Viruses, Vectors, and Villains: Governing the Risks and Rewards of Artificial Intelligence in Virology

Abstract

Artificial intelligence (AI) is rapidly transforming virology by strengthening pandemic preparedness, enhancing our molecular understanding of virus–host interactions, and accelerating the discovery and development of novel antiviral therapies. Yet, the same technologies also pose urgent biosecurity risks, particularly by enabling the development of bioweapons or identifying strategies that maximize harm. This paper presents a critical content analysis of current and emerging AI applications in virology, including tools used to detect synthetic alterations in viral genomes, assess the severity of new variants, and design clinical vectors for gene therapy. It also highlights the potential for misuse, whether intentional or due to poor data quality and flawed model training. Drawing on case studies, public databases, and documented applications from research institutions and biotechnology firms, the analysis shows that AI can integrate large datasets to reduce reliance on animal testing in drug development, improve therapeutic precision, and allocate resources more effectively during outbreaks. However, the increasing accessibility of AI tools and genomic data also creates vulnerabilities, especially as models become capable of autonomously interpreting the scientific literature and mining bioinformatics databases. To address this dual-use dilemma, the paper proposes targeted and adaptable policy recommendations for governments, research institutions, and commercial biotech firms, emphasizing pre-emptive oversight, responsible innovation, and ethical AI deployment. These recommendations are designed for immediate relevance yet flexible enough to evolve alongside the expanding role of AI in global health.

1. Introduction

“Technology is neither good nor bad; nor is it neutral.”

—Melvin Kranzberg’s first law of technology.

“If we’re bad people we use technology for bad purposes and if we’re good people we use it for good purposes.”

—Herbert A. Simon.

These quotes, coined before the widespread use of artificial intelligence (AI), take on new meaning in the field of virus research. AI is revolutionizing virology, providing groundbreaking tools for pandemic preparedness by predicting viral evolution and guiding outbreak responses, improving our understanding of virus–host interactions, and accelerating the development of novel antivirals (Figure 1). However, alongside these benefits lies an urgent biosecurity dilemma: the same AI-driven models that enhance disease detection and public health responses can also be repurposed to design synthetic pathogens, optimize viral transmission, and evade detection. This dual-use potential is not incidental but embedded in the computational framework itself; a single line of code can shift an algorithm from minimizing a pathogen’s spread to increasing it. This presents an immediate governance challenge as AI becomes more accessible across research institutions, commercial biotechnology firms, and non-state actors.

AI’s integration into life sciences has yielded remarkable benefits, from pandemic tracking to precision medicine. Yet, its democratization and dual-use potential demand proactive governance. Rather than stifling innovation, policymakers must pre-empt misuse while fostering AI’s benefits in global health and security. Despite international agreements addressing biological weapons, AI’s rapid advancement is outpacing regulatory frameworks, leaving gaps in oversight. Current governance mechanisms are also not equipped to regulate flawed models trained on biased datasets that exacerbate health inequalities or to counter AI-driven biosecurity threats. Lowering the technical barriers to harm also risks fueling public distrust in research institutions, especially in the wake of the COVID-19 pandemic.

To mitigate these challenges, this paper proposes targeted governance measures that integrate ethical AI standards into global health policy, emphasizing the need for multilateral cooperation and UN-led oversight in major areas for policy regulation: the establishment of proper protocols for training AI in medical research and diagnostics, central oversight framework for firms providing synthetic biology services that could generate lethal pathogens, and integration and automated processing of viral surveillance networks.

While not intended as an exhaustive systematic review, this paper employs a critical content analysis of the AI–virology landscape to establish the evidentiary basis for the governance framework proposed. The selection of literature and AI architecture was guided by three core objectives: (1) identifying high-impact AI tools currently utilized in clinical and research virology; (2) assessing the dual-use potential of generative models in pathogen design; and (3) evaluating the technical feasibility of proposed oversight mechanisms.

The analysis was structured around a representative survey of multidisciplinary research indexed in PubMed, bioRxiv, and technical AI repositories (e.g., arXiv), primarily between 2018 and 2025. Search queries focused on the intersection of “AI-assisted virology,” “biosecurity,” and “generative protein design.” Rather than a comprehensive bibliography, tools were selected based on their prominence in the field and their capacity to illustrate specific “risk–reward” dualities. These findings were synthesized thematically into the four functional domains discussed in the following sections: diagnostics, surveillance, drug discovery, and synthetic design.

2. Background

The integration of AI into the life sciences began in the 1970s with systems like MYCIN, which was developed to diagnose bacterial infections from basic input and recommend treatments [1]. While limited by the computing power of the time, these early systems demonstrated AI’s potential to augment medicine. The field entered a new era in the 2010s with the rise of machine learning, large language models, and neural networks, hereby grouped under the umbrella term AI for ease of discussion. Models such as AlexNet in 2012 showcased AI’s ability to perform image recognition tasks [2], paving the way for applications in medical imaging diagnostics while normalizing the use of “black boxes,” that is, millions to billions of internal computational parameters beyond the complete comprehension of human supervision. Today, tools like BioGPT parse the vast scientific literature to generate hypotheses, further blurring the line between human and machine-driven research [3].

