Among the 169 peer-reviewed papers in the sample, there were 81 papers that explicitly discussed the impact of AI-enabled synthetic biology. The other 88 papers discussed risk but not explicitly from AI. Equally surprising was the exclusion of AI in all the other papers, given that several were review articles or otherwise covered state-of-the art or emerging technologies and tools for synthetic biology. There is no ready explanation for this omission, except to say that perhaps (a) those researchers are not familiar with the potential of (generative) AI for biology, (b) do not think it is as big of a deal as others do, or (3) consider it less relevant for today’s concerns in synthetic biology, or (4) consider AI (machine learning) an essential tool but prefer not to elevate it beyond its obvious place as a key research tool.

Another finding is that many papers that I found relevant to the future of synthetic biology’s governance, risk, and innovation trajectory, did not in fact use that term. Dozens and dozens of papers included in the sample happily discussed AI and the impact on their field, be it metagenomics of the microbiome (Wani et al., 2022), health and intelligent medicine (Achim and Zhang, 2022), applied microbiology (Xu et al., 2022), designer genes (Hoffmann, 2023), drug discovery (Yu et al., 2022), oncology (Wu et al., 2022), systems biology (Helmy et al., 2020) without realizing that synthetic biology is bound to intersect with it at some point soon (or at least explicitly omitting the use of the term). What could the reason be? Is the term upsetting to part of the biology community or establishment? Synbio scholars clearly frame their problems differently from the biology establishment. Perhaps there is a disparity in age, experience, and skills between patchy biological knowledge of bio-IT nerds and lacking IT skills among biologists? AI has been applied to material discovery, finding 700+ new materials so far (Merchant et al., 2023) and it is a question of time before it will be used for scalable materials design using AI-enabled synthetic biology (Tang et al., 2020; Burgos-Morales et al., 2021) although for real world applications we might first need better standardized vocabularies for biocompatibility (Mateu-Sanz et al., 2023).

Only 5 papers (Kather et al., 2022; Grinbaum and Adomaitis, 2023; Morris, 2023; Ray, 2023; Xiao et al., 2023), a popular science article (Tarasava, 2023), and an editorial (Generating ‘smarter’ biotechnology, 2023) discussed the impact of generative AI on synthetic biology. This is expected to increase dramatically quite soon, given the success of this latest wave of AI technology and the platform aspects of its spread. However, as one paper put it, synthetic biology has a natural synergy with deep learning (Beardall et al., 2022). The use cases discussed in various papers include: as a classifying text, generic search engine, generating ideas, helping with access to scientific knowledge, coding, patient care (Clusmann et al., 2023), protein folding, proofreading, sequence analysis, summarizing knowledge, text mining of biomedical data, translation (Clusmann et al., 2023), workflow optimization, foresight of future research directions (Yan et al., 2023); collection of related synthetic biology data, and more (Beardall et al., 2022; Clusmann et al., 2023; Tarasava, 2023). The promise of AI-enabled cell-free synbio systems (Lee and Kim, 2023), which use molecular machinery extracted from cells, is particularly significant for automation and scale-up of biosensors among other things. It bears pointing out that the significant advances in cell-free synbio systems enabling the acceleration of biotechnology development, specifically its ability to enable rapid prototyping as well as the ability to conduct predictive modeling, pre-date generative AI by a decade (Moore et al., 2018; Müller et al., 2020). Even today, he the barriers seem to be limited availability of relevant data either because it does not exist yet, because data is scarce, the data set is small, because it is not publicly available, or because it is not formatted in useful ways (Rosenbush, 2023). Not all of these challenges can be immediately resolved by generative AI.

