Global Productivity Loss from Watching Reels and Its Cure
Published:
With the rapid development of learning-based technologies, the reel industry (YouTube, Facebook, Instagram, etc.) has taken the world into the booze of cocaine-like drugs. Even the best minds of a country are regular customers for hours every day. Let us first understand mathematically (in a very non-rigorous way) why reels are so addictive.
The Mathematical Framework
Consider the average person watching reels as the Audience, and there is an Agent (the drug lord) who has designed an algorithm $F(\cdot|H): \mathcal{Z} \to \mathcal{Y}$, where $H$ is history, $\mathcal{Z}$ is the set of design parameters, and $\mathcal{Y}$ is the set of strategies. Step by step, let us define the sequence of events that take place while watching reels:
Event Sequence
1. Content Generation
The Agent runs the secret algorithm and generates a particular reel at time \(t\) (say \(x_t\)) from a pool of finitely many reels (say, set \(\mathcal{X}\)).
2. Audience Interaction
After the audience watches it, they give likes/comments/swipes to the next reel without watching it in full. All of these actions of the audience are formulated in terms of rewards \(r_t(x_t)\) at time \(t\). For example, for a given \(x_t\):
- ‘Like’ \(: r_t(x_t) = 1\)
- ‘Swiping without watching’ \(: r_t(x_t) = 0\)
Note that the Agent always wants to maximize the reward.
3. Regret Formulation
We define the notion of regret (\(R_T\)) after $T$ rounds as:
\[R_T = \max_{x \in \mathcal{X}} \left(\sum_{t=1}^{T} r_t(x)\right) - \mathbb{E}\left[\sum_{t=1}^{T} r_t(x_t)\right]\]where:
- The first term is the best possible achievable cumulative reward (if the Agent could have known the reward function \(r_t(\cdot)\) beforehand every time; in other words, if the agent had known what the audience likes most before showing a reel)
- The second term is the actual realized cumulative reward after \(T\) rounds
The Addiction Mechanism
The goal of the Agent is to develop an algorithm where \(R_T\) grows sub-linearly with $T$, i.e., in the order of \(\mathcal{O}(\log T)\) (for example).
Note that if \(R_T\) grows sub-linearly with \(T\), it implies asymptotically that \(R_T\) is almost constant, which means every new \(x_t\), as given by the algorithm, is approximately the optimal one. When this happens, it means you like almost every reel you are watching and are under the arrest of an addiction.
The drug lords know the secret formula to design such algorithms that want to steal away your precious time by engaging you in reels.
Proposed Solution
With every technology, a grave amount of responsibility comes into the hands of the innovators. Therefore, each user must be given a hyper-parameter (say \(p\)), which will signify the time control for watching reels on any platform.
Implementation
Once a user sets \(p = 1\) hour (for example), the Agent behind the algorithm changes its strategy after 1 hour to maximize the Regret (\(R_T\)), which will eventually show such reels that the user starts to dislike and stay away from watching it further.
This simple modification transforms the optimization objective from:
\[\text{Current: } \min R_T \quad \text{(minimize regret = maximize engagement)}\]to:
\[\text{Proposed: } \max R_T \quad \text{(maximize regret = minimize engagement after time } p\text{)}\]Summary
By introducing user-controlled time limits that trigger a strategic shift in the recommendation algorithm, we can transform addictive engagement mechanisms into self-limiting systems. This approach respects user autonomy while addressing the societal problem of attention exploitation.
Copyright: All rights reserved.
Citation:
Mallick, A. (2025). Global Productivity Loss from Watching Reels and Its Cure.
Retrieved from https://amallick-tudelft.github.io/posts/2025/02/global-productivity-loss-reels/
What are your thoughts on this proposal? Share your ideas in the comments below.