In 2021, AlphaFold transformed computational biology by predicting protein structures from chemical composition [4], accelerating drug discovery from structural biology without years of experimentation. In drug development, generative AI designs novel molecules, and models repurpose existing drugs, such as BenevolentAI’s identification of baricitinib as a COVID-19 treatment [5]. The use of AI in medical diagnostics presents several challenges. While algorithms detect tumors in medical images with accuracy rivaling human experts [6], a review of 415 studies using models to diagnose COVID-19 found only 62 passed quality standards, and none were suitable for clinics due to flaws in design and training datasets [7].

In genomics, computation has become indispensable for analyzing viral evolution and predicting emerging threats. Tools like DeepVariant improve the identification of variants from genome sequencing [8], while platforms such as Nextstrain provide real-time tracking of pathogen spread, as during the COVID-19 pandemic [9]. These applications enable health agencies to respond quickly to outbreaks. For biodiversity, AI expanded the annotated fraction of viral protein families by nearly a third [10].

The accessibility of AI tools has democratized research but also amplified risks. Open-source protein biology models like ESM [11] and public databases of emerging viral variants such as GISAID [12] empower scientists worldwide. However, they could be used in combination to find clinics with new viral isolates that can be propagated for misuse. As a proof-of-concept for database misuse, researchers in 2022 used the publicly available MegaSyn tool to generate 40,000 potential toxins in hours, including deadlier derivatives of agents banned under the Chemical Weapons Convention [13].

Historical precedents illustrate the dual-use dilemma. The 2011 H5N1 gain-of-function studies, which enhanced the transmissibility of avian flu in mammals, sparked global debate about balancing scientific freedom with biosafety [14,15]. Lax screening by gene synthesis firms allows orders for pathogenic sequences to go unchecked [16], even decades after synthetic material was used to generate infectious viruses [17,18,19]. The Virology Capabilities Test recently demonstrated that several common models, such as ChatGPT and Google’s Gemini, could substantially outperform expert researchers in troubleshooting lab procedures, such as how to grow and quantify viruses [20]. These examples underscore the urgent need for international governance structures capable of monitoring AI-assisted biological research while balancing scientific innovation and security. As AI continues to lower technical barriers, policymakers must develop multilateral oversight mechanisms to regulate AI-driven virology research, ensuring ethical deployment while mitigating misuse. To provide a structured evaluation of the current landscape,

Table 1. Analysis of AI integration in global virology.

Strengths Accelerated discovery of antivirals. High-accuracy protein folding predictions. Automated, real-time genomic surveillance. Enhanced detection of synthetic viral modifications.	Weaknesses “Black-box” nature of AI decision-making in clinical settings. Heavy reliance on high-quality, non-biased training datasets. High computational costs and energy requirements for AI processing. Lack of interpretability in complex neural network outputs.
Opportunities Creation of a UN-led global biosecurity clearinghouse. Standardized synthetic biology screening protocols. Integration of environmental metagenomics with AI forecasting. Bridging the diagnostic gap in resource-limited settings.	Threats Democratization of dual-use design for malicious actors. Rapid evolution of AI outpacing international law/policy. Potential for “hallucinations” in generative biological models. Misalignment of private sector incentives with global safety.

summarizes the internal functional capacities (strengths/weaknesses) and external environmental factors (opportunities/threats) governing AI’s role in virology.

3. Governance of AI Use in Clinical Virology Diagnostics

AI is increasingly used in clinical virology to enable faster diagnostics, outbreak forecasting, and therapeutic discovery. Tools such as convolutional neural networks for medical imaging, transformer-based models for viral genome classification, and deep generative models for drug–target interaction prediction are increasingly embedded in clinical workflows. However, regulatory frameworks have not kept pace with these innovations, resulting in fragmented oversight, inconsistent performance evaluation, and risks to clinical reliability and patient safety.

Trained models have the potential to revolutionize speed and accuracy in diagnostics, even by inferring the presence of a pathogen by measuring patient data without directly detecting pathogen material [21,22,23]. However, many tools have been developed and validated on narrow datasets, leading to overfitting and poor generalizability across geographic regions or new viral variants. During the COVID-19 pandemic, many AI-based diagnostic tools (e.g., chest CT scan classifiers or serological test predictors) exhibited high variance in performance across populations due to confounding variables not accounted for in the training phase. In fact, of 731 models, 535 had a high risk of bias due to reasons such as inadequate sample sizes or incomplete model evaluation [24]. This raises ethical concerns regarding overconfidence in AI-based results.

