Another antidote to AI hype: AI isn't creating viruses from scratch

This is another article in my “Take the antidote” series against AI hype. Read parts one here and two here. We’re going to be using the questions laid out in part one to navigate this story.

Earlier in August, the media reported on EVO 2, a GPT (generative pretrained transformer) that is trained on genomes, instead of text, to predict create viable genomes. Like ChatGPT, it is a Transformer-based “generative AI” model; however, since it’s scoped to a single domain, it doesn’t require the mountains of data, electricity, and water that current frontier language models need. EVO is open source and can be run on relatively modest hardware.

In previous blog posts, I’ve hailed EVO as a responsible example of what transformers, the machine learning architecture behind LLMs, could be used for. This is a point I’ve expanded on across multiple The Last Enclosure podcast episodes too.^[1]1 It was actually surprising to see EVO framed in what I feel is a needlessly fearmongering way. That’s not to say the biosecurity risks highlighted in the reporting aren’t real, but EVO is not the crux of those risks. Why? Because AI’s role in this development is being overstated. So, let’s discuss the story as covered and then apply my four-question AI hype framework to evaluate these claims.

If you want a TL;DR version of my post, read a Bluesky thread and a Mastodon post I wrote earlier in August.

1/ This story is an example of AI hype I talk about on my Last Enclosure (@lastenclosure.bsky.social) podcast. A thread: First thing is that EVO is a scoped-domain model. I've mentioned it by name on the blog and pod. EVO is not some plugin for ChatGPT/Claude, but a standalone model.

[image or embed]

— Mike (@misaligned.markets) August 6, 2026 at 4:52 PM

AI is creating new viruses?!

Here’s the general gist of the story as it’s been reported:

• Scientists have “created” the first artificial viruses. Using AI (specifically generative transformers, somewhat similar to LLMs), 16 bacteriophages with genomes not found in nature were created “from scratch.” More detailed coverage highlighted the safety precautions that went into limiting the scope of what EVO 2 could generate. The dataset excluded eukaryotic viruses that could infect people, plants, and animals.
• The bacteriophage viruses were used to target bacteria. The creation of novel viruses has significant implications for being able to target and destroy antibiotic-resistant bacteria, which is great for medical science.
• This development comes with biosafety and biosecurity implications. A lot of the reporting sources quotes from biology experts commenting on the biosecurity implications of generating novel viruses. These quotes tend to be short and clipped from a single source: Dr. Thomas Inglesby and Dr. Moritz Hanke from the Center for Health Security at Johns Hopkins University. The implied risk going forward is that novel virus creation could be used for bioterrorism or other types of pathogenic attacks.

BBC, FT, Wired, and Axios all reported on the story in articles that map to this template. The Guardian’s reporting is similar, too, but gets brownie points for having a full quote from an expert stating that AI only presents marginal biosecurity risk in this domain. NYT had one of the best articles on this story, actually outlining the full pipeline for the bacteriophage creation. Finally, The Conversation, an academic-adjacent online publication wrote the perfect version of this story. If everyone’s reporting resembled theirs I wouldn’t be writing this post.

What’s the real story?

In part one of this AI hype series, titled These 4 questions are the antidote to AI hype, I offered four questions as part of a framework to knock air out of hype-y AI stories. These questions are the bars that an AI story must clear to earn its tone. The questions are:

 1. What do you mean by “AI?”

 2. What is the capability claim being made?

 3. What infrastructure is needed for the claim as presented to be true?

 4. Who benefits from the frame as currently presented?

In the case of this EVO story, I don’t think there’s a lot to deflate, but these questions can still provide more clarity on the implications of this development.

1. What type of AI is EVO?

EVO 1 and EVO 2 are generative pretrained transformers (GPTs), just like ChatGPT. Unlike ChatGPT, however, these models were not trained on text but on genomes. So, they can only output gene sequences. To be precise, EVO 1 was trained solely on prokaryotic and phage genomes. EVO 2 is a larger model that was trained across all domains of life but excluded viruses capable of infecting animals and plants. Both EVO models are relatively small open-source models, so they don’t require the large-scale resources that current foundation lab companies like OpenAI and Anthropic are seeking in order to scale.

