The AI-Native EHR Has a Data Pollution Problem

Healthcare is moving from electronic records that humans document to clinical records that machines help generate. That shift can reduce administrative burden and improve access. It also introduces a new class of risk: AI output can enter the permanent record, influence care and payment, become a label in analytics, and later return as evidence for another AI system.

The core argument

The industry is treating each AI use case as a bounded product decision. The actual risk is cumulative. Healthcare needs a formal data pollution control layer before AI-generated documentation becomes indistinguishable from direct observation.

An ambient scribe drafts a note. A coding model extracts diagnoses. A risk engine interprets those codes. A quality program measures the resulting population. A research team builds a cohort. Another model trains on the same data. At every step, machine inference can harden into apparent clinical fact.

Human review remains essential, but a signature alone cannot supply the missing lineage. Health systems need machine-readable provenance, separate treatment of observed and inferred facts, correction propagation, downstream-use controls, and measurable limits on untraceable AI content.

The electronic health record was already an imperfect account of care. It contains direct observations, patient reports, clinician judgments, billing-oriented classifications, imported results, templated language, and text copied forward from earlier encounters. Generative AI adds another author, but the record rarely identifies that author in a form downstream systems can reliably use.

That is the danger.

An AI-generated sentence can look cleaner than the evidence that produced it. It can convert uncertain conversation into declarative prose, turn a differential diagnosis into an active condition, omit a negative finding, or add a clinically plausible detail that was never stated. Once a clinician signs the note, most downstream systems see a single status: authenticated. They do not see which words came from the patient, which came from a device, which were inferred by a model, which were edited by a clinician, or which were carried forward from an older note.

Authentication answers who accepted responsibility for the record. It does not reconstruct how every assertion was created.

Healthcare is therefore approaching a structural break. The chart is no longer only a source for AI. It is becoming an output of AI, a source for other AI, and a potential training set for future AI. The information loop is closing while provenance remains open.

The industry has a name for polluted environmental data, compromised laboratory samples, and corrupted software dependencies. It still lacks a widely operationalized control model for clinical data whose apparent origin is human but whose actual origin is partly generative.

Ambient documentation deserves its momentum. A pragmatic randomized trial involving 238 outpatient physicians across 14 specialties found modest documentation-time improvement for one of two evaluated scribes and potential improvements in burnout, cognitive load, and work exhaustion. The same study noted that EHR time measurements did not include editing performed inside vendor platforms, and that frequent user feedback is an unrealistic sole quality-assurance strategy for nondeterministic, versioned systems.

The benefit and the risk can be true at the same time.

Another evaluation found that AI-authored notes could score close to physician-authored reference notes on overall document quality while remaining less succinct and more prone to hallucination.

These studies do not prove that AI scribes are unsafe. They prove that polished note quality and factual lineage are different properties. A note can be organized, readable, and useful while still containing an omission or unsupported assertion that matters downstream.

Healthcare procurement often rewards the visible benefit: less time typing, shorter after-hours work, better clinician experience, faster encounter closure. The downstream cost is dispersed across departments and time. A documentation error may surface later as a problem-list discrepancy, coding change, denial, risk-score shift, quality-measure variance, research misclassification, or biased training example. Because no single team owns the entire chain, the organization can celebrate local productivity while accumulating enterprise data debt.

"The clinician reviews every note" has become the standard answer to AI documentation risk. It is necessary. It is also inadequate as the only control.

Clinicians review under time pressure. They may focus on whether the note supports the immediate plan, not whether each sentence is suitable for every future analytic use. They may correct a medication error while leaving a subtle statement of certainty unchanged. They cannot predict every payer rule, quality measure, natural-language-processing pipeline, research phenotype, retrieval system, or future training job that will consume the text.

The signature then performs a dangerous transformation. It changes a machine-assisted draft into an authenticated medical record without preserving enough detail about the transformation itself.

This is not an argument against clinician accountability. It is an argument against forcing a clinician signature to carry technical meaning it was never designed to carry. Legal attestation, clinical approval, data provenance, and fitness for secondary use are separate controls. Collapsing them into one checkbox creates false confidence.

The industry learned a related lesson from copy and paste. Reviews of EHR documentation have warned that copied text can spread outdated or inaccurate information and contribute to adverse events. AI generation increases the speed, volume, and apparent originality of the content. Traditional duplication detection may identify repeated text. It will not reliably identify a newly worded assertion derived from an older error.