Systems used in medical contexts must be explainable to gain clinician trust, though many current systems are black-box architectures. Without clear interpretability, clinicians may hesitate to act on AI-based predictions or misinterpret their outputs. Transparency and explainability of clinical models should be regulated under AI-specific governance policies [25].

AI is now commonly used in the early-stage prediction of antiviral candidates. This includes repurposing small molecules that inhibit viral infection [26], designing new compounds [27] and antibodies [28], or personalizing protocols to patients [29]. While this offers speed during emergencies such as pandemics, many of these in silico predictions lack robust validation prior to being promoted in clinical trials or emergency use pathways. There have been hundreds of deals totaling billions of dollars for AI in the pharmaceutical sector [30], yet hardly any real comparisons of efficacy and safety standards. In the absence of stringent guidelines, such acceleration could lead to wasted resources or off-target toxicity.

Without structured standards for AI use in diagnostics and drug prediction, clinical virology is exposed to unreliable results, unequal access, and unvalidated treatment pathways. The UN and its agencies play a vital role in setting global norms and incentivizing adherence to standards that promote safety and efficacy while accelerating innovation. There are several areas in which policymakers can support AI-based clinical virology:

1. 

Standardize AI Model Development and Validation for Diagnostics.

• Training Dataset Transparency.

  ○

  Developers must disclose training dataset sources, size, and diversity. The metadata of each dataset should be standardized before incorporating multiple databases into the training process. Data should reflect geographic, demographic, and temporal variation in virus circulation.

  ○

  International standards must be set for sample size and tolerable variability between controls, for instance, to avoid comparing scans of infected adults to healthy children, so an algorithm identifies age instead of infection [31].

• Performance Benchmarking.

  ○

  The World Health Organization (WHO) or national labs should establish standardized pathogen reference panels for cross-platform evaluation.

  ○

  Diagnostic platforms should be evaluated by network analysis tools such as those described by the “Foundation for Innovative New Diagnostics” [32].

  ○

  Software competitions awarding successful models should be promoted [33].

• Adaptive Revalidation.

  ○

  Mandate periodic model reassessment, particularly when new viral variants emerge. Models should progress towards desired outcomes, rather than just acceptable ones, for instance, following the rationale in “target product profiles” established by the UK government for COVID-19 test development [34].

2. 

Support Clinically Meaningful AI Integration.

• Context-Dependent Approval.

  ○

  Define specific thresholds for validation, such as rapid triage in emergencies vs. long-term care standards, allowing models with lower sensitivity but high specificity in triage settings to reduce overload during outbreaks.

• Explainability Requirements.

  ○

  Models must include interpretable outputs, such as attention heatmaps, feature contribution rankings, decision trees, and confidence intervals.

  ○

  Experts must determine the balance between performance speed and the availability of transparency outputs [35]. Should logic be given on demand, delaying calculations, or does output speed supersede human understanding in emergencies, with rationale being analyzed afterwards?

• Post-Deployment Feedback Loops.

  ○

  Require monitoring of false positive/negative rates over time and integrate feedback from lab personnel and clinicians.

  ○

  The WHO’s Global Laboratory Initiative could serve as a host for feedback frameworks.

3. 

Preclinical Validation Pipelines for AI-Predicted Drug Candidates.

• Minimum Evidence Standards.

  ○

  Form open, crowdsourced networks modeled after the COVID Moonshot initiative, where labs across institutions collaborated to rapidly identify unpatented antiviral compounds [36] rather than relying on singular experts.

  ○

  Establish standard in vitro testing protocols (e.g., viral replication assays, pseudovirus entry systems) to assess AI-predicted inhibitors.

  ○

  Set guidelines that in silico results alone are insufficient; require at least:

  ▪

  One functional antiviral assay and one test of toxicity for fast-tracked drugs.

  ▪

  One biophysical binding method to target molecules for prolonged clinical use.

• Time-Based Reassessment.

  ○

  Emergency deployments must expire after a set time, say, a reasonable window of 6 months, unless follow-up clinical data are collected.

  ○

  Models must be updated when underlying viral proteins change due to recombination or mutation.

4. 

Emergency Conditional Deployment Mechanisms.

• AI-Evidence-Based Conditional Approval.

  ○

  Permit emergency use if a candidate:

  ▪

  Has clear AI model provenance and performance history;

  ▪

  Shows in vitro antiviral activity when suitable models exist [i.e., infectable cell lines and viral isolates];

  ▪

  Has acceptable toxicity profiles from existing data (e.g., via Open Targets [37] or ChEMBL [38]).

• Clinical Registry and Oversight.

  ○

  Create a WHO or UN-hosted Registry for AI-Prioritized Antivirals, listing:

  ▪

  Chemical structure;

  ▪

  Mechanism-of-action hypotheses;

  ▪

  Source model and dataset details;

  ▪

  Testing status and outcomes.

5. 

