In May 2026, The Lancet published one of the most significant indictments of AI use in scientific research to date. A team from Columbia University analysed nearly 2.5 million PubMed-indexed papers and discovered that approximately one in every 277 papers published in the first seven weeks of 2026 contained at least one fabricated citation, a sixfold increase from 2023 when the rate was one in 2,828 papers. [1] Review articles, the papers that form the evidence base for clinical guidelines and HTA submissions, showed a 57% higher fabrication rate than other paper types. The hallucinated references were not easily identifiable: they appeared with realistic paper titles, real author names, correct journal formatting, and plausible DOIs. [3] They were, in almost every way, indistinguishable from legitimate citations.
This is the source quality crisis that generic AI tools have introduced into scientific and clinical research in 2026. For pharma teams using large language models to assist with evidence synthesis, competitive intelligence, pharma market monitoring, pharma clinical strategy development, regulatory dossier preparation, and medical communications, the implication is direct: when the AI tool is not source-controlled, every output carries the risk of containing a citation that looks real but is not, a data point that sounds correct but was generated rather than retrieved, or a clinical claim that is plausible but unverifiable. At Pienomial, we built source controlled AI research capability into KnolAI specifically because we recognised that speed without source control is not a research capability. It is a liability.
1. The Source Quality Crisis: What the Data Shows
The scale of the fabricated citation problem in 2025 and early 2026 is not a fringe concern about occasional AI errors. It is a systemic disruption to the integrity of the scientific record that researchers from Nature, The Lancet, and multiple research institutions have now documented independently. [1]
A cross-institutional study estimated that nearly 146,900 hallucinated citations were present in research papers hosted across four major scientific repositories, including arXiv, bioRxiv, SSRN, and PubMed Central, in 2025 alone. [2] The fabrication rate accelerated sharply from mid-2024, coinciding with the widespread adoption of generative AI writing tools. By the first seven weeks of 2026, the rate of fabricated references in published papers had reached 56.9 per 10,000 papers, up from 4 per 10,000 papers in 2023. [3] The researchers found that more than 98% of papers containing fabricated references had not received any correction, expression of concern, or retraction at the time of the audit.
The clinical consequence is direct. If a systematic literature review used as the evidence base for an HTA submission includes papers with fabricated citations, the evidence chain underpinning the submission is compromised. The HTA body reviewer cannot verify the claimed source. The NICE technical team issuing a clarification request about a clinical claim discovers a dead-end citation. The G-BA scientific advisor finds a reference that does not exist. The submission fails on the evidence integrity standard before the clinical data is even evaluated. [4]
2. What Source-Controlled AI Research Means
Source controlled AI research is a specific design principle: every claim that an AI research system generates or retrieves must be linked to a verified, real, accessible primary source, and that link must be maintained at the claim level through the entire research workflow from source ingestion to output delivery. [9]
Source control has three operational components. First, source validation at ingestion: every document entering the research AI's knowledge base is validated against its claimed source before ingestion. A clinical publication is confirmed to exist in PubMed with the attributed author and DOI before it can be used as the basis for any output claim. A regulatory filing is confirmed to exist in the FDA or EMA database. An HTA decision is confirmed against the NICE or G-BA assessment database. Documents that cannot be independently verified do not enter the knowledge base.
Second, claim-level source linking: every factual claim that the AI system generates is linked to the specific entity-relationship triple in the knowledge base from which it was derived, and that triple links to the specific source document and its location. Not document-level attribution, meaning the source document is mentioned in a reference list. Claim-level attribution, meaning the specific claim links to the specific sentence, table, or data point in the source document that it represents.
Third, source currency controls: the knowledge base is continuously updated with new validated sources, and sources are flagged when they are superseded by newer evidence. For competitive intelligence and HTA landscape intelligence, a source that was current at the time of ingestion may be superseded six months later by a new regulatory decision or trial readout. Source control includes the mechanism to identify and update these superseded facts rather than silently serving outdated intelligence as current.
The source quality crisis in scientific literature is a direct consequence of how large language models generate text. LLMs do not retrieve facts from verified sources. They generate text that statistically resembles text about the topic based on patterns in training data. When an LLM generates a citation, it produces the most statistically probable citation-shaped text for the topic, which may or may not correspond to a real paper with those attributes.
This is not a training quality problem that can be solved by using a better model. It is a structural property of probabilistic text generation. A model trained on more clinical literature will generate more clinically accurate-looking citations, but the generation process remains probabilistic. The model is not checking whether the cited paper exists. It is producing text that looks like a valid citation because that is what statistically similar text in its training data looks like. The result is a citation that appears legitimate because it was generated by a model trained on legitimate citations, even though it refers to nothing real.
