rtificial intelligence, as the latest IT marketing term is called,
has always been a great way to render information
untrustworthy. Newspapers and X users use it to create
unpleasant-looking cartoons of their political enemies.
Artists use it to turn celebrity promo shots into porn.
Now we know it’s also a good way of sneaking results of
questionable utility into the scientific literature.
The latest claim is that it’s creating new viruses. The article is paywalled and it rapidly disappeared off Science’s main page, which means that even Science mag is ashamed of it.
The authors used two chatbots called Evo 1 and Evo 2 to create mutated versions of an E. coli phage called ΦX174, a phage commonly used in evolution experiments. (A phage is a small circular DNA that infects bacteria.) They created thousands of candidates, of which only 300 were worth synthesizing and only 16 of which were actual viable phages. The authors say they found that a “cocktail,” i.e. an unknown mixture of the products, was partially able to overcome ΦX174 resistance in bacteria. This makes sense if the product was no longer similar enough to ΦX174 that the bacterium no longer recognized it. A resistance gene, if there is one, could also have come from one of the other phages in the database and been inserted by chance. In other words, their “AI” program took a perfectly good phage, proposed mutations on it, and turned between 95% and 99% of them into slop.
Phage ΦX174 map showing the restriction sites. The user’s DNA is inserted in a restriction site. The orange circles are the 11 gene products, many of which are non-essential. The origin of transcription (Ori) is in A. Source: New England Biolabs
Creating new viruses is not only routine in science but practically mandatory these days. Most DNA vendors have software that allows you to select the various parts: you tell it the start codon you want, select the linker sites and antibiotic resistance gene you want, then paste in the sequence you want to test. There’s almost no need to mutate individual bases anymore. It’s something a first-year biochem student could do.
Of course, the paper is mainly a proof of principle. It was a nice demonstration that a chatbot doesn’t necessarily need words but can work on nucleotide bases. But it also goes down the same path that “AI” has taken: taking perfectly useful information and turning it into junk. The starting phage already worked: if the chatbot had done nothing at all, the result would have been between 18 and 128 times better.
What this paper actually shows is that you can train a chatbot by putting thousands of perfectly good sequences in and get 16 good ones and 284 useless ones out along with thousands of sequences that are garbage. That’s what passes for progress in AI-land.
One possibility is that it was a stunt designed to panic people into passing laws against AI. Since AI doesn’t yet exist, that would be like passing a law against time travel. Or maybe they just forgot to edit out all the extra em-dashes.
It’s tempting to dismiss it as more fake news like those tabloid stories that say “WW3 fears explode” and “Panic in Russia!“; and indeed one now says “Panic as AI creates never-before-seen viruses—risk of ‘serious harm’.“ In fact, it’s an example of how researchers like to try out new techniques to see if they work.
It’s worth remembering that the software has no idea what any of this means. Like the AI at some company that supposedly went berserk and started hacking into their competitors’ systems, what they’re calling AI is just another computer program. It knows nothing, it decides nothing, and there’s zero chance of consciousness or even rudimentary awareness ‘emerging’ from it. While someone could theoretically use it to create deadly new diseases, doing it this way would be very inefficient: even the Wuhan Institute of Virology had to start with a bona fide disease virus and use conventional techniques to increase its tropism, species selectivity, and pathogenicity.
The authors describe it as a new computational technique like simulated annealing. The claim is that a chatbot is a simple technique for producing mutations that can be used for directed evolution. And indeed, it’s a promising new tool for studying evolution. So far it appears to have the random mutation part down cold.
aug 09 2026, 5:38 am