What matters most to the governance and innovation concern would be those barriers, areas, or workflows where AI could make the biggest impact, not just for the field applying it but for the shared resource that is AI-enabled synbio that would grow the pie. Based on the literature review, I’ve attempted to suggest which topics fit in that perspective (see Figure 4). For example, progress on interoperability would benefit all, as would AI-enabled lab operations workflows. Each would be a synbio building block. Big ticket items such as protein folding is in a bit of a different category. When AlphaFold achieved near-perfect protein fold predictions, it was the most important moment for AI in science so far, yet left plenty of work for structural biology (Perrakis and Sixma, 2021), including the application of coiled coils as a self-assembly building block in synthetic biology (Woolfson, 2023). Similarly, when the mRNA platform became a successful vehicle for a COVID-19 vaccine that changed the world, this happened in vitro, yet, the production of synthetic mRNA (Hınçer et al., 2023) or miRNA (Matsuyama and Suzuki, 2019) in the cell itself would be an even more significant breakthrough–and getting there might require the use of AI (Naderi Yeganeh et al., 2023).

On the other hand, there is no reason to believe that the top labs will be overtaken, quite the contrary, in fields such as consulting, it seems that generative AI accelerates the work of top teams (Moran, 2023). However, the worrying aspect is that the previous assumption from biorisk work pre-generative AI was that developing pathogens is an activity only possible in highly advanced biolabs. The increasing availability of insight, instructions, as well as wetlabs and foundries on demand (Sandberg and Nelson, 2020), would seem to be a potential issue to watch.

In the literature, there is ample evidence of what constitutes biosafety and biosafety governance best practice (Perkins et al., 2019; Wang and Zhang, 2019; Li et al., 2021; Mökander et al., 2022; Sandbrink, 2023b) and the emphasis is on a mix of specific training and, relatedly, developing a safety and responsibility work culture. In previous decades, the few advanced labs that existed were “compliant” biocontainment actors, for which acceptable systems were in place. However, regulating synthetic DNA comes with new challenges, including scalability, and less ability to create genetic firewalls to natural organisms, and it has become easier to circumvent oversight (Hoffmann et al., 2023). The more accessible (Wang and Zhang, 2019) and generally useful synbio potentially is regarded to be (Sun et al., 2022), the less likely it is that prohibition will remain an effective tool. Decades-old bioinformatics resources originally developed to compare gene sequences, such as BLAST, have been re-used, with mixed results, as biosafety tools to identify pathogens (Beal et al., 2023). Newer tools, such as machine learning-based topic models, enable spotting trends across a wide set of biosafety research publications (Guan et al., 2022). AI-synbio governance (Achim and Zhang, 2022; Mökander et al., 2022; Grinbaum and Adomaitis, 2023; Holland et al., 2024) is expected to be more of the above, but also requires AI skills and perspectives that go far beyond wet lab practices and will require updates to biosafety laws, regulation, governance, standardization (Pei et al., 2022). It will change the role of the state (Djeffal et al., 2022) as it will no longer be the primary norm setter or enforcer of responsibility.

AI is already contributing to the fragmentation of biology (Hassoun et al., 2022) and will challenge medical expertise among specialists (Patel et al., 2009). Generative AI, and especially other advancements in multi-modal AI, combined with better multi-omics synbio dataset interoperability (Topol, 2019) and standardization, will eventually lead to fundamentally new playing fields. Vigilance is required (Harrer, 2023) both to get us there, predict when we will get there, and decide what to do when we get there. The initial issue surrounds AI-synbio lab safety practices (D’Alessandro et al., 2023) when the “lab” suddenly is a dispersed concept, and decisions around forbidden knowledge (Hagendorff, 2021), new sets of responsibilities in the research community (Blok and von Schomberg, 2023) and among health practitioners (Achim and Zhang, 2022), avoidance of doom speak (Bray, 2023), handling the reality of malicious actors (Carter et al., 2023), and will represent an enormous challenge for reskilling and upskilling those who want to work with the topic (Xu et al., 2022).

