Artificial intelligence is increasingly being used to support toxicological literature searches, data extraction and preparation of Health-Based Exposure Limit assessments. A new paper published on 12 October 2026 argues that these tools should be governed according to their actual influence on the HBEL decision rather than automatically subjecting every supporting AI application to a full computerized-system-validation package.
The article, Risk-based governance of AI-tools used in health-based exposure limit derivation, was published in Regulatory Toxicology and Pharmacology.
It addresses a highly practical GMP question:
What level of validation and governance is appropriate when AI assists qualified toxicologists in establishing PDEs, ADEs or occupational exposure limits?
Why HBELs Matter in Pharmaceutical GMP
Health-Based Exposure Limits are central to risk management in shared pharmaceutical manufacturing facilities.
Examples include:
- Permitted Daily Exposure – PDE;
- Acceptable Daily Exposure – ADE;
- Occupational Exposure Limit – OEL.
HBELs can support decisions concerning:
- cross-contamination risk;
- cleaning limits;
- facility and equipment sharing;
- containment;
- operator protection;
- product-specific risk controls.
The resulting limit is not produced by a simple calculation alone.
It requires integration of toxicological, pharmacological, clinical and other scientific evidence with expert judgment.
Why Companies Are Exploring AI
HBEL derivation can require review of large and heterogeneous evidence sources.
AI tools can potentially assist with:
- literature identification;
- screening of search results;
- extraction of study information;
- organisation of toxicological evidence;
- identification of possible critical effects;
- drafting of structured sections;
- consistency checks.
For organisations managing large portfolios of active substances, intermediates and cleaning programmes, these efficiencies can be attractive.
But the paper identifies several new risks.
The Main Risks Are Not Only Numerical Errors
AI-supported toxicological work can introduce problems such as:
- hallucinated publications;
- incorrect citations;
- misattributed study findings;
- loss of important context;
- overconfident interpretation;
- scientifically shallow but fluent narrative.
These risks are particularly serious in HBEL derivation because the output can influence GMP cross-contamination controls.
A plausible-sounding toxicological conclusion is not sufficient.
Every critical scientific statement needs to remain traceable to reliable source evidence.
A Key Distinction: The HBEL Monograph Is an Expert Scientific Assessment
The paper makes an important regulatory distinction.
It argues that an HBEL monograph is primarily:
a scientific expert assessment that informs GMP Quality Risk Management.
It is not automatically equivalent to a transactional electronic GxP record generated directly by a manufacturing or laboratory system.
This distinction matters when determining the appropriate level of computerized-system assurance.
According to the author, the use of an AI tool during HBEL preparation does not automatically mean that the complete tool must always undergo the same full CSV approach expected for a critical operational GMP system.
This Is Not an Exemption from Control
The paper’s argument should not be misunderstood.
It does not mean that any public AI tool can be used without assessment.
It means that control should be proportional to:
- the intended use of the AI;
- the importance of the task it performs;
- the degree of human verification;
- the potential effect of an AI error on the final HBEL.
The relevant sequence is:
AI Function → Influence on Scientific Assessment → Detectability of Error → Human Review → Required Assurance
A Practical Risk-Tier Interpretation
A practical implementation of this principle could distinguish several levels of AI use.
Lower influence
AI assists with administrative or organisational tasks such as:
- formatting;
- language editing;
- organising references;
- creating non-substantive summaries for later verification.
The final scientific assessment does not depend on the AI output without independent review.
Intermediate influence
AI assists with:
- literature screening;
- data extraction;
- classification of evidence;
- preparation of structured tables.
Errors could omit or misrepresent relevant evidence, so source-level verification is required.
Higher influence
AI contributes to:
- selection of the critical effect;
- selection of the point of departure;
- interpretation of uncertainty factors;
- derivation of the numerical limit;
- final scientific conclusions.
These activities have direct influence on the HBEL and require the strongest controls and expert scrutiny.
This tiering is a practical interpretation of the paper’s risk-proportionate principle, not a formal regulatory classification issued by EMA or FDA.
Source Verification Is the Central Control
For AI-assisted HBEL work, the most important control may be verification against the original scientific evidence.
A qualified toxicologist should be able to confirm:
- that the cited study exists;
- that the study population and design are described correctly;
- that doses and findings are accurate;
- that the critical effect has not been misrepresented;
- that contradictory evidence has not been omitted;
- that the source supports the conclusion.
AI-generated references should therefore be treated as unverified until checked against the original source.
Human Toxicologist Review Remains Non-Negotiable
The paper states that experienced human toxicologist review and peer review remain the primary quality controls.
This is important because HBEL derivation involves scientific judgments that cannot be reduced safely to text generation.
These include:
- weight-of-evidence assessment;
- selection of a critical toxicological endpoint;
- interpretation of clinical relevance;
- selection and justification of adjustment factors;
- evaluation of uncertainty;
- assessment of data gaps.
AI may support these activities.
It should not silently replace professional accountability for the final limit.
Peer Review Reduces Dependence on the AI Tool
Independent peer review provides an additional control layer.
