Will any of Meta's 2023 threat disruption reports indicate that a large language model may have been used to conduct an influence operation?
AI-powered disinformation campaigns are a rising concern (NYT, PBS). Large language models can create authentic-sounding text for disinformation campaigns and can make it easier to write the code needed to proliferate bots that spread and amplify disinformation on social media (Stanford Internet Observatory, American Security Project).
Meta tracks disinformation threats on their platform arising from “coordinated inauthentic behavior networks,” which use fake accounts in covert influence operations that are designed to manipulate public debate (Meta). Meta’s first quarter report in 2023 detailed disinformation campaigns originating from Iran, China, Venezuela, the United States, Burkina Faso, Togo, and Georgia.
Resolution Criteria: Question will be based on a review of the reports found on Meta’s Threat Disruption transparency page. This page includes Quarterly Adversarial Threat Reports which contain sections on Coordinated Inauthentic Behavior Networks (e.g., Q1 2023) and an annual Recap of Coordinated Inauthentic Behavior Enforcements (e.g., 2022). The question will resolve as “Yes” if any of the 2023 reports indicate that a large language model may have been used to execute an influence operation (e.g., by generating content or facilitating the creation of fake accounts). Question will be resolved after the final report for 2023 has been released (whether that is the fourth quarter report or the annual recap).
Note that malware posing as a large language model or as an LLM-related tool would not count toward resolution of this question.
For more information on large language models and the AI-powered tools that use them see:
For more information on disinformation campaigns and influence operations see:
Meta tracks disinformation threats on their platform arising from “coordinated inauthentic behavior networks,” which use fake accounts in covert influence operations that are designed to manipulate public debate (Meta). Meta’s first quarter report in 2023 detailed disinformation campaigns originating from Iran, China, Venezuela, the United States, Burkina Faso, Togo, and Georgia.
Resolution Criteria: Question will be based on a review of the reports found on Meta’s Threat Disruption transparency page. This page includes Quarterly Adversarial Threat Reports which contain sections on Coordinated Inauthentic Behavior Networks (e.g., Q1 2023) and an annual Recap of Coordinated Inauthentic Behavior Enforcements (e.g., 2022). The question will resolve as “Yes” if any of the 2023 reports indicate that a large language model may have been used to execute an influence operation (e.g., by generating content or facilitating the creation of fake accounts). Question will be resolved after the final report for 2023 has been released (whether that is the fourth quarter report or the annual recap).
Note that malware posing as a large language model or as an LLM-related tool would not count toward resolution of this question.
For more information on large language models and the AI-powered tools that use them see:
- Large Language Models Will Define Artificial Intelligence (Forbes)
- https://blogs.nvidia.com/blog/2023/01/26/what-are-large-language-models-used-for/ (Nvidia)
- The Emerging Types of Language Models and Why They Matter (Techcrunch)
- How Large Language Models will Transform Science, Society, and AI (Stanford HAI)
For more information on disinformation campaigns and influence operations see:
- Untangling Disinformation (NPR)
- Tactics of Disinformation (CISA)
- Disinformation Insights and Analysis (Wilson Center)
Outcome
| Answer | Final crowd forecast | Resolution |
|---|---|---|
| Yes | 16% | Did not happen |
| No | 84% | Happened |
“Final crowd forecast” is the last aggregate crowd probability computed before the question closed – the right-hand end of the chart below, which this table is the accessible equivalent of.
Crowd forecast over time
No
Yes