The Ethics of AI in Scientific Discovery: Progress Without Principles?

Introduction

As artificial intelligence transforms scientific discovery—accelerating drug development, designing novel molecules, and automating research—we face profound ethical questions. Who owns discoveries made by AI? How do we ensure equitable access to these powerful tools? What happens when AI makes predictions we don't understand? And how do we prevent misuse of technologies that can design both medicines and toxins?

These aren't abstract philosophical puzzles—they're urgent practical questions that will shape the future of science and society. As we embrace AI's potential to accelerate discovery, we must simultaneously develop ethical frameworks to guide its responsible development and deployment.

Authorship and Attribution

Who Deserves Credit?

When an AI system designs a novel drug or discovers a new material, attribution becomes murky:

Traditional science:

• Authors listed on papers
• Credit reflects intellectual contribution
• Clear chain of discovery

AI-assisted science:

• Did the researcher discover it, or did the AI?
• What about the engineers who built the AI?
• Those who generated training data?
• Funders who enabled the research?

Current Practices

Emerging norms:

• AI as tool, not author
• Researchers remain responsible
• Disclosure of AI role required
• Methods section details AI contribution

But complications arise:

• When AI does most creative work
• When multiple AIs contribute
• When AI generates unexpected insights
• Attribution chains become complex

Intellectual Property

Patent questions:

• Can AI be named as inventor? (Currently: No in most jurisdictions)
• Who owns AI discoveries?
• How to handle prior art generated by AI?

Recent cases:

• DABUS AI inventor applications rejected
• But human-guided AI discoveries patentable
• Legal landscape still evolving

Access and Equity

The Resource Gap

AI in science requires:

• Massive computational infrastructure
• Large training datasets
• Specialized expertise
• Significant funding

Concentrates power in:

• Major tech companies (Google, Microsoft, Meta)
• Well-funded academic institutions
• Wealthy countries

The Democratization Challenge

Risk: AI accelerates discovery primarily for those already advantaged

Consequences:

• Widening gap between resource-rich and resource-poor institutions
• Neglected diseases remain neglected (no profit incentive)
• Global South excluded from AI benefits
• Brain drain to companies offering resources

Efforts Toward Equity

Open-source initiatives:

• AlphaFold freely available
• ESM models open-sourced
• OpenMolecules project
• Shared datasets and tools

Cloud computing access:

• Cloud credits for researchers
• Compute time donations
• Collaborative facilities

Capacity building:

• Training programs in developing countries
• International collaborations
• Technology transfer

But challenges persist:

• Latest models remain proprietary
• Compute costs still prohibitive for many
• Expertise gap widening

Reproducibility and Transparency

The Black Box Problem

Deep learning models:

• Millions to billions of parameters
• Complex, non-interpretable
• Difficult to understand why they make predictions

Scientific implications:

• How to verify AI reasoning?
• Can we trust predictions we don't understand?
• What happens when model fails in unexpected ways?

Reproducibility Concerns

Challenges:

• Stochastic training (different runs → different models)
• Hyperparameter sensitivity
• Data versioning
• Computational environment dependencies

Mitigations:

• Detailed methods reporting
• Code and model sharing
• Containerization (Docker, etc.)
• Seed setting for reproducibility

Pre-Registration and Transparency

Proposed practices:

• Pre-register AI experiments (like clinical trials)
• Disclose negative results
• Share failed models, not just successes
• Document data provenance

Bias and Fairness

Training Data Bias

AI inherits biases from training data:

Drug discovery example:

• Most clinical trials historically on white males
• Models trained on this data
• Predictions less accurate for women and minorities
• Perpetuates health disparities

Materials science example:

• Databases reflect researcher interests
• Certain material classes over-represented
• AI focuses on already-studied areas
• Novel chemistries underexplored

Algorithmic Bias

Beyond data:

• Objective functions encode values
• Optimization priorities reflect choices
• What we measure shapes what we find

Example: Drug design optimizing for:

• Efficacy (helps everyone)
• Patent novelty (helps companies)
• Manufacturing cost (affects affordability)
• Trade-offs embed ethical decisions

Addressing Bias

Strategies:

• Diverse training data
• Fairness metrics
• Algorithmic audits
• Diverse research teams
• Community input

Dual-Use Concerns

Potential for Misuse

Same tools can create:

• Medicines or toxins
• Vaccines or bioweapons
• Beneficial materials or hazardous substances

AI lowers barriers:

• Less expertise needed
• Faster development
• Easier to hide intentions

Real Examples

Recent concerns:

• Drug discovery models trivially repurposed to design toxins
• Published study showed 40,000 toxic molecules generated in hours
• Pandemic pathogen prediction could inform bioweapon design

Biosecurity risks:

• Synthesis of dangerous pathogens
• Optimizing viral transmissibility
• Evading detection or treatment

Governance Approaches

Possible measures:

• Publication filtering (redacting details)
• DNA synthesis screening
• Export controls on AI models
• Researcher vetting
• Ethics review boards

But tensions:

• Scientific openness vs. security
• Beneficial applications vs. risks
• International cooperation vs. control

Environmental Impact

Computational Carbon Footprint

AI training is energy-intensive:

• GPT-3: ~1,300 MWh (equivalent to 550 tons CO₂)
• AlphaFold training: ~$100M compute
• Ongoing inference costs

Tradeoff analysis:

• Is one AI-discovered drug worth the carbon cost?
• Compared to traditional R&D emissions?
• Net environmental impact unclear

Experimental Waste

High-throughput screening:

• Millions of experiments
• Chemical waste
• Plastic consumables
• Energy consumption

Optimization:

• Better experiment design reduces waste
• But increased throughput may increase total consumption

Scientific Integrity

P-Hacking and Overfitting

AI makes it easy to:

• Try millions of models
• Overfit to noise
• Find spurious correlations
• Report only successes

Safeguards needed:

• Held-out test sets
• Prospective validation
• Pre-registration
• Negative result reporting

Replication Crisis

AI may exacerbate:

• Hype of preliminary results
• Pressure to publish positive findings
• Difficulty replicating complex models
• Opaque methodologies

Or help solve it:

• Automated replication
• Standardized protocols
• Larger-scale validation
• Transparent workflows

Informed Consent and Privacy

Patient Data

AI drug discovery uses:

• Clinical trial data
• Electronic health records
• Genomic information
• Imaging data

Questions:

• Did patients consent to AI use?
• Re-identification risks?
• Benefit sharing?

Data Sovereignty

Who controls biological data?

• Individuals?
• Institutions?
• Countries?

Biopiracy concerns:

• Genetic resources from developing countries
• Traditional knowledge
• Benefit sharing from discoveries

The Control Problem

Autonomous Discovery Systems

As AI systems become more autonomous:

Concerns:

• Loss of human oversight
• Unintended consequences
• Alignment with human values
• Accountability for failures

Guardrails:

• Human-in-the-loop requirements
• Constraint satisfaction
• Interpretability tools
• Kill switches

Dependency Risks

Over-reliance on AI:

• Loss of human expertise
• Vulnerability to model failures
• Single points of failure
• Deskilling of researchers

Regulatory Challenges

Existing Frameworks Inadequate

Traditional regulation assumes:

• Human researchers
• Interpretable methods
• Slower pace
• Defined risk categories

AI changes:

• Speed of discovery
• Opaque reasoning
• Novel risk types
• Blurred boundaries

Adaptive Governance Needed

Proposals:

• Agile regulatory frameworks
• Regulatory sandboxes
• International coordination
• Stakeholder participation

Examples:

• FDA exploring AI drug regulation
• WHO guidance on AI in health
• OECD AI principles

Responsibilities of Different Stakeholders

AI Developers

Obligations:

• Document capabilities and limitations
• Test for harmful use cases
• Enable interpretability
• Support responsible deployment

Researchers Using AI

Duties:

• Understand tool limitations
• Validate computations experimentally
• Attribute appropriately
• Report failures

Institutions

Roles:

• Ethics review for AI projects
• Training in responsible AI use
• Data governance policies
• Equity considerations

Journals and Publishers

Responsibilities:

• Require AI disclosure
• Reproducibility standards
• Code/model sharing
• Negative results publication

Funders

Leverage:

• Require ethical review
• Support open science
• Fund governance research
• Incentivize equity

Policymakers

Needs:

• Evidence-based regulation
• International cooperation
• Balance innovation and safety
• Ensure public benefit

Toward Ethical AI in Science

Principles

Emerging consensus:

1. Beneficence: Maximize societal benefit
2. Non-maleficence: Minimize harm
3. Autonomy: Preserve human agency
4. Justice: Ensure equitable access
5. Explicability: Enable understanding
6. Accountability: Clarify responsibility

Practical Implementation

Concrete steps:

• Ethics training for AI researchers
• Impact assessments before deployment
• Diverse teams and perspectives
• Continuous monitoring and evaluation
• Adaptive management

Value Alignment

Embedding values in AI:

• What goals do we optimize for?
• Whose values?
• How to handle value pluralism?
• Technical and social challenge

Conclusion

The ethics of AI in science are not obstacles to progress—they are essential to ensuring that progress benefits humanity broadly and enduringly. As AI systems become more powerful and autonomous, the stakes of getting this right increase.

We need not—should not—slow scientific discovery to address ethics. Rather, we must develop ethics at the same speed we develop technology. This requires proactive engagement from all stakeholders: researchers, institutions, companies, policymakers, and the public.

The promise of AI in science—curing diseases, solving climate change, understanding the universe—is too great to squander through shortsightedness. But realizing that promise demands that we ask not only "can we?" but "should we?"—and design our systems accordingly.

At digital speed, we're discovering faster than ever. We must ensure we're discovering wisely.

References

1. Stokes, J. M. et al. (2020). A deep learning approach to antibiotic discovery. Cell, 180(4), 688-702.
2. Urbina, F. et al. (2022). Dual use of artificial-intelligence-powered drug discovery. Nature Machine Intelligence, 4, 189–191.
3. Jobin, A., Ienca, M., & Vayena, E. (2019). The global landscape of AI ethics guidelines. Nature Machine Intelligence, 1, 389–399.
4. UNESCO (2021). Recommendation on the Ethics of Artificial Intelligence.