Gartner predicted in January 2026 that by 2028, 50% of organisations will implement zero-trust data governance due to the proliferation of unverified AI-generated data, noting that as AI-generated content becomes more prevalent, the risk of model collapse grows as AI tools train on outputs from previous models. For pharma teams, the model collapse risk is not hypothetical. Systematic literature reviews that use AI tools drawing on a scientific literature increasingly contaminated with hallucinated citations will propagate those hallucinated citations into HTA submissions, regulatory dossiers, and clinical guidelines, compounding the integrity problem at each stage.
4. The Regulatory Consequence: FDA Does Not Exempt AI Content
The FDA has been explicit that AI-generated content in pharmaceutical submissions does not receive a separate compliance lane. The same rules for fair balance, claim support, and misleading impression prevention that apply to human-authored content apply to AI-generated content. Speed of production does not reduce the compliance requirement.The FDA's January 2025 draft guidance on AI in drug regulatory submissions establishes a risk-based credibility framework requiring that AI models used in regulatory applications demonstrate data provenance and traceable documentation.A submission containing a clinical claim sourced to a fabricated citation fails this framework regardless of whether the fabrication was generated by a human or an AI.
NICE's 2024 position statement requires submitting organisations to take responsibility for all content submitted, regardless of whether AI was used in its generation. If an AI-assisted systematic literature review includes a paper with a hallucinated citation, and that citation appears in the NICE evidence table, the submitting organisation is accountable for the error. The NICE reviewer who cannot find the cited paper has grounds to issue a formal clarification request that delays the appraisal and may trigger a broader evidence quality review.
For pharma teams, the compliance implication is direct: using a generic AI tool for research assistance without source control is not a neutral productivity decision. It is a deliberate acceptance of source integrity risk in a context where source integrity is a regulatory obligation.
5. How KnolAI Implements Source Control by Architecture
KnolAI, the research intelligence module of our Knolens platform, is built on source control as a foundational architectural constraint. There are three mechanisms through which KnolAI enforces source control at every stage of the research workflow.
Mechanism 1, Validated ingestion pipeline: Every document entering the Knolens knowledge graph passes through a validated ingestion pipeline. Clinical publications are verified against PubMed with author, title, journal, and DOI cross-referencing before ingestion. Regulatory filings are confirmed against FDA and EMA databases. HTA decisions are confirmed against the NICE, G-BA, HAS, and ICER assessment repositories. Conference abstracts are verified against the abstract book of record for the named conference. A document that cannot be verified does not enter the knowledge graph. This is not a quality heuristic. It is a hard architectural gate.
Mechanism 2, Claim-level triple storage: When a validated document is ingested, the specific facts it contains are extracted as entity-relationship-source triples. Each triple stores the fact, the relationship between entities, and the precise source location: the specific section, paragraph, table, or figure in the source document from which the fact was extracted. The triple is the atomic unit of knowledge in the Knolens graph. It is not a document summary. It is a verified fact with a verifiable source address.
Mechanism 3, Constrained output generation: When KnolAI generates research output, it retrieves specific triples from the knowledge graph and generates natural language from those retrieved triples. The output generation process is explicitly constrained: the language model cannot introduce claims beyond what the retrieved triples contain. Every claim in the output is linked to its source triple, which links to its verified source document and location. The chain from output claim to primary source is complete, claim-level, and verifiable by any reviewer with access to the source document.
6. The Practical Difference: Source-Controlled vs Generic AI in HEOR
Consider a HEOR analyst preparing an evidence table for a NICE single technology appraisal. They need to document the OS and PFS results from the pivotal trial and from three comparator trials identified in the systematic literature review.
With a generic AI tool: The analyst inputs the four trial names and asks for a summary of OS and PFS results. The AI generates a table with efficacy figures for all four trials. The figures look correct, with realistic hazard ratios, confidence intervals, and p-values. The analyst includes the table in the submission. Two of the four OS figures are from the correct trials. One figure is from a different trial in the same indication that the model confused with the named trial. One confidence interval was generated by the model as a statistically plausible value rather than retrieved from a real source. The NICE technical reviewer identifies the discrepancies during the appraisal and issues a formal clarification request.
With KnolAI: The analyst queries the Knolens knowledge graph for the OS and PFS results from the four named trials. KnolAI retrieves the specific entity-relationship triples for each efficacy result from each trial, sourced to the specific publication and table or figure from which they were extracted. The evidence table is generated from those retrieved triples. Every figure links to its source. The NICE reviewer who checks any figure can verify it against the source document in seconds. No clarification request is generated for evidence integrity issues.
