data poisoning

Researchers at OpenAI and Microsoft have published a paper describing a 'doom loop' in which large language models, trained on data scraped largely without consent from human creators, degrade the open web and eventually poison their own training data. The paper reportedly acknowledges that people will come to see the models' wholesale ingestion of creative work as an unprecedented act of theft, reigniting debate over copyright and the sustainability of generative AI.

A widely shared essay argues that large language models and data poisoning are being weaponised by marketers and powerful interests to manufacture false consensus online. The piece warns that synthetic content can bend public perception of reality, flooding training data and information channels with coordinated narratives that look organic. Readers are debating how easily AI-generated influence operations can distort what appears to be public opinion.

Commenters are arguing that coverage of Iranian AI-generated disinformation serves to legitimise expanded surveillance. The claim is that framing foreign actors as poisoning media with AI builds public support for collecting mass data to detect such campaigns, infrastructure that critics say is then turned against domestic populations. They add that platforms flagging coordinated behaviour train on data gathered under these justifications.

Researchers Timnit Gebru and Emily Bender are highlighting a CNN report arguing that large language models are entering military systems quietly, not through humanoid robots but through corrupted or manipulated data that could skew military decisions. Commenters warn that poisoned information fed into defence systems carries escalation risks, potentially up to a global conflict, and say the issue deserves far more media attention than it is getting.