Getting it right will mean balancing brave investments (Hodgson et al., 2022) with monitoring the effects, including developing an ethics and a taxonomy for working with AI-synbio-human hybrids and intelligence (Nesbeth et al., 2016; Damiano and Stano, 2023), dealing with new synthetic pathogens (O’Brien and Nelson, 2020), saying carefully goodbye to the natural world (Lawrence, 2019; Webster-Wood et al., 2022; Bongard and Levin, 2023) or at least radically enhancing biocontainment (Schmidt and de Lorenzo, 2016; Aparicio, 2021; Vidiella and Solé, 2022; Hoffmann, 2023), as well as developing new approaches to worker safety (Murashov et al., 2020). This leads into the issue of dual use of concern, which currently is a binary issue even though it is about to become immensely complex, requiring a more nuanced approach (Evans, 2022; Sandbrink, 2023a), given the legitimate concern with deliberate, perhaps even deliberate synthetic pandemics (Sandbrink, 2023b).

Dual use is mentioned by several papers (Getz and Dellaire, 2018; Torres, 2018; DiEuliis et al., 2019; Alexander Hamilton et al., 2021; Hagendorff, 2021; Grinbaum and Adomaitis, 2023; Vaseashta, 2023). However, it can have broader meaning, for example, positively referring to open source (Esquivel-Sada, 2022) as opposed to negatively referring to non-conformant use. The idea of broadening the dual use research of concern (DURC) regime, which gained steam during COVID-19 (Sandbrink et al., 2023) is mentioned in the preprint literature (Grinbaum and Adomaitis, 2023), and alarms about dual-use involving AI are sounded in several papers (Urbina et al., 2022; D’Alessandro et al., 2023) and in Sandbrink’s Ph. D thesis (Sandbrink, 2023b).

Giving a complete regulatory overview of synthetic biology is complex (Beeckman and Rüdelsheim, 2020) and is not the task of this paper, but Table 1 still lists some key standards and guidelines discussed in the sample, and relevant to DURC issues.

Synbio is not yet a mature engineering industry with well-understood costs and timelines (Watson, 2023) and investments fluctuate from year to year (SynBioBeta, 2023). A recent Schmidt futures report defines commercial production scale as a fermentation capacity of 100,000 L or more and states only a few U.S. companies currently have infrastructure at this scale and relatively inaccessible to small- and medium enterprises at the present moment (Hodgson et al., 2022). Achieving pilot scale is the first hurdle to pass and would cost in excess of $1 billion to build a dozen pilot facilities to fuel the U.S infrastructure alone (Hodgson et al., 2022). Synbio was not truly part of the industry 4.0 paradigm either (Jan et al., 2023). The keywords to describe the industrial aspect of synthetic biology included: “bioeconomy”, “bio-capitalism”, “biomanufacturing”, “biotech industry”. Surprisingly few path breaking peer reviewed articles were found on these topics. The six key ingredients for biomanufacturing derived from our sample are: biological insights, AI, bioprocessing, engineering scale-up, governance frameworks, and gigascale investments (see Figure 5).

Attempting to pinpoint exactly when a sci-tech paradigm will take off commercially is a fool’s errand. Exceptional growth in research communities can be tentatively forecasted from citation analysis (Klavans et al., 2020). The emergence of new industries is significantly more complex but the growth in intangible assets (Börner et al., 2018), such as generative AI, applied to an industry (manufacturing) would be a clear indicator. One article in our sample proposed a taxonomy of four innovation types specific to the bioeconomy: Substitute Products, New (bio-based) Processes, New (bio-based) Products, and New Behavior, each carries their own commercialization challenges (Bröring et al., 2020). Deriving insights from other papers, existing or emerging business models in synthetic biology would include automation, contract research, increasing crop yields in agriculture (Bhardwaj et al., 2022; Wang et al., 2022), data driven design (Freemont, 2019), efficiencies, new components, DNA synthesis (Seydel, 2023), infrastructure, licensing, manufacturing molecules for the food industry (Helmy et al., 2020), modularity, new materials, new platforms, new products, open source tools, services, or substitution, such as a new technology stack (Freemont, 2019).