Recommended International Support Structures.

A successor project to COVID Moonshot, the AI-driven Structure-enabled Antiviral Platform (ASAP), has several programs funded by the US National Institutes of Health via the Antiviral Drug Discovery (AViDD) U19 Centers, part of the Antiviral Program for Pandemics. However, funds for years 4 and 5 of AViDD centers have been reallocated to other programs [39], highlighting the need for stability stemming from international and multilateral cooperation for such initiatives.

• Global Repository of Validated AI Tools.

  ○

  Hosted by the WHO, Coalition for Epidemic Preparedness Innovations (CEPI), or another neutral body.

  ○

  Includes metadata on AI model validation, geographic bias, and clinical deployment records.

• Capacity Building for Low- and Middle-Income Countries.

  ○

  Provide funding and cloud infrastructure/edge-computing tools to enable AI diagnostic use in low-resource settings.

  ○

  Technologies can be deployed via Google’s Open Health Stack [40] following the WHO’s eHealth Technical Package guidelines [41].

4. AI in Synthetic Virus Biology

Synthetic DNA technologies, particularly the synthesis of viral genomes or genome fragments, have become dramatically more accessible in the past decades due to falling costs, expanding provider networks, and the rise of benchtop synthesizers and design software. These tools offer great promise for public health, providing materials for vaccine and diagnostic development without the need to start from pathogenic material. However, these same capabilities enable the creation or enhancement of pathogenic viruses, raising urgent biosafety and biosecurity concerns.

While synthetic DNA screening frameworks have existed since 2010, notably from the U.S. Department of Health and Human Services [42] and the non-profit International Gene Synthesis Consortium [43], these efforts are voluntary, inconsistently implemented, and insufficiently resourced to meet modern risks. While regulations against the development of specific pathogens exist across the globe, no country legally mandates universal DNA synthesis screening, and current guidance documents remain vague on technical standards and enforcement [16]. Many screening restrictions can be bypassed by ordering single-stranded DNA or multiple pieces that are smaller than the cutoffs for screening mandates [44]. Algorithms can also adapt non-critical residues to avoid high homology to databases of pathogenic material or design flanking sequences with benign fillers that can later be removed. Even within nations such as the US, labs not receiving federal funding are exempted from more recent, broad compliance requirements [45].

Moreover, manual follow-up screening has emerged as a critical bottleneck. Evaluating flagged sequences often requires advanced bioinformatics expertise, institutional familiarity, and labor-intensive contextual review, including scrutiny of a customer’s publication history, institutional affiliation, and intended use case. This step can delay scientific progress and deter companies from implementing robust screening protocols. Firms that voluntarily invest in high-quality screening may incur competitive disadvantages, including higher overhead and longer turnaround times. As synthesis volumes increase and sequences become shorter and more modular, the existing model of case-by-case evaluation is becoming unsustainable.

While more nations are developing voluntary national strategies, such as the UK [46], others rely on non-binding participation in industry-led protocols. Even though the creation of viruses through synthetic biology is a WHO biorisk scenario [47], there is no unified or enforceable global mechanism that aligns incentives, supports capacity building, or ensures parity across borders. Without a globally recognized framework or tangible benefits for good actors, many providers either underperform or outsource critical responsibilities to downstream institutions with inconsistent oversight.

In this context, the UN has an essential role to play. It can coordinate international policy harmonization, foster development and adoption of automated screening pipelines, and realign the economic and political incentives around DNA synthesis oversight. This policy framework focuses on viral sequence synthesis and aims to address these systemic shortcomings without hampering scientific innovation or imposing unrealistic regulatory burdens. These policies can also be a groundwork for more biologically complex organisms, such as bacteria and fungi, whose genomes have or are soon nearing complete synthesis [48,49,50]. The proposed framework has the following goals:

1. 

Strengthening Sequence Screening at the Point of Synthesis.

As emphasized in previous reviews, the most effective point of intervention is the screening of sequences at the synthesis provider level [16,51]. Policies for this should include:

• Customer identity verification and establishment of cases for licensed exemption in virology research. Include third-party customer verification and safety officer contact at the purchasing institution for assistance in reviewing flagged orders [52].
• Sequence-based screening using established databases for organisms of concern that are regularly updated, as well as defining risk levels [53]. A standardized ruleset for the length of sequences screened and the inclusion of single-stranded DNA should be adopted, as company protocols currently vary considerably [54]. There can also be screening for patterns that suggest attempts to avoid detection by ordering multiple small fragments that can later be joined together [55].
• Contextual review when sequences are flagged, including defining a threshold of risk level violations for reporting to authorities.
• Examples of protocols that can be adopted include those from the U.S. Department of Health and Human Services [42,56], International Gene Synthesis Consortium [43], and UK Department for Science, Innovation and Technology [46].
• Establishing an internationally recognized database for pathogens and sequences of concern, for instance, based on the Australia Group control list or the US Federal Select Agent Program [57,58]. This can also include databases of functional elements to proactively detect the synthesis of toxic factors from organisms or artificially designed proteins not yet classified as pathogenic [59].
• Incorporation of built-in safeguards of benchtop synthesizers used by individual labs to produce synthetic DNA [60,61].