1. The scribe-to-chart loop

The system converts speech into a transcript and the transcript into a clinical note. Noise, accents, overlapping speakers, ambiguous negation, and missing context can alter the source before summarization begins. The model may then compress uncertainty or omit details to produce a coherent note. If the organization retains only the final note, later reviewers cannot determine whether an error arose in capture, transcription, summarization, template mapping, or human editing.

2. The chart-to-code loop

Clinical documentation feeds diagnosis and procedure coding, severity capture, utilization review, prior authorization, and payment. An inferred condition can become a coded condition. The code then appears more structured and authoritative than the prose from which it came. When later systems treat the code as independent confirmation, one generated assertion has effectively voted twice.

3. The chart-to-measurement loop

Risk adjustment, quality measures, population health, utilization management, and operational dashboards depend on fields extracted from the record. If AI increases documentation detail unevenly across specialties, locations, languages, or clinicians, apparent changes in acuity may reflect documentation behavior rather than a change in patient health. An organization can improve its data completeness and simultaneously break comparability with its own historical baseline.

That is not a reason to preserve poor documentation. It is a reason to measure the intervention. Without lineage and time-aware baselines, leaders may mistake better capture for worsening disease, model lift for documentation inflation, or performance improvement for a denominator change.

4. The chart-to-model loop

EHR data supports model development, validation, retrieval, and fine-tuning. If AI-generated notes enter those datasets without labeling, developers cannot reliably separate human-observed evidence from prior model output. The evaluation set may contain text produced by the same model family being evaluated. Performance can appear to improve because the model recognizes its own style.

Research on recursive training has shown that indiscriminate use of model-generated data can cause generative models to lose information about the original distribution, including rare patterns. That finding does not establish that clinical EHR models are currently collapsing — conditions differ, and mixtures of real and synthetic data can mitigate the effect. The warning is still highly relevant: without reliable origin labels, healthcare organizations will not know the proportion, lineage, or recurrence depth of machine-generated content in their own datasets.

Rare presentations and underrepresented populations are exactly where a quiet loss of distributional detail would be most dangerous.

5. The model-to-chart-to-model loop

The most consequential loop begins when a predictive system influences the label later used to update another model. A risk model prompts an intervention. The intervention changes clinical behavior. The changed behavior alters the recorded outcome. The organization retrains on the resulting data as if it were an independent observation.

Healthcare machine-learning research has demonstrated this feedback-loop problem in simulation using real ICU data. Under a scenario in which clinicians fully adopted predicted labels, false-positive rates grew uncontrollably across model updates. Production environments are more complex, but the lesson is difficult to dismiss: once a prediction changes the data-generating process, conventional retraining can confuse the model's influence with clinical truth. Generative documentation makes that distinction even harder, because the model's influence can be embedded directly in the text used as evidence.

The industry is preparing for the spectacular failure: the fabricated drug, the impossible diagnosis, or the sentence that clearly contradicts the encounter. Those errors matter, but they are also the easiest to notice.

Data pollution is more likely to arrive as thousands of plausible microdistortions:

• ▸uncertainty rewritten as certainty;
• ▸a ruled-out diagnosis retained without its negation;
• ▸a family-history condition attached to the patient;
• ▸a copied condition that is no longer active;
• ▸a normal examination inferred from a familiar template;
• ▸a symptom omitted because the model judged it less relevant;
• ▸a severity term standardized differently across populations;
• ▸a plan paraphrased in language that triggers a different coding or authorization rule.

Each error may appear minor. The cumulative data product is not.

Healthcare analytics is especially vulnerable because scale converts small classification shifts into material population effects. A one-point change in the prevalence of a condition can alter outreach volumes, staffing forecasts, contract performance, quality denominators, and expected cost. If the shift comes from AI-mediated documentation rather than patient health, the organization may take confident action on a synthetic signal.

The danger is epistemic before it is statistical. The organization gradually loses the ability to answer a basic question: what did we actually observe?

Healthcare data platforms usually organize information by source system, encounter, date, patient, and resource type. AI-mediated data requires an additional dimension: method of creation.

These classes should not be hidden in a vendor audit log. They must travel with the data so that clinical applications, warehouses, quality engines, research platforms, and model pipelines can enforce different rules.

