the challenge of making content generated by AIs discoverable

Should we settle for a world where it is impossible to distinguish content generated by artificial intelligence (AI) from that produced by humans? The question is more burning every day: bluffing texts have proliferated since the launch, in November 2022, of the chatbot ChatGPT, and misleading photos like that of the Pope in a white down jacket are set to multiply with the rise of software like Midjourney.

In response, some are looking for ways to make this synthetic content detectable. The challenge is complex, but topical: on Tuesday May 23, the software giant Microsoft announced solutions in this direction and the French Minister of the Economy, Bruno Le Maire, raised the issue in Paris with Sam Altman, the CEO of OpenAI, the creator of ChatGPT.

[Rendre détectables les contenus créés avec l’IA] would help combat cheating at university, or the m generation of propaganda and misinformation with the aim of, for example, flooding blogs with comments supporting the invasion of Ukraine,” To argued, in a conferencein November 2022Scott Aaronson, the researcher responsible for working on this question at OpenAI.

“Maintaining distinctions is an ethical imperative for reasons related to the uses of AI, in education, health or law, but also, at the philosophical level, to delimit what is human responsibility and what is made by machines, adds Alexei Grinbaum, member of the National Digital Ethics Pilot Committee and author of Word of machines (Humensciences, 192 pages, 17.90 euros).

Built-in or external tools

With this in mind, Microsoft has announced the integration of a “invisible cryptographic watermark” (Or water mark) in the images created by its Designer and Bing Image Creator software: by consulting the metadata of a photo or video – i.e. the information attached to this file –, “the user will be able to see that it was created with an AI”, explains the group.

Available ” in the coming months “, this indication of the ” origin ” of content is based on a standard called C2PA. This has also been integrated by Adobe into the image editing tool thanks to the AI ​​available in its famous Photoshop software. OpenAI is also studying watermarking techniques for Dall E 2, its software for generating images from a text description. And its competitor Midjourney has adopted a metadata system created by IPTC, a media industry standards body.

You have 61.93% of this article left to read. The following is for subscribers only.

Source link

Leave a Reply