2. 

Development and Adoption of Automated Screening Pipelines.

Promoting the integration of automated tools will scale oversight to meet increasing market demand and reduce manual bottlenecks. The following can be considered as starting points:

• Pipelines should be open source so the community can continue to develop them alongside advances in both natural and synthetic biology, such as SeqScreen [62].
• Commercial examples include UltraSEQ [63] from Battelle, offering tiered threat assessments and curated metadata, and SecureDNA [64], which is a free, confidentiality-preserving system for screening orders.

Centralized benchmarking and test datasets are essential for the deployment of any computational tools. The UN, the WHO, or institutions such as the Nuclear Threat Initiative should be encouraged to:

• Develop and host standardized test datasets for screening tools.
• Certify tools against transparent performance benchmarks (e.g., sensitivity to sequences of concern, false positive rates).
• Update these benchmarks regularly to reflect emerging viral threats.

3. 

Incentivizing Voluntary Compliance.

a.

Certification and Recognition Programs.

• Establish a UN-endorsed “Responsible Synthesis Partner” certification for providers who adhere to recognized screening protocols and participate in benchmarking.
• Publicize certified providers in a centralized registry accessible to national biosafety authorities and grant-making agencies.

b.

Economic and Strategic Incentives.

• Encourage government research agencies and public health labs to procure synthetic DNA only from certified providers.
• Provide preferential access to shared sequence databases, updated regulatory lists, or screening pipelines as a benefit for certified participants.
• Offer small grants or tax incentives to companies developing or adopting automated screening infrastructure, particularly in low- and middle-income countries.

4. 

Extend High-Risk Classification to Pathogen Design AI Models.

• Push to amend the European Union Artificial Intelligence Act [65], and similar national legislation, to classify AI models trained to generate, evolve, or optimize viral genomes as high-risk systems. Legal definitions of biological misuse can be based on recent guidelines of the US AI Safety Institute [66].
• Regulatory requirements for AI in designing novel organisms for research:

  ○

  Documentation of training data provenance.

  ○

  Disclosure of filtering or safety layers used to restrict misuse.

  ○

  Risk assessments by independent technical reviewers prior to deployment.

  ○

  In line with operative paragraph 9 of UN resolution A/RES/78/265 [67], the private sector should be encouraged to develop safe AI systems, for example, widely accessible models with inherent safeguards against biological design. Such considerations have been outlined by the Nuclear Threat Institute [68] and are already considered for xAI and OpenAI [69].

5. Strategies for Epidemiological Monitoring and Response

Viruses are the most numerous biological entities on the planet and have continuously demonstrated the ability to mutate and evolve with their hosts or to infect new ones, sometimes with devastating consequences. The natural risks that viruses pose are only expected to increase with climate change, driving the spread of animal vectors and human encroachment into natural habitats. As thousands of species cross-over events are predicted in the coming decades [70], there is great potential for new viruses to establish epidemics in addition to new strains of traditional pathogens of concern. Furthermore, the ability to design viruses in silico or alter them via AI-assisted optimization presents a growing challenge to public health infrastructure. While current biosurveillance focuses on natural evolution, future pandemics may involve viruses with engineered features—whether from vaccine research, industrial synthesis errors, or misuse.

Artificial intelligence is already being deployed to meet these challenges. AI models now assist in the real-time monitoring of health networks, the modeling of disease outbreaks, and the detection of early warning signals from hospital records, wastewater analysis, and genomic data. Early warning programs such as BlueDot were able to predict the potential spread of SARS-CoV-2 by monitoring medical mailing listservs [71], while wastewater surveillance tools in the state of Texas (US) could detect measles virus before cases in the recent outbreak were reported [72]. These capabilities demonstrate our ability to act quickly and strategically against biological threats if response frameworks are developed.

But AI’s utility also introduces a new dilemma. The same techniques used to model natural viral evolution and transmission can be directed toward enhancing virulence, escaping immune detection, or targeting tissues. A growing number of open-access tools now allow for the design or modification of viral genomes in silico. Despite these advances, our systems for detecting and attributing synthetic or engineered viruses remain fragmented. No global standards exist for identifying artificial genomic features or for distinguishing engineered viruses from naturally occurring ones with high confidence. Outbreak attribution remains largely reliant on epidemiological methods and contact tracing, not molecular forensics. If we are to prepare for a future in which AI and synthetic biology become routine components of life science, we must also develop robust, internationally coordinated safeguards to identify, assess, and respond to novel biological threats regardless of their origin. Such tools would fit into the goals of the WHO’s R&D Blueprint and Global Outbreak Alert and Response Network (GOARN). Particular policy recommendations are:

1. 

