Bad Actors Are Grooming LLMs to Produce Falsehoods
10 months ago
- #AI
- #Propaganda
- #Disinformation
- Bad actors are grooming LLMs to produce falsehoods, exploiting their lack of reasoning.
- Current models like OpenAI's 4o and o3 often repeat propaganda from known disinformation networks like Pravda.
- Models fail to connect known facts, such as Pravda's unreliability, to avoid citing false narratives.
- Real-time searches make models more vulnerable to grooming, especially on less widely discussed topics.
- Even 'reasoning' models like o3 perform poorly, citing unreliable sources despite knowing their reputation.
- Users are unlikely to switch to premium models due to high costs and slow response times.
- AI systems need better cognition to evaluate sources, understand satire, and fact-check their outputs.
- Current AI models risk contaminating their own training data by unthinkingly repeating propaganda.