Algorithmic ranking: the limits and meaning of user control
In a long post titled ‘You and the Algorithm: It Takes Two to Tango”, Nick Clegg laid out this week his vision for the future of algorithmic ranking. Similar to how algorithms constitute the ‘recipe’ for what we see every day in our Feeds- Clegg argues people have to know what goes into the recipe and be able to exercise control.
The framing of ‘user control’ in the context of feed ranking is worth unpacking.
A few patterns are pointing to the limited benefit — for both people and society — of full user control.
First, there is the need to protect people from harm whether they as individuals agree with the rules or not. Airplane passengers have very limited control over what goes on in the plane, for the safety of the other passengers and the crew. In the context of Feed ranking, there are types of harmful content (for example terrorist content) that users simply should not have an option to see more or less of- for the good of the community at large. Where the line is drawn between controllable and non controllable topics will depend on each platform. Google lets users decide whether adult content should be filtered out of their Search results through the ‘Safe Search’ control whereas Twitter does not provide that option.
Second, there is the inherently complex nature of modern technologies and the unconscious cognitive tradeoffs we make as humans to understand them based on their utility. As humans we interact day to day with complex technologies — from Wifi to GPS — which we have grown accustomed to despite not truly understanding them. In these cases, transparency or control do not matter so much as a “it works for me” and the theoretical ability (theoretical because we could not really live without Wifi or electricity- just like we couldn’t live without ranking) to turn things off if needed. We accept to relinquish full control because in the cognitive tradeoff between fully understanding the technology and reaping its utility without understanding it, the latter wins.
What is the meaning of user control?
First, the end goal of introducing controls should be for people to feel- and rightly so- in control. If Netflix introduced a dozen new controls to allow users to filter in and out specific types of results from search results- would users feel in control or simply overwhelmed? As controls are introduced, we have to consider whether the person we’re trying to empower is actually feeling empowered i.e. whether their sense of agency is actually increasing as a result of the new controls.
Once you consider the end goal of giving users agency, different interventions come to mind besides straight up ‘Controls’. Think about hiring a chef, or ordering a Uber. Do you need to take the wheel and impose directions or a particular way of cooking? Or do you really need to provide feedback and see that your feedback has been implemented? Thinking of feedback and preference as part of ‘agency’ opens new ways of empowering users- so long as avenues for feedback are deterministic and users know where to find them and use them consistently.
Finally, let’s consider trust. How much of our feeling of being in control is unlocked by trust? We as humans feel in control when it comes to our Wifi, our planes and our Ubers partly because we trust these technologies to work for us and for society as a whole. In other words, the technology itself has to be demonstrably trustworthy and safe (or at least neutral) before feelings of agency can develop. In the case of social media platforms, all the ‘controls’ or all the ‘user agency’ in the world will not matter if at the end of the day, people believe the underlying machinery consistently gears towards bad outcomes for people and society.
To recap, user control is a good vision for the future of algorithmic ranking- as long as a judgement call is made regarding which parts of the algorithms users will be able to control and how these controls will make people feel. There are other ways to create a sense of control than providing direct control. To start with, the algorithm has to be seen as trustworthy.