Establish an Integrated AI-Assisted Surveillance Network.

To effectively mitigate emerging viral threats, global health institutions must establish an interoperable, AI-augmented biosurveillance system. Such a network would integrate diverse data sources, different regional formats and languages, and predictive models to identify, localize, and contextualize outbreak signals in real time. The key features include:

• Health System Monitoring: AI tools from the firms BlueDot and Metabiota already scan global health services, such as ProMED-mail [73], news, and social media, to flag outbreaks [74].
• Environmental Surveillance: Existing wastewater and metagenomic sampling programs can upload their data to centralized servers. These can then be parsed with machine learning models to detect deviations from baseline viral loads. Several tools already exist for identifying viruses from such datasets, where most of the isolated material is likely nonviral. These include ViraMiner [75], DeepVirFinder [76], LucaProt [77], and EdeepVPP [78]. Many tools can predict completely new viruses with no known reference sequence [79]. Companies such as Ginko Biosecurity have airport wastewater monitoring programs that can directly flag potential spread of viruses.
• Zoonotic Risk Modeling: AI-assisted mapping of animal reservoirs and vector habitats, alongside climate and land-use data, can forecast crossover risk zones. Major groups such as the U.S. Department of Agriculture, the World Organisation for Animal Health, and the Food and Agriculture Organization of the United Nations Emergency Prevention System conduct surveillance of zoonotic diseases in livestock, wildlife, and migratory species, often in partnership with regional and national veterinary services. Linking their datasets with AI platforms could enable earlier detection and support predictive modeling of emerging zoonotic threats as part of an integrated global biosurveillance network. AI tools have even been developed to classify mosquito vectors from images of wings, simplifying fieldwork [80].

2. 

Automate Pipelines for Identifying Viruses and Variants.

The methods of collecting viral sequences vary considerably, and each comes with its own advantages and limitations. Targeted amplification by PCR can be done rapidly and at low cost but may miss minor variants within a sample or fail outright if the target is significantly different from what is expected. Even after a viral genomic sequence has been obtained, it takes specialized software and database tools to cluster it into a viral family tree to trace origins or determine if new mutants have more pathogenic potential. Being able to determine the origins of a virus can help in containment, for instance, by identifying animal reservoirs or hallmarks of synthetic design. Several tools could provide the basis for an integrated analysis network that has the following major objectives:

• Analyze and Classify Viral Genomes.

Pipelines must be able to integrate large sets and various data formats from clinical and environmental surveillance. They should automate classification and determine if viral loads are above baseline values. Tools as starting points for consideration are:

• Freeware such as GARSA [81], GARD [82], and UShER [83] for real-time creation of viral family trees and detection of large recombinations;
• PyR0 to identify relative geographic prevalence and fitness of viral strains [84];
• NetMHCpan [85] to identify if new viruses are less visible to the immune system;
• BioLaboro [86] to automate PCR-based diagnostics against flagged sequences.

b.

Detect Patterns of Synthetic Alteration Models.

There was much debate during the SARS-CoV-2 pandemic about the lab origin of the virus, even among scientists. Such discussions served to foster a growing distrust of research institutions and diverted analytical resources away from other avenues. However, identifying a synthetic virus would be important in containment responses and tracing origins. While there are no widely accepted tools to flag synthetic alterations, development could be based on comparing new isolates to natural, closely related viruses for the following patterns:

• Restriction sites flanking key genes of interest that indicate genome editing.
• A change in the global nucleotide patterns. This includes variation in GC-richness, codon usage and codon-pair biases that may optimize a virus for a particular host, and dinucleotide patterns, particularly CpG and ApU.
• New functional elements and genomic configurations. This can include gene expression factors such as promoters, internal ribosomal entry sites, peptide cleavage or frameshifting signals from unrelated viruses, or genes arising in different configurations or overlapping with prototypical sequences.
• Mutational hotspots in one region of the genome and not in others, particularly those that change the corresponding amino acids.
• A panel of experts should be formed to manually validate sequences flagged by a validated autonomous system as synthetically altered, and output formats should be graphical and easily understood.

c.

Proactively Identify Variants of Concern.

• New sequences can be checked against large biological databases of protein and small molecule interactions to infer potential drug targets or gain-of-function mutations. Such databases include STRING [87] and BioGRID [88].
• Protein folding tools, such as AlphaMissense [89], DDMut [90], FoldX [91], or ESM-2 [92], combined with automated protein–protein interaction predictions, such as HADDOCK [93] or ClusPro [94], could theoretically be used to determine if new viral strains bind better to host factors or calculate what mutations would improve viral fitness before they are seen in the clinic.

  ○

  These tools can also predict mutations in viral genes that preserve biological function but bypass diagnostics from PCR or antibody tests.