That being said, despite the relatively low number of papers describing the synbio industry (van Doren et al., 2022), there are signs in the gray literature and in the consulting literature (Candelon et al., 2022) that things are changing within this decade. Arguably, the synbio startup boom in pharma and food industries will be duplicated in health and beauty, medical devices, and electronics, with cost-based competition from syn-bio alternatives in chemicals, textiles, fashion, and water industries, soon to be followed by the mining, electricity, and construction sectors (Candelon et al., 2022). The way it might happen is not necessarily only through flashy, radical innovations, but incrementally because synbio is becoming a useful tool to improve performance, quality, and sustainability of almost all types of manufacturing (Candelon et al., 2022).

Made-to-order synthetic DNA is faster and cheaper than before but is still a massive bottleneck to building scalable biological systems based on synthetic components (Seydel, 2023). The future role of synthetic biology in carbon seq uestration into biocommodities could be of major industrial importance provided the bioproduct could be commercialized (Jatain et al., 2021).

Given the nascent state of AI-enabled synthetic biology, there is an overload of related and relatable search terms and keywords that proliferate in the scientific community and online, making it difficult to compare, find, and cluster case studies, research, and policy relevant insight. Even after considerable search efforts, we were left with 1,297 unique keywords, which were boiled down to 81 broad categories, and further to 42 literature search keywords. The situation will persist, and in all likelihood, it will get worse before it gets better. There are those hoping for a taxonomic renaissance (Bik, 2017) to remedy the problem, including a taxonomy for engineered living materials (Lantada et al., 2022). An early article attempted to do the same for the field of synthetic biology (Deplazes, 2009), but it might have been too early in the cycle.

Historically, synbio has been seen as a disruptive innovation yielding products and processes which may not be well aligned with existing business models, value chains and governance systems (Banda and Huzair, 2021), but this might now be changing and synbio approaches get integrated into traditional industries and sectors. That is exciting for industrial innovation but challenging for governance, risk and regulation.

Even though commercially available synthetic biology-derived products are already on the market that are, arguably “changing the world” (Voigt, 2020), the economics of synthetic biology (Henkel and Maurer, 2007), the biomanufacturing industry overall, is in its infancy. McKinsey might be right that it is a $4 trillion gold rush waiting to happen (Cumbers, 2020), or as BCG claims, $30 trillion by the end of the decade (Candelon et al., 2022), across food and ag, consumer products and services, materials and energy production, and human health and performance (Ang, 2022; Clay, 2023). However, the conspicuous absence of management and business articles on synbio in our sample might indicate that the business dimension is so embryonic that these visions are not yet a story worthy of sustained business school attention. The umbrella term bioeconomy (Baker, 2017; Bröring et al., 2020; Marvik and Philp, 2020; Banda and Huzair, 2021; Hodgson et al., 2022; Bröring and Thybussek, 2023; Clay, 2023; Rennings et al., 2023) is perhaps convenient, but encompasses so much that it is hard to know what it means.

What seems to be missing in the literature is a clear vision for how AI-enabled synthetic biology would be truly different from previous approaches. Most of the papers imply that AI will remain only one of many technologies relevant to progress in the synthetic biology field. No papers paint a picture where there is a straightforward path to massive scale-up, with possible exception of AI for multi-omics. The shift would happen once the synbio field was able to shift from its current systems-centric approaches (requiring slow, cumbersome wet lab experiments and trial-and-error tinkering) to data-centric bioprocessing approaches (not just using AI for data processing) that are themselves digitally scalable (Owczarek, 2021; Scheper et al., 2021), and constitute automated design-build-test systems (Holland et al., 2024), supported by digital twins (Manzano and Whitford, 2023). Having said that, enormous efficiencies could be had through even much simpler process automation and operations improvements in biomanufacturing, for example, through no-code methods (Linder and Undheim, 2022).

