Technologie · ArchivesL’IA a créé 16 virus vivants qui n’avaient jamais existéRédigé 7 septembre 2026 à 21:22 UTCRegarderTranscription0:00Researchers at Stanford University and0:02the Arc Institute in California have0:04used AI to build viruses from scratch0:07and brought them to life in the lab. 160:09functional viruses that never existed in0:11nature before in a first demonstration0:13published in the journal Science. An AI0:15model called Evo, trained [music] on0:17massive amounts of genetic data to read0:19DNA patterns like grammar, generated0:22hundreds of thousands of [music] viral0:23genomes from scratch. Scientists built0:25and tested 285 of those designs and 160:28proved fully functional, able to infect0:30[music] bacteria and reproduce0:31themselves. The pipeline started with0:33Evo learning the grammar of DNA, then0:36generating new genomes that researchers0:38synthesized [music] and tested for0:39infection. Those working viruses are0:41bacteriophages based on a harmless0:43[music] phage that infects only0:44bacteria, not humans. Researchers via DW0:47News say that [music] is why they are0:49not a threat to humans, and the phages0:51target bacteria including Vibrio and E.0:53coli.0:53>> [music]0:53>> The team deliberately excluded human0:55viruses from Evo's training data, but0:58the study itself notes the new0:59capability raises biosafety, [music]1:01biocontainment, and biosecurity1:03considerations, and that current1:05strategies to manage those risks have1:07limitations.1:08>> [music]1:08>> DW science reporter Zulfikar Abbany1:11contrasted Dolly the sheep as a copy of1:13a living animal, while these viruses1:15were made from scratch [music] and never1:16existed before, making an AI1:18hallucination like generation1:20consequential [music] once rendered into1:21a living organism.1:28Supporters via DW News say the same1:30design [music] power could help attack1:32and kill bacteria, diagnose infections1:34by showing bacteria exist, and combat1:36drug resistance where antibiotics fail1:38with [music] potential for new medical1:40treatments, gene therapies, and biotech1:41tools. Abbany warned it would not take a1:44huge leap for other viruses to be1:45developed. It comes down to what you1:47prompt the AI and how you train it,1:49noting people had already tested AI1:51systems to develop or try to develop1:53biological weapons in the post-Chat1:54GPT-3.5 [music]1:56and 4.0 era. Though he noted an AI model1:59on a screen is one [music] thing and2:00creating a living organism is the next2:02stage that still needs humans in the2:04lab. [music] Governments and regulators2:06are still trying to figure out where the2:07guardrails should be2:09>> [music]2:09>> via DW News. The European Union has its2:11AI Act aiming to ensure trust in AI, but2:14no specific [music] enforced guardrail2:16for living AI generated viruses is2:18verified. Leaving the method published2:20and replicable while the human stakes,2:22potential medical gains against the risk2:24that [music] intentional misuse or2:25inadvertent consequences outpaces2:27oversight remain unresolved.Lire l’articleChronologie de l’histoireHeyDexLe format court ↗
