In 2014, I was among Ukrainian media people, the graduates of one of the projects of the Mohyla School of Journalism in Kyiv, who created the StopFake project. This was our reaction to the avalanche of “fake news” about Ukraine. We and some other Ukrainian teams have begun to raise this issue with policymakers at the international level. But, in the very beginning, just a few of them listened to us: it was hard to believe that such scale of disinformation was possible. Perhaps we would not believe it ourselves if we did not face it.
We didn’t imagine that this was just the beginning and that disinformation campaigns would soon spread around the world. And this would happen in particular through the use of new technologies in the field of artificial intelligence and machine learning.

In four years I stood on the TED stage in Vancouver speaking about our experience. The international audience listened to me with much more attention. It was already known about the unauthorized access of Cambridge Analytica spin-doctors to personal data of Facebook users, as well as the influences of "troll factory" on the US elections in 2016 and Brexit. After that, the phenomenon of spreading disinformation online became one of the most discussed topics globally.

How does it work? The technology collects the data about the websites we visit, as well as our geolocation, purchases, and publications to integrate them into our consumer profiles. On this basis, algorithms define us as the audience for certain messages. In the case of Cambridge Analytica, this was done using an online quiz that gathered user information. AI systems test thousands of messages looking for the ones that will convince each audience. If old-style agitation is used to target the average supporter of a politician or party, then micro-targeting reaches all the ends of the ideological spectrum as well as audiences who have not previously voted. The messages may be the opposite. This allows attracting as many votes as possible. Technology can do the same for a propagandist who wants to divide society or undermine trust in individuals or institutions. This became possible because of the synergy of the powerful AI, availability of user data and lack of media literacy among users.
Following the Cambridge Analytica scandal, a number of countries have tightened their personal data protection laws. But there is still a lot of work to be done. A number of studies show that some governments and individual politicians can continue trying to use the power of social media to undermine public trust and wreak havoc.

Revolution and evolution of media consumption

It all started with a change in the process of news production and, consequently, its consumption. In the early 2000s, when I started my career in journalism, a whole army of journalists, editors, proofreaders stood in the user’s way to information. They filtered out overtly false or unscientific content. At that time there were no more censors of the Soviet era, although there were cases of state censorship. However, any media outlet potentially contained other types of censorship: publishers’ and editors’ censorship, as well as the most difficult for detecting - journalists’ self-censorship.
The advent of blogs and social networks has "emancipated authorship" - everyone has become a medium. At the time, the media industry saw this as an utopia, the creation of a new censorship-free world. There was a lot of enthusiasm for the idea of a "viral editor", where witnesses and the most competent experts will create content, the audience will check it out, and the demand will determine the rating. Editors were predicted to disappear as a profession, as they would only have to become moderators of user-generated content.

It also turned out that it has never been so easy to manipulate the audience as in the age of social media. Today, "filter bubbles" create different realities for different social, cultural or ideological groups. In these realities, fake news and people who can refute it may never meet each other.
The evolution of the Internet has gone through three stages. It was technically difficult to publish anything on the first websites. In the second stage, intuitive blogging platforms for text, video, audio content, etc. appeared. In the third phase, these platforms evolved into social networks that brought users together to interact with content and with each other: Facebook, YouTube, Twitter. The list of these brands is constantly updated, because progress is accelerating. New formats for old and new audiences are being tested all the time. In 2021, for example, Clubhouse became world-famous, and TikTok surpassed Facebook and Google as the world's most popular web domain. Messengers perform certain functions of social media. Some of them, such as Telegram, combine the advantages and threats of both. Platform algorithms affect the results and interfere with the reproducibility of research. Messengers are the "grey area" for researchers and fact-checkers, not least because end-to-end encryption aimed to protect privacy makes it almost impossible to track the emergence and spread of disinformation. A study found that 50% of Canadian respondents regularly received fake news in messengers. A balance is needed between privacy, personal security and countering disinformation.
Dr. Žiga Turk from the University of Ljubljana described five stages of the news process - creating, editing, publication, amplifying, and consuming. In the traditional model, the content was created mostly by professionals and the responsibility for the next three stages laid with media outlets. Today there is no single point of control, but most often all these stages are implemented on social media platforms. It gives them a powerful position to influence this process.
Services that allow even a child to create a fake photo or video will not surprise anyone. AI has learned to identify signs of editing or search for original content. The new technologies, called the weapons of information wars of the future, such as deepfakes, are coming into play today. Deepfake is a video where the face of one person is convincingly replaced by a face compiled by a computer from fragments of their photos, videos, and voices that are already on the Internet. They bring a lot of benefits by saving time on video creation, making online learning individual, etc. But at the same time, their potential to create fake news is frightening.
Text generators work in a similar way. Their potential is almost unlimited: they can serve as smart assistants, write texts and programs, process datasets etc. At the same time, in a few years, their ability to create disinformation is predicted to surpass that of any troll factory. In the summer of 2020, OpenAI introduced a language generator that complements and creates from scratch English texts no different from ones written by humans. According to a May 2021 report by the US Center for Security and Emerging Technology, the program generates convincing fake tweets and news stories that shift the tone and manipulate narratives, for example, for and against Donald Trump. The mass production of such content can affect public opinion. Such disinformation will require less time and effort, and the reach and effectiveness of campaigns will increase. That is why the authors of the technology do not open mass access to it.
The platforms, together with leading universities, are working on the deepfake detection software. Google has created a database of about 3,000 such fake videos, using various methods of their creation, to teach AI to identify them. Facebook and Adobe created similar databases, and new ones continue to appear. The job ahead is to teach algorithms to detect any AI-generated material, be it a deepfake video, a text or images. So far, most of these tools are still in development and are not ready for public use. Many platforms have banned the posting of deepfakes. But it is still quite easy for them to slip through the moderation systems.
Availability of apps for creating deepfakes for everyone who has a mobile phone is a matter of time. Responsible creators of such technologies embed special markers into their products to warn viewers that the video is being manipulated. But if it works with video, it may not work with text - at least as of today.
So, there is room for manipulation at each stage of the news process. While developers are working on new detection tools, the platforms also have enough tools to stop them.