  ○

  Generated structures could be tested for resistance to antiviral drugs [95].

3. 

Model-Driven Outbreak Control for Predicting Transit Hubs and Spread.

In parallel with detection and attribution, early containment depends on understanding how and where viruses may spread. AI models using human mobility data can identify transit hubs that may serve as amplifiers of infection and prioritize containment measures geographically. Simulations and risk maps can incorporate airline routes, commuter data, and anonymized telecom signals to project real-time virus trajectories, particularly in cases of panic where people may be fleeing and spreading a localized outbreak. This modeling can guide both national responses and global coordination.

• Establish tools to prioritize transit hubs. In the event of an outbreak, total shutdowns are economically and politically unfavorable. Certain routes could still facilitate trade without spreading disease, or their closure could be ineffective in controlling diseases with animal vectors. Combined with biological data such as transmission routes (e.g., droplet or by mosquitoes) and time between infection and onset of symptoms, several tools can be used to predict the spread and, therefore, which transit hubs should be monitored or restricted or which destination hubs should be on alert. In addition to the commercial firms mentioned in part 1 of this section, computational tools include GLEaMviz [96] and EPIRISK [97], based on major transportation hubs, EpiModel for customizable factors [98], or tools using cellular networks to map mobility [99,100].
• Support the development of national response plans. Local agencies should develop action plans in the event of an outbreak. These should account for the transmission modes of the virus and testing protocols. A UN or WHO-backed bulletin outlining effective strategy goals could normalize the efficacy of these plans.

6. Conclusions

Artificial intelligence is a double-edged tool in virology: it holds enormous potential to improve diagnostics, outbreak preparedness, and therapeutic discovery, but it also presents significant risks through its capacity to design, optimize, or obscure pathogenic viruses. As these capabilities continue to diffuse across sectors and borders, traditional governance models based on static protocols and manual, voluntary review are no longer adequate. AI has lowered the technical barrier to both helpful and harmful applications of synthetic biology [101], and governance must evolve accordingly (Table 1).

To confront this dual-use dilemma, the UN and its partners should lead the development of a flexible, tiered governance framework that balances scientific freedom with security. This includes integrating AI-attribution tools into WHO and UN outbreak protocols, standardizing the preclinical validation of AI-derived diagnostics and therapeutics, and establishing international incentives for responsible DNA synthesis oversight. These measures can guide innovation while mitigating emerging biosecurity threats (Figure 2).

Several challenges remain. Current incentives for biotechnology firms and synthetic biology providers are misaligned, as voluntary compliance with screening protocols often adds cost without tangible benefit. For academic researchers, mandatory validation requirements could slow the pace of beneficial innovation. A drug discovery program delayed by biosecurity screening may miss critical windows for pandemic response, and diagnostic tools held in regulatory review may fail to reach clinics in time to alter outbreak trajectories. Implementation capacity is uneven across countries, particularly in low- and middle-income regions where much of the viral spillover risk resides. Stringent screening requirements and certification mandates may disproportionately burden smaller biotechnology firms and under-resourced academic institutions in these regions, potentially widening rather than narrowing the global research capacity gap. A framework optimized in a wealthy location but unaffordable in a developing region becomes a mechanism for exclusion rather than protection. Moreover, if governance structures are perceived as instruments of Western regulatory control rather than genuinely multilateral partnerships, they risk provoking resistance from the very nations whose participation is most critical to global biosecurity.

Global coordination mechanisms to harmonize AI and biosecurity standards are fragmented and under-resourced. Even well-designed frameworks fail without sustained political commitment and adequate financing—outcomes that depend on factors beyond the framework’s technical merits and can change rapidly. The regulatory challenge is therefore not merely to design optimal policies on paper but to calibrate interventions that are risk-proportionate, economically feasible, and politically achievable across vastly different national contexts. Overly stringent governance may prove counterproductive if it drives dual-use research into less transparent settings or discourages beneficial applications that outweigh their misuse potential.

To address these gaps, a coalition of stakeholders, including the WHO’s Global Initiative on AI for Health and Center on AI for Health Governance, FAO, UNEP, and regional centers, like Africa CDC, should be empowered to develop and manage an interoperable biosurveillance network supported by automated data pipelines and open-access AI tools. The UN Technology Facilitation Mechanism and the Pandemic Fund could serve as institutional platforms for aligning technical capacity with financing. Aid agencies such as USAID, the Global Fund, and Gavi can contribute to infrastructure and workforce development for AI-driven health security in vulnerable regions. Meanwhile, neutral certification bodies under UN or CEPI leadership could incentivize private sector compliance through procurement preferences, access to global screening datasets, and multilateral recognition programs.

