What are the implications for biodiversity, regulation and democratic oversight?
Ever since artificial intelligence (AI) has become broadly available, its use has been a source of constant controversy. As in other fields, much of the debate in the biotechnology sector centres on (bio)security gaps, particularly in view of the growing importance of biological AI tools and developments in genetic engineering (GE).
The convergence of artificial intelligence and genetic engineering, abbreviated as AIxGE, is gaining a lot of attention, as it promises great benefits. However, since it also enables re-writing complete genomes virtually from scratch, or re-designing entire ecosystems, it also holds considerable risk potential.
There are many concerns, especially about biosecurity i.e. the deliberate misuse of AIxGE. Economic aspects are also becoming more and more important. The US and China, in particular, are engaged in political and economic competition over AIxGE, which is seen as key to technological superiority, prosperity, security and military supremacy. With the current wave of deregulation in the biotechnology sector, the EU is also taking a stance in the competition for global leadership in this field.
However, what is largely missing from these discussions is biosafety: the risks to people and the environment and biodiversity when AI-designed organisms are released into the wild.
This article makes that case in three steps. Firstly, it looks at what is changing: how AI and genetic engineering are converging, and what that makes possible. Secondly, five examples – from engineered trees to redesigned viruses – demonstrate the risks that current safety checks cannot capture. Finally, how regulation is being loosened just as the risks grow, and what needs to happen instead.
What is changing: AI meets genetic engineering
In recent years, there has been increasing convergence between GE methods, AI and the automation of development processes. This is driven by increased availability of molecular and biological big data, as well as expanded AI model capabilities to analyse and design molecular entities and biological functions.
AlphaFold, an AI model that predicts protein structures, is now widely used. Multimodal ‘foundation models’ such as Evo or Bacformer are capable of generating or re-designing entire genomes. Alongside this, self-driving labs are emerging – so-called ‘biofoundries’ run experiments with minimal human input.
Combined with new genomic techniques (NGTs) such as the CRISPR/Cas ‘gene scissors’, these tools are very powerful. AIxGE is now being promoted for agriculture, environment and food industries, as well as animal or plant breeding. It is usually presented as a transformative solution to pressing problems and offering opportunities for sustainability that can be rapidly implemented.
Many of these claims could be described as ‘hype’. Even worse, releasing AIxGE organisms outdoors carries various risks to health and the environment, as the following examples show.
What could go wrong
Each example below shows the same pattern: organisms with genetic changes that are new to nature, released into environments where their effects cannot be predicted.
Plants that could weaken wild populations or disrupt ecosystems
With the help of AI, genetically engineered poplars with a reduced lignin content have been developed. These are intended to make wood processing and paper production easier. At the same time, they exhibit a previously unknown combination of genetic changes and are new to the environment, which is why their release may involve certain risks.
Poplars readily spread via pollen, seeds, or broken branches, and they grow almost anywhere. A lower lignin content is usually associated with lower stability and less resistance to tree pathogens. Therefore, the uncontrolled spread of NGT poplars, and the transfer of these traits with possible negative trade-offs to wild populations, could become a threat to protected species such as the black poplar.
The same applies to NGT plants with an altered flower structure to facilitate pollination carried out by AI-trained robots. These NGT plants also exhibit a previously unknown combination of genetic alterations that are unlikely to occur through conventional breeding. However, colour, scent and structure of flowers are crucial to pollinator preferences. Engineering them could shift the preferences of pollinators or even disrupt established plant-pollinator interactions – with cascading effects on entire ecosystems.
Nevertheless, in many countries AIxGE plants can be put on the market without environmental risk assessment and without any containment measures having been put in place.
Engineered animals that could displace wild populations
Similar trends can be seen in animal breeding, where AIxGE is used as a base technology for future animal genomics and smart agriculture, in order to accelerate genomic ‘improvement’ in livestock. The EU is preparing to loosen rules here too: the European Food Safety Authority (EFSA) has confirmed that there were no “novel hazards”, and the EU livestock strategy advocates for “cutting-edge genomic and animal breeding techniques [that] are a win-win for both economic and environmental outcomes.”
The most prominent animal-related projects at present, however, are attempts by companies like Colossal Biosciences to ‘bring back to life’ extinct species, such as woolly mammoths, dire wolves or moas. With the help of AIxGE they are altering genes that are responsible e.g. for shaggy hair, bigger size, or larger jaws. However, what they are actually creating are genetically engineered versions of existing wildlife species like Asian elephants, grey wolves, South American birds, and laboratory mice with ‘woolly mammoth fur’ – all of them featuring genetic changes previously unknown in these species.
If these genetically engineered wildlife, such as the ‘dire wolf’, were to escape into natural populations of grey wolves, it is possible that they could reproduce. As these animals are larger and stronger than their natural counterparts, the genetically engineered grey wolves and their offspring would have a fitness advantage, and thus be capable of displacing natural populations. This would have significant consequences – not only for the protected wolf species, but also for ecosystems, and possibly humans if the genetic changes result in changed hunting behaviour.
