Agents for Fighting Misinformation Spread on Twitter: Design Challenges

Abstract

Containing misinformation spread on social media has been acknowledged as a great socio-technical challenge in the last years.Despite advances, practical and timely solutions to properly communicate verified (mis)information to social media users are an evidenced need. We introduce a multi-agent approach to bridge Twitter users with fact-checked information. First, a social bot, which nudges users sharing verified misinformation, and a conversational agent that verifies if there is a reputable fact-check available and explains existing assessments in natural language. Both agents share the same requirements of evoking trust and being perceived by Twitter users as an opportunity to build their media literacy. To this end, two preliminary human-centred studies are presented, the first one looking for an adequate identity for the bot, and the second for understanding preferences for credibility indicators when explaining the assessment of misinformation. The results indicate what this design research should pursue to create agents that are consistent in their presentation, friendly, engaging, and credible.CCS Concepts: • Human-centered computing → Natural language interfaces.

Exploring the impact of automated correction of misinformation in social media

“…The observation that alerting , factual , and friendly templates were more likely to engage users with the bot corrections correlates with the result in Piccolo et al. (2021) where such templates were preferred over the other templates. However, in our analysis, this association is only observed for the most active users in our data.…”

“…2020; Piccolo et al. 2021; Wang 2022; Huang and Wang 2022; Velez, Porter, and Wood 2023; Boukes and Hameleers 2023). One advantage of lab experiments is that complex variables and scenarios can be investigated.…”

“…Most misinformation correction studies involved some type of controlled lab experiments (Fridkin, Kenney, and Wintersieck 2015;Wood and Porter 2019;Ecker et al 2020;Piccolo et al 2021;Huang and Wang 2022;Velez, Porter, and Wood 2023;Boukes and Hameleers 2023). One advantage of lab experiments is that complex variables and scenarios can be investigated.…”

“…The base templates we chose are listed in Table 1, which are designed around different message tones (e.g., use of informal or sympathetic language). These templates are based on marginally modified versions of the ones proposed by Piccolo et al (2021). To form a reply, the bot randomly selects one of these templates, and substitutes <FACT-CHECK-URL> token with the fact-checking URL that corresponds to the shared misinforming URL, and the <VERDICT> token with the fact-checker rating of the shared misinforming URL.…”

“…An initial limited user study by Piccolo et al. (2021) on the best tone to use for correcting Twitter/X users showed that individuals preferred factual and alerting corrective messages as they were identified as not being aggressive and confrontational whereas suggestive and empathetic messages were judged more negatively as being judging, aggressive, and condescending. The findings were limited and not tested on real misinformation sharers.…”