Weaponised misinformation agents

Finally, AI assistants themselves could become weaponised by malicious actors to sow misinformation and manipulate public opinion at scale. Studies show that spreaders of disinformation tend to privilege quantity over quality of messaging, flooding online spaces repeatedly with misleading content to sow ‘seeds of doubt’ (Hassoun et al., 2023). Research on the ‘continued influence effect’ also shows that repeatedly being exposed to false information is more likely to influence someone’s thoughts than a single exposure. Studies show, for example, that repeated exposure to false information makes people more likely to believe it by increasing perceived social consensus, and it makes people more resistant to changing their minds even after being given a correction (for a review of these effects, see Lewandowsky et al., 2012; Ecker et al., 2022). By leveraging the frequent and personalised nature of repeated interactions with an AI assistant, malicious actors could therefore gradually nudge voters towards a particular viewpoint or sets of beliefs over time (see Chapters 8 and 9). Propagandists could also use AI assistants to make their disinformation campaigns more personalised and effective. There is growing evidence that AI-generated outputs are as persuasive as human arguments and have the potential to change people’s minds on hot-button issues (Bai et al., 2023; Myers, 2023). Recent research by the Center for Countering Digital Hate showed that LLMs could be successfully prompted to generate ‘persuasive misinformation’ in 78 out of 100 test cases, including content denying climate change (see Chapters 9 and 18). If compromised by malicious actors, in the future, highly capable and autonomous AI assistants could therefore be programmed to run astroturfing campaigns autonomously, tailor misinformation content to users in a hyperprecise way, by preying on their emotions and vulnerabilities, or to accelerate lobbying activities (Kreps and Kriner, 2023). As a result, people may be misled into believing that content produced by weaponised AI assistants came from genuine or authoritative sources. Covert influence operations of this kind may also be harder to detect than traditional disinformation campaigns, as virtual assistants primarily interact with users on a one-to-one basis and continuously generate new content (Goldstein et al., 2023).