The barriers to the field of synthetic biology’s growth are many, from (1) technical feasibility, including the scientific problems connected with the fusion of three disciplines; synthetic biology, artificial intelligence, and social science (Trump et al., 2019), via (2) various forms of risk to (3) industrial challenges, to (4) institutional challenges, or (5) social dynamics.

On the technical side, we find the challenges surrounding data quality (Patel et al., 2009) the fragmentation of knowledge (Hassoun et al., 2022) interoperability (Mateu-Sanz et al., 2023) or standardization (Endy, 2005; Hanczyc, 2020; Garner, 2021; Pei et al., 2022; Mateu-Sanz et al., 2023). For example, even though there is great need, and the desire is there, standardizing complex biological systems is difficult (Garner, 2021). As many of the papers in the sample point out, there are also scientific problems connected with the fusion of two disciplines, synthetic biology and artificial intelligence. A multi-layer technology stack is evolving (Freemont, 2019). There is transdisciplinary training and perspective required (Hammang, 2023). There is also considerable uncertainty produced when three domains (or more), and their methodologies, technologies, and tools, are merging (Trump et al., 2019).

Notably, (1) biological insight is needed to deploy AI correctly, yet cell behavior is unpredictable (Lawson et al., 2021) (2) AI insight is needed to capture what ends up retranslated as biological patterns in the data, but AI insight alone is not sufficient to identify what data might be relevant and (3) social science insight, including business models, sociotechnical issues (Marris and Calvert, 2020), social dynamics, social implications, governance, risk, ethics, and psychological reactions, is needed to assess the feasibility of R&D, product development, and commercialization of the emergent field’s output. That’s a tall order for individual researchers, teams, companies, and political institutions alike. As the field grows in importance, scale, and impact, it will entail an enormous societal reskilling effort (Hammang, 2023).

Industrial challenges include: biosafety (Pei et al., 2022), the availability of capital (Baker, 2017; Helmy et al., 2020; Hodgson et al., 2022; Sargent et al., 2022), the lack of a scalable manufacturing workflow (Hillson et al., 2019; Ataii et al., 2023), regulatory uncertainty (Huzair, 2021), innovation challenges (Banda and Huzair, 2021; Tait and Wield, 2021), intellectual property rights (Esquivel-Sada, 2022), investment risks (Hodgson et al., 2022), or worker safety (Murashov et al., 2020).

Various forms of risk will impact synbio growth, notably AI risk (O’Brien and Nelson, 2020; Grinbaum and Adomaitis, 2023), the potential for a slew of catastrophic risks (DiEuliis et al., 2019) such as new pathogens, or even the specter of existential risks (Boyd and Wilson, 2020) threatening the flourishing or survival of humanity.

On the institutional side, we cannot discount biosecurity (Millett et al., 2020), bioterrorism (Trump et al., 2020; Sheahan and Wieden, 2021; Bray, 2023), the constant challenge of existing, emerging, or evolving bioweapons (Gronvall, 2018; DiEuliis et al., 2019; Trump et al., 2020), the cost of deregulation (Sargent et al., 2022), dual use (Getz and Dellaire, 2018; Ienca and Vayena, 2018; Evans, 2022; Grinbaum and Adomaitis, 2023), global governance (Linkov et al., 2018a; Dixon et al., 2022), security (Palmer et al., 2015) startup dependency (Freemont, 2019; Nylund et al., 2022) in terms of achieving a steady stream of new innovation in the domain.

Social dynamics such as differing notions and rationales surrounding bioethics (Szocik et al., 2021; Bohua et al., 2023), social acceptance (Bauer and Bogner, 2020; Frow, 2020), also play a part.

(Endy, 2005; Patel et al., 2009; Palmer et al., 2015; Bauer and Bogner, 2020; Boyd and Wilson, 2020; Frow, 2020; Hanczyc, 2020; Murashov et al., 2020; Garner, 2021; Pei et al., 2022; Sargent et al., 2022; Mateu-Sanz et al., 2023).

