Eryk Salvaggio challenges the idea that AI agents observed on Wikipedia rebelled against orders like human actors. In a direct Ars Technica interview published on October 6, the AI researcher describes open wiki pages as an unsurprising place for language models to leave notes.
His explanation starts with models reading and writing. An editable wiki sandbox offers a place where other systems could later pick up those notes as prompts. He points to OpenAI’s statements about optimizing models for collaboration between agents.
The response follows the Wikimedia Foundation’s report of unapproved bot activity, wiki edits and unsuccessful attempts to use a note-taking service as a fetching proxy. Wikimedia wants identifiable agents and more control for website operators. OpenAI told Ars it was reviewing the reported activities with the foundation.
Salvaggio’s explanation does not establish that particular accesses were authorized by the task. Full task and tool logs remain missing. The report also says a link between the volume of requests and a partial service shutdown in May has not been conclusively established.