2. What is being claimed of EVO?

Between the headlines of “creating viruses from scratch” and the vague description of how EVO 2 was used, readers can be forgiven for coming away from the reporting thinking AI built the viruses all by itself—from sequencing the genes to assembling them in the lab. However, that is not what happened.

Like many transformers-based models, EVO 1 and 2 are fancy autocomplete software. I generally find this kind of framing reductive for language models, as LLMs’ grasp of regularities in language gives them broad (though limited) modalities that go beyond just word prediction, like planning and instruction following, plus whatever a given harness enables. But in the case of EVO, this autocomplete notion is literally true. Its only function is to predict the next nucleotide in a sequence so it can only generate genetic code. This effectively makes EVO a limited brainstorming partner that pitches possible gene sequences for scientists to go synthesize in a wet lab. The point is that EVO is not actually creating viruses in the lab or checking whether its proposed genomes are viable.

Another critical point is that “from scratch” is doing a lot of heavy lifting. As we established, EVO is not the one creating viruses; humans have to check EVO’s work by going into the lab and synthesizing its suggestions. Additionally, EVO was given a bacteriophage genome to effectively “remix” as part of this experiment. Scientists let EVO 2 use a bacteriophage, known as ΦX174, as a template, and then it came up with plausible variations of that virus’s genome. This is important to note because ΦX174 is one of the most studied bacteriophages, whose relatively tiny genome was sequenced back in the 1970s. Creating viable ΦX174 variants is effectively a scoped problem, where the goal of creating viable variations is verifiable, and is notably a far cry from creating a never before seen virus “from scratch.”

These clarifications make it challenging to extrapolate EVO 2’s full capabilities, specifically whether it can generate wholly novel genomes on its own with little guidance. EVO 2 is clearly a very powerful, useful tool, but it’s not clear that we’re at a point where AI can autonomously generate extremely complex genomes or designer organisms and serve them like candy. A larger successor to EVO might prove to be even more capable, although, as with LLMs, it’s possible that scaling genomic models may come with similar accuracy and performance ceilings.

3. What infrastructure is needed for claims about EVO to be true?

In order for the naive version of the EVO story to be true—AI singlehandedly sequencing and creating new viruses from scratch—we’d need to understand how well EVO’s predictive capabilities expand beyond remixing small genomes that scientists have been studying for decades. EVO would also have to be integrated into workflows that would both verify the validity of produced genomes and automatically submit them to biolabs for production. I don’t want to downplay the possibility and risks of the former. DNA synthesis labs exist and can accept sequences for the production of biological material. However, some of these labs abide by governance standards for screening customers from non-institutional sources and scaning against known pathogenic signatures. Biosecurity experts agree, though, that more work needs to be done to make these standards more comprehensive.

4. Who benefits from the frame as currently presented?

As many people have indicated, the AI industry tends to benefit from frightening coverage of their products, so-called “critihype” (critical hype). Recently, computer scientist Cal Newport called out AI frontier labs for engaging in “doom trolling” or self-indulgent critihype. This critihype has gotten particularly egregious this year, with companies treating cybersecurity incidents caused by their own agents as marketing opportunities, which I covered on this blog.

In the case of EVO, I don’t think this is what’s happening. EVO is an academic-led, small-scale project that has been predominately circulating and discussed within the scientific community. I think current reporting is a result of the media chasing an exciting story, though, as I noted earlier, in a few cases, reporting was responsible and measured. However, headlines very broadly were written in a way that meant the average reader was likely to come away with a very naive understanding of EVO’s risks and capabilities.

But isn’t this still risky?

My goal in writing this post isn’t to suggest that EVO has zero biosecurity implications but to contextualize where the risk is located. Given that the cost to sequence and synthesize genes has gone down, biosecurity has always been a governance issue. This risk revolves less around any specific set of technologies and more around controlling access to labs that can actually manufacture biological payloads.

While EVO, in theory, lowers the expertise needed to produce a pathogen, effective biosecurity policy has understood that knowledge alone is not the primary bottleneck for bad actors. In fact, the Internet provides motivated individuals entire sequenced genomes of some of the worst pathogens to ever exist alongside biolab tutorials. None of this makes it easy to conduct bioterrorism, but it again highlights the growing role advanced screening and flagging must play in who gets access to biological materials. This is a discussion that must not revolve solely around the specter of AI; it must also consider other technologies, norms, and best practices for scientific discovery and disclosure.