7. Source Control for Competitive Intelligence: The Strategic Risk Dimension
The source quality risk in pharma AI is not limited to HTA submissions and regulatory dossiers. It applies equally to competitive intelligence, where an incorrect claim about a competitor's clinical trial result, pipeline programme, or regulatory status can influence portfolio investment decisions, clinical protocol design, or market access strategy.
A competitive intelligence brief produced by a generic AI tool that cited a competitor's Phase III result against a hallucinated comparator arm would be indistinguishable from a correct CI brief until a team member independently verified the source, which research consistently shows most teams do not do consistently under time pressure. The consequence of acting on incorrect CI is a strategic decision made on false evidence. The portfolio investment that should have been made in response to a real competitor readout is not made. Or the clinical protocol adjustment that was unnecessary is made at substantial cost.
KnolAI's source control architecture applies to competitive intelligence outputs with the same rigour it applies to evidence synthesis. Every competitive claim in a KnolAI CI brief links to a specific, verified primary source: the ClinicalTrials.gov registration, the regulatory filing, the conference abstract, or the peer-reviewed publication from which the data was extracted. Senior leadership receiving a KnolAI CI brief can trust the figures in it, not because we ask them to, but because the architecture guarantees that every figure has a verified source.
8. Source Control and the Gartner Zero-Trust Data Governance Prediction
Gartner's January 2026 prediction that 50% of organisations will implement zero-trust data governance by 2028, driven by the proliferation of unverified AI-generated data, is a direct reflection of the source quality crisis that the Lancet fabricated citation study documents empirically.Zero-trust data governance in the AI context means treating all AI-generated content as unverified until its source provenance is confirmed, regardless of how credible the content appears.
For pharma organisations, zero-trust data governance is not a 2028 aspiration. It is a current operational requirement. Every AI-generated clinical claim in a regulatory submission, HTA dossier, or payer presentation must be verified against its primary source before it is used. The question is whether that verification happens as a manual post-generation step, which is what most pharma teams are doing with generic AI tools today at substantial analyst time cost, or whether source control is built into the AI architecture so that verification is structural rather than procedural.
KnolAI's architecture implements zero-trust data governance by design. No claim enters a KnolAI output without a verified source. No source enters the Knolens knowledge graph without passing the validated ingestion pipeline. The zero-trust principle is not a governance policy layered over the AI tool. It is the foundational architectural constraint from which the entire system is built.
9. How Fast Can Your Team Switch to Source-Controlled AI Research?
Transitioning from generic AI tools to source controlled AI research with KnolAI does not require your team to rebuild their research workflows from scratch. KnolAI integrates into existing HEOR, CI, and regulatory research workflows as the governed intelligence layer that replaces the uncontrolled source risk of generic AI tools. Most teams are running their first fully source-controlled research outputs within two weeks of onboarding.
Sprint 1, Weeks 1 to 2, Validated knowledge layer active: The Knolens knowledge graph is loaded with validated clinical, regulatory, and HTA content for your indication. Every source has been verified through the ingestion validation pipeline. Your team runs the first KnolAI queries and receives outputs where every claim links to a verified source. The fabricated citation risk is eliminated from the first session.
Sprint 2, Weeks 3 to 4, Source verification integrated into submission workflows: KnolAI's source-linked evidence tables are configured for your primary submission use cases. NICE evidence table format, G-BA dossier extraction tables, and ICER evidence report sections are generated from the validated knowledge base with claim-level source links embedded. Manual post-generation source verification is replaced by structural source control.
Sprint 3, Weeks 5 to 6, Continuous source quality monitoring live: The Knolens continuous monitoring layer tracks new publications and source updates for your indication. When a new trial result is published that updates a fact in the knowledge base, the existing triple is flagged for update review. When a previously cited publication is corrected or retracted, the affected triples are flagged and removed from active use until reviewed. The source quality of your research outputs is maintained continuously, not audited periodically.
Conclusion
The fabricated citation crisis documented by The Lancet in 2026 is not an abstract threat to scientific integrity. It is a direct operational risk for every pharma team that uses a generic AI tool to assist with research, evidence synthesis, or document preparation. One in 277 papers published in early 2026 contains a fabricated citation. Review articles, the papers that form the evidence base for clinical guidelines and HTA submissions, show a 57% higher fabrication rate than other paper types.
At Pienomial, we built source controlled AI research into KnolAI because we believe that in pharma, the only research AI that is actually useful is research AI whose every output can be verified against a real, accessible primary source. Speed without source control is not a productivity gain. It is a liability that grows with every output produced. KnolAI delivers the speed of AI-assisted research with the source integrity that regulated life sciences demands.
CTA: See how KnolAI eliminates source risk from your research workflows. Book a demo with the Pienomial team.
