# created ```dataview LIST WITHOUT ID file.link FROM -"windsurf" WHERE file.cday = date(this.file.name) AND !startswith(file.folder, "windsurf") SORT file.cday ASC ``` # modified ```dataview LIST WITHOUT ID file.link FROM -"windsurf" WHERE file.mday = date(this.file.name) AND !startswith(file.folder, "windsurf") SORT file.mday ASC ``` --- # [[diary and thoughts]] #thought #diary --- # Insidious LLM Propaganda It is clear that generative AI can be used to create propaganda. In May 2025, we saw a crude example: [[system prompt]] tampering led XAI's chatbot Grok to insert comments about alleged "white genocide in South Africa" (Preston, 2025) in its responses to completely unrelated questions, such as those about cats or cartoons. In this talk, I consider how LLMs might become instruments of propaganda in more insidious ways. I argue that, because it is unlikely that LLMs are intentional agents in any interesting way, it is implausible to think that they could themselves be propagandists. If this is correct, then treating them as if they are like human propagandists may not produce effective propaganda, or, at least, may be a sub-optimal way of producing propaganda with an LLM.  I then consider two training phases of modern LLMs: pre-training and post-training (e.g. alignment training). I suggest that post-training may not be able to produce propaganda with a fine enough grain as to be very effective, and argue instead that pre-training is likely more effective because it leads to a particular type of _semiotic media_. By this I mean that pre-trained LLMs are grounded on learned statistical patterns—relationships between concepts (paradigmatic) and sequences (syntagmatic) extracted from [[training data]]. Such a medium could be corrupted during pre-training through selective corpus design: adjusting co-occurrence frequencies, privileging certain framings, or systematically excluding alternatives. Such manipulation would embed bias directly into the model’s probability distributions, [[not just]] its surface behaviour. [[The result]] would not be overt propaganda, but systematic tendencies that privilege certain ideas and frames by making them more salient. Therefore, the most effective method of producing LLM propaganda may well be to [[target the]] pre-training substrate where meaning-making patterns are first formed. ### References - Preston, D. (2025, May 16). Grok’s white genocide fixation caused by ‘unauthorized modification’. [[The Verge]]. [https://www.theverge.com/news/668220/grok-white-genocide-south-africa-xai-unauthorized-modification-employee](https://www.theverge.com/news/668220/grok-white-genocide-south-africa-xai-unauthorized-modification-employee)