HL7 FHIR already provides part of the technical foundation. The Provenance resource records the activity, agents, and entities involved in creating or revising a resource and supports assessments of authenticity, reliability, and trustworthiness. AuditEvent addresses related event-level activity. The standards do not implement policy by themselves, however. A Provenance resource that is never populated, exchanged, validated, or consumed is ceremonial compliance.

Machine-readable lineage. Every AI-assisted write should identify the vendor, product, model, model version, prompt or workflow version, source artifacts, creation time, initiating identity, reviewing identity, and disposition. The lineage should survive export and exchange. "Created in the EHR" is not sufficient when the EHR invoked several models before saving the result.

Assertion-level traceability. Document-level labeling is too coarse. A single note may combine patient speech, imported results, clinician reasoning, copied text, and generated language. High-impact assertions — diagnoses, allergies, medications, procedures, negated findings — need traceability to their supporting source. Systems should be able to show whether an assertion was spoken, observed, imported, inferred, copied, or generated.

Separate evidence from inference. AI suggestions should not enter the same field and status as confirmed clinical facts merely because the interface makes acceptance convenient. Risk scores, candidate diagnoses, code suggestions, and generated summaries require explicit state transitions. A model output can be useful without masquerading as observation.

Correction propagation. Correcting the note is only the first step. The organization must identify every derived artifact that consumed the error: problem lists, codes, claims, quality calculations, registries, feature stores, embeddings, vector indexes, model-training snapshots, and outbound exchanges. Material corrections need a propagation and retraction process with measurable completion time.

Downstream-use policies. Every secondary-use pipeline should be able to include, exclude, or weight AI-derived content by provenance class, model version, validation state, and use case. Data suitable for immediate care may not be suitable as a research label. Data suitable for retrieval may not be suitable for training.

Protected reference datasets. Organizations should maintain versioned evaluation sets composed of well-characterized, human-origin or independently adjudicated evidence, isolated from production AI output and from the systems they evaluate. Otherwise, benchmark contamination can convert familiarity into apparent accuracy.

Continuous surveillance. FDA has emphasized that AI performance can change with clinical practice, patient demographics, data inputs, infrastructure, workflow, and guidelines, and has sought practical approaches to real-world monitoring and drift detection. Monitoring should include content origin and feedback exposure, not only conventional accuracy metrics. A model can maintain aggregate performance while the data beneath it becomes less independent.

Contractual control. Vendor agreements should specify whether prompts, transcripts, drafts, clinician edits, and accepted notes may be retained or reused for training. They should require model-version notice, exportable provenance, correction support, incident disclosure, subgroup testing, and transition assistance. A vendor that cannot identify which model created a clinical assertion cannot provide an adequate audit trail for that assertion.

Healthcare executives manage error rates, infection rates, denial rates, data-quality thresholds, and cybersecurity exposure. AI-originated clinical content needs the same operational discipline.

A practical pollution scorecard should include:

1. ▸Provenance coverage — percentage of AI-assisted records with complete, machine-readable origin metadata.
2. ▸Unsupported assertion rate — clinically meaningful statements that cannot be traced to an encounter source, imported result, or documented human judgment.
3. ▸Omission rate — required or clinically material facts present in the source but absent from the generated artifact.
4. ▸Human edit distance — the volume and clinical severity of changes between model draft and signed record.
5. ▸Copy and generation depth — how many prior notes or model outputs sit behind a current assertion.
6. ▸Model-on-model exposure — proportion of training, validation, retrieval, and analytic inputs containing AI-derived content.
7. ▸Correction propagation latency — time required to retract or update downstream derivatives after a material source correction.
8. ▸Subgroup variance — differences in omission, unsupported assertion, and correction rates by language, accent, specialty, setting, age, disability, and other relevant characteristics.
9. ▸Independent evidence rate — proportion of high-impact structured facts supported by a source not generated by the same model workflow.
10. ▸Downstream retraction completeness — percentage of known dependent artifacts successfully corrected after the source changes.

The target cannot always be zero. AI transformation will be embedded throughout clinical work, and synthetic data can be valuable when governed deliberately. The objective is to make exposure visible, bounded, and reversible.

An organization that cannot calculate these measures does not know whether it is accumulating AI data pollution. It only knows that no one has reported it yet.

The ONC decision-support certification framework requires support for source attributes — intended use, training-data relevance, external validation, performance, fairness, maintenance, and risk management — for predictive decision support interventions supplied through certified health IT. Those requirements advance transparency for the model.

