There鈥檚 no doubt that recent demonstrations of how machine learning can generate believable fakes are striking. From to, it is natural to worry that these techniques will make it much easier to manipulate political discourse and public opinion.
But, it鈥檚 worth taking a step back. Just because a technological capability exists does not mean that it will be widely used, or that it will make a serious impact. In this case, there are good reasons to believe that聽鈥� for the near future聽鈥� such 鈥渄eepfakes鈥� enabled by artificial intelligence will see limited use, and have limited impact.
To that end, I鈥檓 taking part in a and betting that by the end of 2018, we will not have seen a political hoax generated by machine learning acquire more than two million views before being discovered. Given that this is a year in which US voters go to the polls again for their mid-term elections, when we might expect advanced disinformation campaigns will be highly likely, why do I think I鈥檓 backing the right horse?
Advertisement
Fakes on a budget
The economics of fakery are important here: state and non-state actors engaging in online propaganda want to achieve the most influence at the lowest cost. The use of the latest developments in artificial intelligence must be compared against other available techniques for generating believable hoaxes.
We need to recognise that rudimentary techniques can already have a major influence on public discourse. attest to the fact that simply reusing an old video and asserting that it is something that it is not can be sufficient to fool many people. job can be enough to spread disinformation as well.
At the same time, AI remains a relatively costly tool for generating hoaxes. The current breakthroughs in AI rely on two major inputs, large datasets and computational power, both of which can be expensive to acquire in order to pull off a high-fidelity fake. Moreover, the techniques behind the more impressive demonstrations of AI-driven fabrication, known as generative adversarial networks (GANs), are famously temperamental. If improperly configured, GANs often exhibit a phenomenon known as, in which these systems fail to generate desired outcomes.
Deep diversion
This all suggests that, in the near term, AI-constructed fakes will be inferior to the existing arsenal of tools available to the propagandist.
This gives us perhaps more reasons to say there isn鈥檛 a unique and inevitable threat posed by deepfake techniques. For one, it gives time to create effective methods for detecting this type of fakery. For another, it enables society to adjust to the knowledge that everyday video can be faked in this way, and to be on guard before it sees widespread deployment.
Ultimately, it is possible that deepfakes are a diversion, something that portends a great threat but actually distracts us from more crucial questions. It is possible that we may eventually see AI driven hoaxes become an attractive option for the technology-minded propagandist. But, it will be the draw of the underlying narratives pushed by these fakes that determine their actual influence. Better understanding the behavioural factors driving belief in hoax narratives will be the only way to avoid playing an unwinnable game of technological whack-a-mole.
