Toxicology is deliberately slow for good reason. A poor decision quickly becomes a product, public health, regulatory, or legal issue—often with irreversible consequences.
However, while toxicology's principles endure, the field is no longer working in the same landscape.
What has changed is the decision-making environment. As noted previously, safety assessment now spans a more complex and diverse body of evidence—including legacy animal data, in vitro systems, computational models, and mechanistic insights. The expectation remains: conclusions must be transparent, traceable, and defensible.
Against this backdrop of accelerating change and complexity, the rise of artificial intelligence highlights the need for careful assessment.
Not because AI lacks potential, but because toxicology cannot risk misplaced confidence.
This article marks the first in our AI in Safety Assessment series, where we examine how artificial intelligence can be integrated into toxicology workflows with scientific discipline, regulatory awareness, and practical realism.
A Discipline Defined by Consequence
Other fields judge automation by efficiency. Toxicology uses different criteria.
Every conclusion must withstand:
- Scientific scrutiny
- Regulatory review
- Legal challenge
- Long-term real-world validation
This causes natural resistance to overstatement. It’s not conservatism—it’s accountability.
Toxicology workflows handle imperfect and conflicting evidence. Decisions are rarely binary and require interpretation and explicit management of uncertainty.
This is where much of the current AI narrative becomes problematic.
Assuming more data processing leads to better decisions is flawed. The challenge is not accessing information, but interpreting it.
AI Enters a Constrained System
As discussed previously, artificial intelligence is already being evaluated across regulatory and industry environments.
Programmes at the FDA, EFSA, and EPA show a shift from exploration to structured implementation. Industry is exploring how AI reduces manual workload in literature screening, data extraction, and dossier preparation.
This shift is already visible in regulatory practice. The FDA’s AI4TOX programme includes initiatives such as AnimalGAN for synthetic data generation, SafetAI for predictive modelling, and BioBERT for document analysis. EFSA’s AI@EFSA and AI4NAMS initiatives similarly focus on structured evidence extraction and integration of non-animal methodologies. The U.S. EPA is also piloting generative AI within study evaluation workflows.
The direction of travel is clear.
Yet the tone of these initiatives remains measured.
Industry discussions reflect the same measured approach. ECETOC’s 2024 workshop on AI in chemical safety assessment focused on data quality, governance, and regulatory trust rather than capability alone. Strategic analyses from McKinsey similarly highlight that AI delivers value only when embedded into structured workflows supported by curated data and scientific oversight.
Across regulators and scientific forums, AI is consistently framed as:
- An assistive capability
- A tool for improving evidence handling
- A means to increase consistency and throughput
It does not replace scientific reasoning.
This critical distinction is often overlooked in AI discussions.
AI may raise the floor on tasks, but cannot replace judgment.
Where Hype Collides with Reality
The gap between expectation and reality becomes most visible when examining how toxicology actually works.
Imperfect Evidence Is the Norm
Toxicology does not operate on clean, complete datasets. Evidence is:
- Fragmented across studies and sources
- Generated under varying experimental conditions
- Often limited to novel or innovative substances
AI processes fragmented information efficiently but cannot overcome fundamental limitations in evidence. More data does not guarantee clearer conclusions.
This limitation is already visible in practice. Recent 2025 NGRA-aligned research and regulatory guidance confirm that non-animal safety assessments are increasingly built on integrated NAM workflows combining in silico modelling, PBPK/QIVIVE frameworks, and mechanistic bioactivity data, with a strong emphasis on quantitative uncertainty characterization (e.g., probabilistic modelling, inter-method variability) and transparent, structured reporting of assumptions and data limitations to support regulatory decision-making.
Relevance Is Not a Technical Problem
Determining whether a study is applicable requires:
- Understanding of exposure context
- Knowledge of study design limitations
- Interpretation of biological significance
These require judgment from domain experts—not pattern recognition.
Uncertainty Must Be Explicitly Managed
Regulatory toxicology does not eliminate uncertainty—it characterises it.
This includes:
- Defining applicability domains
- Identifying data gaps
- Justifying assumptions
- Explaining limitations
AI can support, but not replace, responsibility for defining and communicating uncertainty — which is why structured toxicology consulting remains the backbone of defensible safety conclusions, regardless of the tools used
The Risk of Overclaiming
One of the fastest ways to lose trust in toxicology is to overstate certainty.
The same applies to AI.
Calling AI “automated assessment” or “decision-making” mismatches capability with expectations. When exposed, especially under regulatory scrutiny, credibility is hard to restore.
Credible regulatory and industry voices now adopt a more restrained stance:
AI is valuable when it is:
- Interpretable
- Fit-for-purpose
- Validated against real tasks
- Governed with clear oversight
AI loses value when lacking transparency or accountability.
A More Realistic View of Value
Without hype, areas of genuine AI value become clear.
AI can improve:
- Evidence discovery and literature screening
- Deduplication and organisation of large datasets
- Structured extraction of key study attributes
- Integration of data into standardised formats
These are real improvements. They help manage the volume and complexity of evidence in modern toxicology — as demonstrated in practice in our case study on enhancing chemical safety and compliance through strategic collaboration, where structured workflows and expert oversight delivered measurable gains for a global consumer goods company.
At the same time, the discipline’s core remains unchanged.
AI does not replace:
- Weight-of-evidence evaluation
- Contextual interpretation of findings
- Determination of critical endpoints
- Construction of defensible conclusions
This boundary is a necessary control, not a limitation.
Alignment Across Regulators and Industry
One of the most notable developments is the level of alignment across stakeholders.
Regulators emphasise:
- Transparency and explainability
- Validation and performance evaluation
- Risk-based governance
Industry focuses on:
- Data readiness and curation
- Integration into existing workflows
- Maintaining human oversight
Scientific forums reinforce:
- The importance of trust and interpretability
- The need for domain-constrained systems
- The risks of overgeneralised AI applications
This convergence reflects a shared truth: in toxicology, credibility comes from discipline—not speed. That discipline extends beyond the lab into how companies approach scientific and regulatory decision support as a structured function, not a reactive one.
Reframing the Question
The central question is not whether AI should be used.
It is:
Where does AI strengthen toxicology, and where must its role remain limited?
This requires moving beyond vague capability discussions and focusing on use cases, constraints, and validation.
That is the focus of the next articles in this series.
Moving Forward
AI in toxicology is not one solution. It encompasses predictive modeling, language processing, and generative systems—each with unique strengths and limits.
Recognizing these differences is essential.
In a consequence-driven field, misunderstanding capability isn’t theoretical; it has real costs.
Final Perspective
Toxicology does not resist innovation. It evaluates it under constraint.
Artificial intelligence will play an increasing role in safety assessment workflows. However, its long-term value will not be determined by how quickly it is adopted, but by how carefully it is applied.
The discipline does not need faster conclusions, but better-supported ones.
It needs better-supported ones.
And that distinction will determine whether AI strengthens toxicology—or undermines it.