Creation of ‘dark matter’
AI can now design genetic material with no natural equivalent. It can analyse genetic interactions and generate novel regulatory sequences. This has already led to the successful integration of artificial regulatory sequences into several bacterial species without prior experimental data.
Going further, researchers have used the genomic language model Evo to develop a ‘semantic design’ for bacteria. To this end, novel genes were generated that also incorporate changes to the genomic context. Unlike other genetic engineering methods, this approach does not rely on existing gene sequences to create new variants. Instead, a certain effect or function is specified as the target and the AI generates the corresponding genetic code. The results bear no resemblance, in sequence or structure, to anything known, showing that these models can reach well beyond natural biology. Reports on the project describe the creation of “a genuine ‘dark matter’ protein […] without significant matches in existing sequence or structural databases […], thereby transcending the confines of established biological knowledge”.
This makes risks to the environment and ecosystems very hard to predict.
Genetically engineering microbes so that they produce more of one substance may unintentionally alter other microbial characteristics, and with them the way they interact with their environment. Such considerations become even more relevant if new gene functions are established that can create completely novel proteins. Furthermore, the specific effects of the ‘dark matter proteins’ cannot be predicted through comparison with existing data, for example, on toxins. Therefore, current methods and guidance are insufficient when it comes to assessing health and environmental risks.
Rewriting entire viral genomes
AIxGE can also completely redesign entire genomes, not just individual genes or gene functions. This has been shown by the generation of functioning viruses from scratch. Here, the AI not only provided instructions for changing individual DNA segments, but also reassembled the virus’s entire genetic material, altering many functional details. The resulting AIxGE viruses were not only fully functional, but were also capable of overcoming resistance and exhibited greater virulence than comparable viruses.
Since viruses continuously evolve, engineered versions may acquire new properties, such as becoming pathogenic to other organisms that are essential for life. And although the AI in this case was not trained with data from human pathogenic viruses, the same approach could in principle be applied to viruses that are dangerous to humans.
Genetically engineering microbes directly in the wild
AIxGE is also being used to change how species interact. Computer-aided design tools are being built to create organisms with artificial metabolic pathways that exchange signals with other engineered species – even across the boundaries between microorganisms, plants and animals.
Another approach transfers a CRISPR mechanism from one bacterium to another. This means whole microbial communities could be genetically engineered directly where they live – in the soil or in the gut – including species that are poorly understood or even unknown. The same technology could be used to transform the root microbiome of plants directly in the environment – as well as the microbiomes of humans and animals.
Why oversight is failing
New AIxGE methods are appearing so fast that strategies for containment and oversight are lagging dramatically behind. Existing methods and guidelines for the assessment of risks to humans and the environment are far from sufficient. Even EFSA has noted that “Even with the complete genetic information of a synthetic microorganism, it is beyond the capacity of any existent bioinformatic analysis to fully predict the capability of a synthetic organism to survive, colonise and interact with other organisms under natural conditions, given the uncountable diversity of potential microhabitats and their temporal variability.” Consequently, the release of AIxGE organisms into the environment would certainly present multiple challenges or even be considered irresponsible.
There is no track record with these organisms and far too little data, so neither AI models nor their developers can reliably predict what will happen. The release of such organisms, especially the ones designed to alter other organisms, would turn the biosphere into an open-air laboratory without effective means of control, posing risks that cannot be assessed.
Meanwhile, large biotech companies and their umbrella organisations are lobbying, largely out of public view, for the removal of any regulatory hurdles. And they are succeeding, as these demands from industry have been taken on board quite successfully, for example, by the EU, which recently deregulated NGT plants. However, with the parallel launch of further “simplification of EU rules” in terms of Omnibus Packages and Biotech Acts to strengthen Europe’s competitiveness – such as the current attempts to deregulate genetically engineered microorganisms – it becomes clear that this is only the beginning.
Conclusion
AI has changed genetic engineering in kind, not merely in degree: whole genomes, and the relationships between species, can now be (re)designed at breathtaking speed. The environmental consequences of that shift remain the least examined. Weakened trees, displaced wolves, proteins and viruses with no natural precedent – these are the foreseeable consequences of releasing organisms whose behaviour neither established methods nor the models themselves can anticipate. And as uncertainty deepens, oversight is being dismantled, with industry lobbying and the pursuit of competitive advantage steadily eroding the rules that remain.
As AIxGE widens the scope for harm, the standards applied to it should rise in proportion, not fall away in the name of competitiveness. There is an urgent need for international cooperation and governance in order to contain the associated risks. Otherwise, without effective measures of control, we risk losing democratic oversight and the ability to prevent lasting damage to health and the environment.