The descriptions of tools in this manuscript are not specific endorsements but rather a demonstration that the entry of AI into virology has already occurred. Indeed, AI has already been utilized to create novel viable viruses against bacteria [102]. Ultimately, these tools must be evaluated and deployed to advance global health rather than destabilize it. This requires proactive investment in AI governance not just to regulate misuse, but to actively shape the next generation of computational biology. The tools already exist and continue to grow in number by the year. What is needed now is a collective commitment to use them wisely before the next biological crisis.

7. Future Directions and Policy Limitations

Several limitations constrain this policy analysis. First, this framework remains conceptual pending empirical validation; pilot studies in select jurisdictions are needed to assess implementation feasibility and cost-effectiveness. Second, our analysis relies on the published literature and policy documents current to 2025. As AI capabilities continue to evolve rapidly, regulatory frameworks may require continuous updating to maintain relevance. Third, the economic burden of manual expert validation introduces practical constraints. While expert panels incur costs related to financial overhead, time delays, and potential human error, the catastrophic risk of allowing highly pathogenic engineered sequences to bypass detection far exceeds these expenses. A tiered urgency-based triage system prioritizing review based on epidemiological urgency, submitter credentials, and metadata fidelity can optimize resources by reserving intensive scrutiny for high-risk cases while enabling verified research to proceed. Such systems could integrate into existing infrastructure, like the WHO BioHub Initiative, to distribute costs internationally [103]. Finally, political will and international coordination challenges remain substantial, particularly regarding enforcement across non-signatory nations and potential resistance from major AI powers.

The transition from theoretical frameworks to operational implementation requires addressing several critical gaps. First, bridging the knowledge divide between machine learning engineers and domain virologists demands integrated review structures where AI predictions are grounded in biological synthesizability and experimental validation. Industrial trials have demonstrated AI’s transformative potential: Moderna utilized computational biology to design its mRNA-1273 vaccine less than a day after the SARS-CoV-2 genome was released, with the first clinical trial 66 days after the genome was available [104], while BenevolentAI employed knowledge graph analysis to identify baricitinib as a COVID-19 treatment within weeks of the first known cases [105]. However, ensuring AI-based predictions transcend “black-box” outputs requires standardized documentation of model limitations, interdisciplinary training programs pairing virologists with AI developers, and open platforms linking predictions with synthesizability databases.

Second, addressing the “free-rider” problem among non-signatory nations remains paramount. Historical precedents reveal recurring failures: the Biological Weapons Convention lacks verification mechanisms due to the 2001 protocol collapse over proprietary and security concerns [106]. Learning from these failures, our framework incorporates supply chain controls for DNA synthesis equipment and cloud infrastructure, combined with economic incentives such as “Certified Responsible Provider” designations, liability protections, and fast-track regulatory approvals that make biosafety compliance a competitive market asset. The WHO’s ongoing implementation of its 2022 Global Guidance Framework, including 2023–2024 pilots in regions like Uganda [107], provides valuable real-world testing of these approaches. Future research should evaluate these pilots’ effectiveness, develop automated risk assessment tools for AI-generated sequences, conduct cost–benefit analyses of various governance structures, and explore incentive mechanisms that align commercial interests with biosecurity objectives. Third, the framework’s design consciously draws on lessons from dual-use governance in other domains. Nuclear technology regulation through the International Atomic Energy Agency (IAEA) demonstrates that technical verification mechanisms combined with benefit-sharing programs (civilian nuclear power access) can achieve sustained international compliance even among geopolitical rivals. Key features transferable to AI biosecurity include: (1) mandatory reporting of dual-use materials and capabilities; (2) inspection protocols based on technical standards rather than political negotiation; and (3) graduated sanctions that preserve pathways for return to compliance. Our framework incorporates analogous mechanisms where automated DNA synthesis screening logs function as technical verification, while economic incentives (certification benefits, fast-track approvals) create tangible rewards for participation. Civilian engagement can be encouraged, and overall system effectiveness can be measured through competitions similar to hackathons, where groups strengthen model security through red teaming to expose flaws and vulnerabilities.

Conversely, cybersecurity governance illustrates pitfalls to avoid. The Budapest Convention on Cybercrime, despite being the primary international treaty addressing computer crime, has struggled with adoption beyond Europe and North America. Major cyber powers, including China, Russia, and India, remain non-signatories, citing sovereignty concerns and inadequate involvement in treaty drafting. This fragmentation has limited the Convention’s effectiveness in addressing transnational threats. The lesson for AI biosecurity is clear: frameworks developed primarily by Western nations risk similar legitimacy deficits. Our proposed UN-led approach deliberately prioritizes inclusive multilateral development, recognizing that buy-in from major AI and biotechnology powers (US, China, EU, India) is a prerequisite for effectiveness. The modular design, allowing partial adoption without requiring full treaty ratification, responds directly to the Budapest Convention’s rigidity problem, as nations can implement DNA synthesis screening or diagnostic validation standards without committing to the entire framework, creating multiple entry points rather than an all-or-nothing adoption barrier.