Disinformation is a rapidly evolving process, and it is becoming more sophisticated with advances in and successful integration of Machine Learning (ML) and Artificial Intelligence (AI) technology. Where AI represents machine intelligence as compared to human intelligence, ML is the subset of AI which represents algorithms that perform specific tasks without explicit instructions, as they rely instead on inference and patterns. This article examines the ways ML/AI have transformed the art of disinformation campaigning, and to what extent this development will effectively shape public opinion and political discourse. ML/AI technologies are increasingly sophisticated in manipulation of both the rate and nature of information production. It is subsequently more crucial than ever to strike a balance between countering disinformation with user freedoms and overall credibility.
By Rana Al-Nusair
People have always used the term “seeing is believing” and technology professionals now speculate whether this term may have become obsolete due to progressively sophisticated iterations of ML/AI technology. Deepfake technology is a current challenge to the aforementioned points in its use of ML.[1] This software alters audio and video, to the extent that an individual with enough skill, understanding, and technological tools can create personally-targeted content which can make the subject appear as if he/she has said or done something which has not been said or done.[2] This software is particularly dangerous in the political sphere and can result in the undermining of leaders. [3]
Disinformation campaigns are currently occurring both at unprecedented quantitative and qualitative rates. It would certainly be difficult to argue that people are not experiencing a never-achieved-before state-of-the-art information war. Can the same tools that are used to create highly sophisticated false content be the same tools used to reliably counter that very content as well? With ML/AI, a rogue actor can create false content that is uniquely and intelligently tailored to the unsuspecting target audience. ML/AI introduces techniques which can easily and cost-effectively penetrate many more layers than ever experienced before in the realm of disinformation campaigns, mainly the cognitive, sociological, and political spheres. With ML, algorithms learn more information at a faster rate about any given individual or a group; this advancement stems from the vast amount of data that is readily available to the algorithms’ training sets.[4] In addition, an actor can acquire tailored information without necessarily having to engage in a physical or any kind of compromising interaction with the targeted person or audience.
OpenAI, a leading AI research company, has created software known as the GPT-2, a disinformation detection tool which spots potential bot accounts on social media based on analysis of linguistic patterns, in addition to evaluation of the spread of bot-generated disinformation content over time. In February 2019, OpenAI’s research lab stated it would not release the full code for the GPT-2, as it might be used by adversaries to generate spam and false content. In November 2019, however, OpenAI decided to release the full GPT-2 software.[5] The GPT-2 had initially trained on eight million web pages, which were not disclosed to the public. After full disclosure of the software, OpenAI revealed that the software now had trained on one and a half billion parameters.[6] Google has expressed the same concerns over misuse of their software when they released a paper revealing that they had put constraints on research software previously shared.[7] As machines become better at mimicking their training data and thereby generating higher quality content, which could be indistinguishable from reality, malign actors –if equipped with the same level of algorithmic sophistication– could increase the quality of their fake content in such a way as the GPT-2 by itself might not be able to reveal said content as either fake or real.
Disinformation detection software is also becoming more sophisticated, using tools such as topic detection, viewpoint clustering, narrative identification, and topic/viewpoint/narrative user interface.[8] What makes disinformation detection software, such as the GPT-2 or Grover, so effective at spotting fake content, is that the software itself is also highly adept at creating fake content, using sophisticated techniques such as signal processing analysis, physics-level analysis, semantic and physiological signals.[9] Cybersecurity professionals regularly penetrate their own software to discover their weaknesses in order to improve upon them and decrease the possibility of real adversarial attacks. As Yejin Choi, a researcher at AI2 says, “the best models for detecting disinformation are the best models at generating it”.[10]
Regarding the quantitative dimension, a study in 2017 estimated that about 15% of Twitter user accounts are bots, and this grew from approximately 8.5% based on an earlier 2014 study.[11] In 2104, this percentage represented 48 million Twitter bot accounts from the entire demographic of Twitter users, a conservative percentage at best. Computational propaganda can be produced with virtually no effort, minimal-to-no cost and is effective and quick. Dissemination of Deepfake videos, for example, is growing exponentially alongside the software’s’ low cost distribution. Arguably, producing these videos is easier than doing something similar on Photoshop, due to the fact that these Deepfake videos rely on machine learning rather than manual design. Deepfake videos are usually produced by combining an already existing video with new audio and image, using a class of machine learning that is called generative adversarial networks (GAN).[12]
ML/AI technology transforms the current cyber races between states, groups and individuals on multiple fronts. ML/AI achieves this change from both its fundamental qualitative and quantitative intrinsic nature. This technology transforms the quantitative aspect of competition as well by advancing autonomous machine-driven propagation, incentivizing, and multiplying adversaries with the democratization of technology in which the costs of acquiring and using the technology rapidly decrease. This ability enables virtually anyone to radicalize or sensationalize an event that could be detrimental, not only to a few targeted individuals but to masses. Additionally, considering the accountability aspect primarily through the notions of plausible deniability, it is an integral part of the transformative aspect to computational propaganda. Transparency from the most powerful social media companies is likely a step towards countering computational propaganda; however, this step has proven difficult to establish, especially for private companies which strongly oppose criticism and are overly protective of their algorithms.
Sources
[1] Simonite, Tom. “Will ‘Deepfakes’ Disrupt the Midterm Election?” WIRED, 2018. https://www.wired.com/story/will-deepfakes-disrupt-the-midterm-election/
[2] Kertysova, Katerina. “Artificial Intelligence and Disinformation: How AI changes the way disinformation is produced, disseminated, and can be countered”, Security and Human Rights in Monitor. Pp. 11. https://www.shrmonitor.org/assets/uploads/2019/11/SHRM-Kertysova.pdf
[3] BuzzFeedVideo, “You Won’t Believe What Obama Says in This Video!”, April 18, 2017 https://www.youtube.com/watch?v=cQ54GDm1eL0
[4] Funk, McKenzie. “The Secret Agenda of a Facebook Quiz,” New York Times, 2016, https://www.nytimes.com/2016/11/20/opinion/cambridge-analytica-facebook-quiz.html
[5] Simonite, Tom. “To See the Future of Disinformation, You Build Robo-Trolls”. WIRED, 2019. https://www.wired.com/story/to-see-the-future-of-disinformation-you-build-robo-trolls/
[6] See tweet by OpenAI at https://twitter.com/OpenAI/status/1191764001434173440
[7] Simonite, Tom. “The Text-Generator that’s too Dangerous to Make Public.” WIRED, 2019. https://www.wired.com/story/ai-text-generator-too-dangerous-to-make-public/
[8] BUSSINESS WIRE. “Datametrex Introduces World Class Bot Detection and Fake News Filter”. MARTECH SERIES, 2019. https://martechseries.com/technology/datametrex-introduces-world-class-bot-detection-fake-news-filter/
[9] Lorica, Ben. “How AI Can Help to Prevent the Spread of Disinformation”. Information Age, 2019. Accessed December 3rd 2019. https://www.information-age.com/ai-prevent-disinformation-123479466/
[10] Kalim, Faisal. “Turning fake news against itself: AI tool can detect disinformation with 92% accuracy”. WNIP, 2019. https://whatsnewinpublishing.com/turning-fake-news-against-itself-ai-tool-can-detect-disinformation-with-92-accuracy
[11] Onur Varol, et. al. “Human-Bot Interactions: Detection, Estimation, and Characterization.” (2017).
[12] Goodfellow, Ian. et. Al. (2014). “Generative Adversarial Networks”. Proceedings of the International Conference on Neural Information Processing Systems (NIPS 2014). pp. 2672–2680.