The reviewer should examine:
- the source evidence;
- the scientific reasoning;
- the calculation;
- the final conclusion.
This means the reviewer assesses the HBEL assessment itself rather than merely confirming that an AI workflow ran successfully.
A strong expert-review process can therefore reduce the need to rely exclusively on internal model transparency.
The Deliverable and the Supporting Tool Should Be Assessed Separately
One useful governance principle is to distinguish:
quality of the HBEL monograph
from:
qualification of the AI tool.
The HBEL monograph needs to be:
- scientifically sound;
- complete;
- traceable;
- transparent;
- independently reviewed.
The AI tool needs controls proportionate to how strongly it influences that deliverable.
A tool used only to help organise references does not necessarily require the same evidence as a model that proposes the numerical PDE.
Full CSV Is Not Automatically Required – But Applicability Must Be Justified
The paper argues that full computerized-system validation is not inherently required for every AI tool supporting HBEL derivation.
This is an important but potentially sensitive conclusion.
Whether Annex 11 or other GxP computerized-system requirements apply depends on factors such as:
- the intended use of the system;
- whether it forms part of a regulated GMP process;
- whether its output is used without independent verification;
- its influence on product-quality risk decisions;
- how it is integrated with controlled systems.
The decision not to apply a full CSV approach should therefore be documented and justified through risk assessment.
It should not be based only on convenience.
Draft Annex 22 Remains Relevant
The paper draws on the principles of draft EU GMP Annex 22.
Relevant concepts include:
- defined intended use;
- Quality Risk Management;
- data quality;
- performance assessment;
- documentation;
- human oversight;
- lifecycle control.
However, draft Annex 22 was developed primarily for AI models embedded in critical GMP computerized-system applications.
An AI tool supporting an expert toxicological assessment may occupy a different regulatory position.
That is why Context of Use and influence on the final decision are essential.
ICH M7 Provides a Useful Precedent
The paper also refers to ICH M7, under which computational toxicology models can support assessment of mutagenic impurities.
This provides an important precedent:
regulatory acceptance of a computational tool can be based on defined applicability, model qualification, transparent documentation and expert interpretation rather than a universal validation template.
The same general logic may help inform governance of AI-assisted HBEL tools.
Data Confidentiality Also Requires Control
HBEL assessments may use:
- internal toxicology studies;
- unpublished clinical information;
- proprietary compound data;
- confidential manufacturing information.
These data should not be entered into an uncontrolled public AI platform.
Organisations should assess:
- where prompts and documents are processed;
- whether inputs are retained;
- whether supplier models learn from submitted data;
- who can access the information;
- what contractual safeguards exist.
An AI tool may be scientifically useful but still unsuitable because its data-handling model is unacceptable.
Traceability Should Focus on Material Scientific Contributions
The level of recordkeeping should also be risk proportionate.
Where AI materially influences the scientific assessment, it may be useful to retain information such as:
- AI system and model version;
- task or prompt;
- source documents supplied;
- substantive AI output;
- human corrections;
- review and approval.
This does not necessarily require retaining every exploratory interaction.
The objective is to reconstruct material contributions to the final HBEL.
A Practical Governance Model
A defensible workflow could be:
Approved AI Tool
↓
Defined Permitted Tasks
↓
Controlled Source Information
↓
AI-Assisted Retrieval or Drafting
↓
Source-Level Verification
↓
Qualified Toxicologist Assessment
↓
Independent Peer Review
↓
Approved HBEL Monograph
This places AI inside the existing scientific-quality framework rather than allowing it to become an independent decision-maker.
Why This Paper Matters
The paper addresses a common problem in pharmaceutical AI governance.
One extreme is to assume that any AI use automatically requires a full validation package equivalent to a critical GMP system.
The other is to treat AI as an informal productivity tool requiring no control.
The proposed alternative is:
risk-proportionate assurance based on Context of Use, decision influence, source verification and expert accountability.
This may be a more practical model for many AI-assisted scientific assessments.
Regulatory Status Note
The publication is an independent scientific article and does not constitute EMA, FDA, ICH or PIC/S guidance.
Its conclusion that full CSV is not inherently required for all AI tools used in HBEL derivation is the author’s regulatory interpretation.
It does not create an exemption from applicable GMP, Annex 11, data-integrity or quality-system requirements.
Each organisation remains responsible for determining and documenting the controls appropriate to its specific AI use and workflow.
Sources
Lovsin Barle E. – Risk-based governance of AI-tools used in health-based exposure limit derivation. Regulatory Toxicology and Pharmacology. Published online 12 October 2026. Article 106275.
https://pubmed.ncbi. … lm.nih.gov/42863558/
DOI:
https://doi.org/10.1 … /j.yrtph.2026.106275
EMA – Guideline on setting Health-Based Exposure Limits for use in risk identification in the manufacture of different medicinal products in shared facilities:
https://www.ema.euro … ts-shared-facilities
FDA – Considerations for the Use of Artificial Intelligence to Support Regulatory Decision-Making for Drug and Biological Products: