AI-generated summary should not silently acquire the epistemic status of the source material it summarized.

Schema compatibility is not provenance equivalence.

#The deeper shift

AI makes evidence production faster.

It also allows evidence to become input to new evidence-production processes almost immediately.

That creates a new failure mode:

the system can begin learning from, validating against and citing its own prior representations of reality.

Public-health information systems therefore need more than data lineage.

They need epistemic lineage: a record of how many layers of interpretation separate an analytical object from the observation it claims to represent.

The core question becomes:

Are we still measuring the population, or are we increasingly measuring previous models' descriptions of the population?

#Limitations and uncertainty

The strongest empirical evidence for model collapse concerns recursive training of generative models, not public-health surveillance or epidemiological databases.^[1] The clinical-autophagy paper is a viewpoint that proposes a theoretical risk and governance response rather than demonstrating population-level harm.^[2] This article therefore advances recursive evidence contamination as a public-health hypothesis and provenance problem, not as an established epidemiological effect. AI-mediated transformation can improve data quality, accessibility and consistency, and provenance requirements should remain proportionate to risk. The proposed evidence-origin vocabulary requires testing, standardization and interoperability work before operational use.