At the stage of creating, it is important to destroy the business model of clickbait, which so far looks quite effective. To do this, it makes sense for large advertising services such as Google AdSense to stop advertising on fake news websites. Neural networks that help find the best audience for advertising can also avoid clickbait resources. In fact, they are already doing this, but so far with insufficient success.
At the stage of publication, community standards come into play. Banning malicious YouTube videos or Facebook posts is an effective way of regulation. Social networks, although in different ways, play the role of guardians when it comes to adult content, copyright infringement, incitement to hatred, and so on. The problem is the enormous amount of content and the lack of qualified personnel to check it. So, platforms rely on algorithms and user complaints. If there are too many, the content is blocked or demonetized. Say, YouTube deprives channels of the opportunity to make money on advertising. The war of complaints replaces the war against malicious content. Another extreme is bringing the watchdog function to the maximum.
And finally, the platforms play the role of amplifiers showing users some posts and hiding others. Facebook shows us less than 10% of all the content we subscribe to. Its visibility depends on an interest in a page, on a post, on a content format – for instance, images and videos are more popular than text, - on its "freshness", interactions, likes, visits and evaluation by the platform. Twitter ranks tweets by posting time and "user relevance". On TikTok, it depends on customer preferences. Many of these parameters can be "hacked” by bots. After all, algorithms are non-transparent. This gives platforms the freedom to choose which news to promote and which to hide. In addition, the algorithms may be imperfect. The dependence of the news feed on preferences only thickens the walls of "echo chambers" around each of us and makes dialogue almost impossible.

Algorithms are created by humans

All methods of combating fake news and manipulations are based on the question of who and how will determine them.

There are two strategies for automatic fake news detection: to base it on the content or the social context. The social context includes the source of information and the nature of its dissemination. The fact that a piece is published on a clickbait site indicates its unreliability. The inauthentic nature of the spreading by multiple accounts may raise suspicions about the use of bots to promote them. However, the algorithm can come to a dead-end if a reputable website posts a fake story by mistake. In turn, the analysis of news content is challenging for AI, as many manipulative news items contain humor, hints, morals, or metaphors, especially when it comes to politics. As a result, the tandem of human and machine intelligence works best. The neural network saves time by scanning huge amounts of content. The human is dealing with just a tiny part of the work that the AI cannot decide on itself.
It is a mistake to assume that algorithms are never wrong. AI is taught on a given set of data chosen by people who may have different intentions. The users have the right not to trust them to filter their newsfeed. Given this, a transparent method of counteraction disinformation is cooperation with independent fact-checkers for checking posts and marking fake news or manipulations. The system warns the user before he or she wants to share such content, but the decision is up to a user. In Ukraine, since 2020, Facebook has started a partnership with two organizations - StopFake and VoxCheck. It allowed marking Covid-related manipulative content, which decreased its visibility and the rating of the pages spreading them. Instagram, Google, TikTok also work together with fact-checkers. Experts say it is better to have several fact-checking teams on the market so that users can choose who they trust more.
The stage of consumption is the most important. As paradoxical as it may appear, the main fake news distributors are ordinary people who believe in them. By clicking the "share" button, each user becomes a middle link in the disinformation chain. If users become invulnerable to manipulations, the chain will be broken. What can each of us do about it? Choose trusted checker services. Choose browser extensions labeling suspicious content in the news feed, as well as photo, video, and text checking programs.

However, in the long run, I believe in education and media literacy, which will not be replaced by any AI. AI can make it easier by offering automatic fact-checking tools. But the foundation of a rational attitude to information should be laid starting with schools and kindergartens, as is done in many countries. In Ukraine, more than 1500 schools are already participating in the program of introducing media literacy in the school curriculum "Learn and Discern", implemented by IREX together with the Ministry of Education in partnership with the NGOs and experts.
Obviously, the fight against disinformation with the help of AI is worth it. But there is the need for caution and comprehensiveness. History shows that disinformation has always been and always will be. It does not stand still and often operates with more money than those who oppose it. The creators of disinformation and those who try to resist it are playing cat and mouse. Every successful means of countering causes further improvements in means of deception. The algorithm invented today may not detect tomorrow's fakes. So, there is no magic technological pill. But it is hoped that new technologies will work in combination with the responsibility of their creators, the platforms that establish fair and clear rules of the game, and governments that regulate the information field as much as necessary for national security, but no more.
These technological means also include cyber defense systems as part of a national cybersecurity system, giving users greater control over the algorithms that determine their news feeds, as well as empowering users to protect their personal data and privacy. Among them is such a simple step as explaining the rules of using online services and privacy policy in plain language. After all, now users often agree with them without reading a few pages of complex legal wording.
 There are several non-technological ways to limit the damage from disinformation: the already mentioned media literacy, support for quality journalism, diversity of the media landscape, balanced legislation, which will ensure the protection of information and other human rights in the digital environment etc. It is important to increase the transparency of algorithms, conducting their independent audit of compliance with democratic principles and human rights, as well as their openness to researchers. This is possible only under control by the society.
Prepared in February 2022, published in February 2023