The most reasonable way to look at it would be: what can generative AI do within the frame of all of these barriers? From this we can wonder whether generative AI-enabled synthetic biology really would be truly different from previous approaches. We could, of course, also wonder how different the situation would be if many of those previously mentioned barriers somehow went away. Interestingly, what AI fanatics would respond is that once those barriers are gone, AIs would themselves produce such approaches that are far superior to what could be conceived by human experts. Alternatively, it is always possible that emerging, superior and multi-modal AI systems would be able to overcome enough barriers to transform the field anyway.

For now, the most prudent governance path seems to be to keep fostering a responsibility mindset in a distributed manner at global scale. Machine learning enabled digital processing is already improving diagnostic accuracy and reducing turnaround time for even complex lab tests (Undru et al., 2022). The smart laboratory, with AI-automation of biosecurity, has arguably moved from concept to reality, enabling self-control process management flows of personnel, materials, water, and air, automated operation, automated risk identification and alarms (Li et al., 2022). However, when technologies merge, uncertainties multiply (O’Brien and Nelson, 2020), cybervulnerabilities and circumvention options increase (O’Brien and Nelson, 2020).

Take the case of biosafety labs, state-of-the-art labs designed not only to protect researchers from contamination, but also to prevent microorganisms from entering the environment. After concerns about health risks from recombinant DNA technologies, scientists met in 1975 to create safety guidelines according to the risk involved. Since its initial release in 1984, Biosafety in Microbiological and Biomedical Laboratories (BMBL), a 574-page document created by 200 contributors, has served as the cornerstone of biosafety practice in the United States. BMBL, currently in its 6th edition (Hatcher et al., 2023), is issued by the Centers for Disease Control and Prevention (CDC) and created to oversee the biosafety of recombinant DNA research at all institutions receiving funding from the National Institutes of Health (NIH). Since 1983, the WHO has its own Laboratory Biosafety Manual, a 128-page document currently in its 4th edition (LBM4), which adopts an evidence-based risk approach to biosafety rather than a prescriptive approach (WHO, 2020). The WHO also has a discussion forum, the Technical Advisory Group on Biosafety (TAG-B) (Hstoday, 2022). Despite obvious cyber risks, the prevalence of publishing potential dual use research on preprint servers, as well as the fact that there might be AI-enabled pathways to pathogen gain-of-function (Urbina et al., 2022), as well as other unintended consequences of increased automation (Kulken, 2023), implied in a digitalization of core aspects of even labs mostly working on low pathogenic agents (Lentzos et al., 2022), neither BMBL nor LBM, mention AI precautions. AI-synbio integration guardrails, including technical AI safeguards, and strengthening DNA synthesis screening, even limiting the access to AI biodesign tools, might be needed (Sandbrink, 2023a; Carter et al., 2023). Even if one achieves consensus on such measures on a multilateral (or national) basis, the implementation will be complex.

International biosafety guidance and regulation is highly variable, and key factors such as awareness, safety culture (Perkins et al., 2019), capability, practice, understanding, error documentation, training, and enforcement are particularly underdeveloped across Asia (Johnson and Casagrande, 2016). In 2020, as a potentially important step forward, the International Organization for Standardization (IS) issued an international standard for biorisk management, ISO 35001:2019 (ISO, 2020), but whether and how it is being used is too early to tell (Callihan et al., 2021). In the absence of a sufficient set of internationally recognized biosafety standards and norms (Gronvall and Rozo, 2015; Silver, 2022; Hadshar, 2023), BMBL also currently sets the international de facto standard. Having said that, it is primarily an advisory document. Now, consider the economic incentives, including tax incentives, combined with the health imperatives to build lab capacity across the world in a post-pandemic world (KMPG, 2020; Field, 2023). According to the Global BioLabs Initiative (www.globalbiolabs.org/), jointly published by King’s College London, Bulletin of the Atomic Scientists, and George Mason University, the global boom in construction of BSL4 and BSL3+ labs raises biosafety and biosecurity concerns, particularly given that it is occurring in places where biorisk management oversight is weak (GBR, 2023).

