A web browser Dillo has proposed a method for filtering out changes prepared via AI. The Dillo project accepts patches created only by humans, but reviewing submitted changes is time-consuming, and it is not always immediately clear whether a patch was created by a person or not. To simplify the filtering of AI-generated patches, participants submitting changes to the project for the first time are asked to provide a recording of their development session as evidence of their work. When using Vim, the session can be recorded, for example, with the asciinema utility.
The idea has not been converted into mandatory requirements and is still in the discussion and experimentation phase. At first glance, a developer trying to deceive the maintainer and present a patch created through AI as their own code might also use AI to generate a recording of the session in asciinema. Experiments have shown that the artificiality of such recordings is immediately noticeable, as they do not reflect the natural actions, mistakes, and thoughts of a person during the work process. Generic AI models are trained on existing code, but their training does not fully encompass the coding process carried out by humans.
Forrás: opennet.ru
