● These new risks have prompted a need for innovative technologies that can distinguish between truth and falsehood and improve the precision of investigators attempting to detect disinformation campaigns.
● Cutting-edge analysis tools powered by machine learning and neural networks have the capacity to detect content generated by artificial intelligence.
“Decoding of information is benefiting more and more from technological advances.”
Hot on the heels of deep fake video that gives the illusion of real footage, artificial intelligence has now brought us audio tools that can convincingly imitate individual human voices when trained with a few soundbites. Entire telephone conversations and radio messages can be faked adding to the already formidable arsenal at the disposal of the creators of fake news, and disinformation campaigns that are often perceived as genuine on social networks. When they are viewed by large numbers of Internet users, fake news and disinformation can generate very real social tension, influence voter preferences and do real damage to corporate and personal reputations. According to a recent report by the American think tank, Brookings, deep fakes now represent a significant geopolitical risk in as much as they can utilized in military and subversive intelligence operations. Worse still, the researchers who authored this report warn of the ease and rapidity with which these deep fakes can now be produced.
Algorithms will soon produce content that is indistinguishable from that produced by humans
The AI arms race
A technological arms race is now underway between the creators of fake news and the developers of systems that can detect and flag mendacious content to audiences that may be exposed to it. New tools based on advanced language models like Chat GPT-3, or applications like FaceSwap and DeepFaceLab, have raised many questions. According to the World Economic Forum, “algorithms will soon produce content that is indistinguishable from that produced by humans”. For Georgetown University researcher, Josh A. Goldstein, there is also a risk that artificial intelligence will be used to produce fake content that is tailormade for individual Internet users, which could, for example, take into account data sourced from their social networks, making it all the more believable. Artificial intelligence also has the capacity to cut the cost of producing fake news, which could pave the way for disinformation on a massive scale.
In Madrid, the media company Newtral has gone as far as deploying its own AI language model, christened ClaimHunter, which has been trained to analyse and fact-check political speeches (especially with regard to data and figures) pronounced by candidates; an effort that has proved successful. Newtral is now working with the London School of Economics and the TV channel ABC Australia to develop a further version of its tool to distinguish falsehood from truth in political discourse.
The need to improve fact-checking
‘The race between fact-checkers and those they are checking on is an unequal one”, points out Tim Gordon in “Wired”, the co-founder of the Best Practice AI consultancy. For Gordon, the generative capacity of AI will make fact-checking, often undertaken by small organisations, well-nigh impossible. At the same time, however, the decoding of information has also come on apace with technology. Since 2018, researchers from the Computer Science and Artificial Intelligence Lab (Massachusetts Institute of Technology) and the Qatar Computing Research Institute thus claim that the best approach against fake news is to look at the actual sources rather than isolated news items. They have developed a machine-learning-based system to detect if a source is pertinent or biased.
In France also solutions are emerging to attempt to contain the fake news problem. Within the frame of the Content Check project started in 2016, four research laboratories and media such as Le Monde are working together to develop software aimed at journalists in order to check facts.
Ioana Manolescu, a computer science researcher at the French National Institute for Research in Digital Science and Technology (Inria), is one of the pioneers of Content Check. “My starting point was the observation that with the development of open data, everyone has access to a lot of information, the researcher confides to Farid Gueham, of the Fondation pour l’Innovation Politique. But this information is widely spread and not always easy to access: it is very complicated to link it all together.”
The team is working for example on software that improves the accessibility of Insee (The French National Institute of Statistics and Economic Studies) data. A crawler (an indexing robot) analyses the website, the data is then extracted thanks to an API and is consolidated in a database by an algorithm that identifies the type of each cell. The software makes it possible to answer a journalist’s research query by returning a value and a link to the original table.
Artificial intelligence and neural networks
At the Institut de recherche en informatique et systèmes aléatoires (Research institute of computer science and random systems), Vincent Claveau, French National centre for scientific research (CNRS) research fellow specialising in natural language processing, is concentrating on fake videos that are circulating on social networks.
Often modified and compressed several times, the content is analysed so as to detect if there are similar images on the web. “A neural network is trained to identify them, by comparing vector representations”, the researcher confides to Industrie & Technologies magazine. Calculation of the difference between the two images then makes it possible to highlight modified areas and to identify the changes made.
His team is also starting to work on image decontextualisation, analysing the characteristics of images and their associated text, again thanks to deep learning. The arms race continues more than ever.