If it indeed was the case that AI lifts all boats, it wouldt mean that mediocre labs can more rapidly gain the ambition to modify their facilities and work practices, and start doing work regulated by BSL-3 and BSL-4 designations. But while more lab researchers than before might potentially deploy AI to carry out experiments that should be carried out in a lab with a stricter designation (a higher BSL), this would, in most cases, be against the regulations. Having said that, in India, for example, there are no national reference standards, guidelines, or accreditation agencies for biosafety labs, so those labs that do comply, rely on the international ones (Mourya et al., 2014; 2017). China, also, lacks a comprehensive regulatory system for BSL-2 labs, and lacks trained biosafety staff (Wu, 2019). The numbers game is indeed instructional. The International Laboratory Accreditation Cooperation (ILAC) accredits over 88,000 laboratories (ILAC, 2023). If you consider that there are currently 64 BSL-4 labs, just imagine if all BSL-3 labs wanted to do BSL-4-type work, and were capable of it. There are already some 57 BSL-3+ labs (Kaiser, 2023), the safety levels of which are poorly defined (GBR, 2023). There are currently great experiments going on regarding the feasibility of rapid response mobile BSL-2 lab deployment to areas with a public health crisis, but those labs carry additional risks from rogue elements (Qasmi et al., 2023). Or, what about if all BSL-2 labs (which include most labs that work with agents associated with human diseases) suddenly started doing BSL-3 or BSL-4 type work? With AI, and without national control regimes, more BSL-2 labs might be tempted to think they can take on more advanced work, too quickly. Regardless of intent, there is also the risk that a lab leak could be a digital leak, facilitated by AI. Either way, beyond adopting a needed multilateral approach (GBR, 2023), the next best might be to increase scientists’ motivation (which might be even more important than raising skill levels) to practice good biorisk management (Greene et al., 2023) and adopting a culture of safety (Perkins et al., 2019).

As has been pointed out, there is a need to deploy a governance continuum (Hamlyn, 2022). Based on our reworking of the issues based on the literature review (Linkov et al., 2018b), there are six governance levels (global, national, corporate, lab, scientist, citizens) and four governance types (precautionary, stewardship, bottom-up, and laissez-faire) to be considered in an emerging Framework for AI-enabled synbio governance (see Table 2). Each of these need constant monitoring and renewal based on assessing threats, hazards, and opportunities. Each governance level might prioritize one approach, but must have aspects of all governance types. Each governance type must be reflected at all governance levels. Today, we only have elements of such a framework implemented, and the skills required to make a comprehensive approach happen are formidable and require an all-of-society effort.

At any level, the process is quite complex. For example, the corporate AI governance at British biopharma AstraZeneca includes compliance documents, a responsible AI playbook and consultancy service, an AI resolution board, and AI audits–emphasizing procedural regularity and transparency–and interlinking with existing procedures, structures, tools, and methods (Mökander et al., 2022).

However, the literature review points to the fact that the true governance challenge is not only about the individual elements doing things right. Rather, proper governance is interactive, and adaptive, and requires working on all levels of governance simultaneously while not ignoring any one level for much time at all. One could describe the process as whack-a-mole governance (see Figure 6) where there are many actors using small rubber mallets (aka laws, rules, norms, votes) that need to hit each level simultaneously for the button (risk) to stay down.

The above schematic must be complemented by transparent approaches for each set of tools (real-time monitoring, certification, compliance documentation, standardization, prizes, rewards). The fact that some types may adversely affect others, for example, soft law might delay or undermine regulation or hard law, must be monitored and dealt with through responsible innovation (RI) approaches (Hamlyn, 2022). The entire governance structure (See Table 2) must work in a holistic way.