The data pollution problem also requires transparency for the output.

Knowing how a model was trained does not tell a downstream analyst whether a specific diagnosis originated in a conversation, a copied note, an inference engine, or a generative summary. Knowing a model's aggregate performance does not reveal whether its accepted outputs later entered its retraining data. Knowing that a clinician could override a recommendation does not show whether the recommendation changed the note, the label, or the care path before that override.

Model governance without output lineage controls the factory and ignores the supply chain. Healthcare needs both. Regulators, accreditors, EHR vendors, AI vendors, and health systems should treat AI-origin metadata as part of clinical-data integrity. The alternative is a growing body of authenticated records whose authorship is technically ambiguous and whose independence cannot be established.

Healthcare organizations should stop allowing AI-generated clinical content to enter the longitudinal record without exportable, machine-readable provenance.

• ▸Reject the claim that a human signature erases the need for origin metadata.
• ▸Prohibit unlabeled AI-derived content from entering reference datasets, outcome labels, and model-validation cohorts.
• ▸Require correction to propagate beyond the note into codes, measures, indexes, and training assets.
• ▸Demand the ability to disable a model version, identify every record it influenced, and reconstruct the evidence available when it acted.

These requirements will slow some deployments. That is a feature. The speed of documentation should not exceed the organization's ability to preserve truth.

The AI-native EHR can still be the right destination. Intelligence embedded in workflow can reduce burden, reveal risk, and help clinicians act sooner. But an intelligent record that cannot distinguish observation from generation is not AI-native in any defensible sense. It is provenance-blind.

And provenance-blind healthcare AI will eventually begin validating itself.

Brandywine Consulting Partners helps healthcare organizations build the control architecture between AI output and enterprise data use. BCP combines healthcare interoperability, cloud-native engineering, data governance, analytics, and AI implementation so that provenance survives the full path from clinical workflow to downstream decision.

Our AI data pollution readiness work can include:

• ▸inventorying every workflow in which AI creates, transforms, extracts, ranks, or writes clinical information;
• ▸mapping how generated content moves into EHRs, FHIR resources, HL7 messages, X12 transactions, warehouses, lakehouses, semantic models, vector stores, and external partners;
• ▸implementing FHIR Provenance and AuditEvent patterns with model, source, user, and version metadata;
• ▸separating observed, reported, interpreted, inferred, and synthetic information in canonical data models;
• ▸establishing AI-content quality rules, exception queues, lineage dashboards, and correction-propagation workflows;
• ▸evaluating subgroup performance and documentation shifts that can affect risk, quality, payment, and care management;
• ▸creating protected validation datasets and controls for model training, retrieval, and fine-tuning;
• ▸designing Azure-based monitoring, audit evidence, secure integration, and managed operational support aligned with HIPAA and HITRUST practices.

BCP's role is not to obstruct useful AI. It is to make AI adoption auditable, interoperable, and reversible before machine-generated evidence becomes inseparable from the clinical record.

Data Innovation. Advanced Analytics. Better Outcomes.

Sources

1. ▸Lukac PJ, et al. Ambient AI Scribes in Clinical Practice: A Randomized Trial. NEJM AI. 2025.
2. ▸Arko L IV, et al. Documenting Care with AI: A Comparative Analysis of Commercial Scribe Tools. Studies in Health Technology and Informatics. 2025;329:337-341.
3. ▸Palm E, et al. Assessing the quality of AI-generated clinical notes. Frontiers in Artificial Intelligence. 2025;8:1691499.
4. ▸Tsou AY, et al. Safe Practices for Copy and Paste in the EHR. Applied Clinical Informatics. 2017;8(1):12-34.
5. ▸Shumailov I, et al. AI models collapse when trained on recursively generated data. Nature. 2024;631:755-759.
6. ▸Adam GA, et al. Hidden Risks of Machine Learning Applied to Healthcare. PMLR. 2020;126:107-121.
7. ▸HL7 International. FHIR R4 Provenance Resource.
8. ▸U.S. Food and Drug Administration. Measuring and Evaluating AI-enabled Medical Device Performance in the Real World. 2025.
9. ▸ASTP/ONC. Decision Support Interventions Certification Criterion.

#HealthcareAI #AIGovernance #ClinicalDataIntegrity #FHIR #Provenance #AmbientScribe #EHR #HealthcareIT #DataQuality #BCP