On hackster.io, a trio of coders (Lucretia Field, Auxence Daillen, and ramon prat) have put together an interesting project they’ve called ‘Digital Poison‘, which is an attempt to influence or ‘infect’ the YouTube algorithm as it pertains to a specific WiFi network, to alter the recommended videos for every device using to said network. The trio lists the hardware and software they use, including the humble Raspberry Pi, as well as a step-by-step guide going into the intricacies of their methodology, so you can potentially recreate the experiment yourself.

The system is ‘discreetly hidden inside a bottle to remain unnoticed in public spaces’, which you can see in the image below, although we’re not sure we agree on the ‘discreet’ part.

It’s unclear exactly what videos were used in which combination, but having a look through the steps the project makers have listed, it seems it included at least the following:

• https://www.youtube.com/watch?v=XQ0B0cjq1-U
• https://www.youtube.com/watch?v=dQw4w9WgXcQ
• https://www.youtube.com/watch?v=3JZ_D3ELwOQ
• https://www.youtube.com/watch?v=VD6xJq8NguY

How do recommendation algorithms actually work on Social Media?

Getting your head around how all this works is far from simple, but if you’re after a bit of in-depth learning, we heartily recommend the essay ‘Understanding Social Media Recommendation Algorithms‘ by Arvind Narayanan on knightcoloumbia.org. To quote:

‘the primary objective of almost every recommendation algorithm on social media platforms is to rank the available content according to how likely it is that the user in question will engage with it.’

What constitutes ‘engagement’ changes between different platforms, with Facebook determining this through a weighted average of comments, likes, reactions, and reshares on a post; Twitter/X similarly combines all the different types of interaction from users on a Tweet; and YouTube (post 2012) primarily emphasises watch time i.e. it prioritizes videos it expects you to spend a longer time watching, based on your viewing history.

‘If a user sees a video in their recommendations and doesn’t click on it, the watch time is zero. If they click on it and hit the back button after a minute, the watch time is one minute. Before 2012, YouTube optimized for click-through rate instead, which led to clickbait thumbnails (such a sexualized imagery) becoming ubiquitous; hence the shift to watch time.’

So it would seem that the creators of the Digital Poison project would have been prioritizing the simulation of a viewer watching these videos all the way through in order to try an influence the algorithm.

If you want to check out some more interesting Raspberry Pi-based projects, check out this AI rock paper scissors game or an impressive Cyberpunk 2077-inspired dashboard.

About the Author

Aaron Ritchie

Aaron's laptop knowledge makes him the go-to guy on PC Guide. But he still finds time for features, deals and much more.