What Does the Algorithm Reward? Our Minds Between Recommendation and Attention

In a book published in 1957, Ferster and Skinner studied how pigeons' behaviour changes with different schedules of reinforcement, that is, when the reward arrives after the act. From those laboratories many have borrowed an analogy for platform design: a notification that may or may not bring something new. Is the analogy accurate?

A popular analogy, narrower evidence

Critics of platform design invoke the logic of "irregular reward" to describe pull-to-refresh and notifications. But it is an analogy, not a direct experimental result on modern platforms in the sources we consulted; Ferster and Skinner's book studied pigeons, and its publishers say its principles apply equally to other species, humans included, a claim to be tested rather than assumed. Safer to look at what platforms actually measure.

What does the algorithm reward?

Its working yardstick is engagement: a click, a pause, a reshare. In 2023 Smitha Milli and colleagues compared engagement-based ranking on Twitter with reverse-chronological ranking, and found that the former amplified emotionally charged content hostile to the opposing political group, and that participants did not prefer the political tweets the algorithm chose. When content was ranked by what users say they want, angry content fell, though that may reinforce content that agrees with their views. No ranking is free of cost; what matters is knowing what is being measured.

Separately, the study by Soroush Vosoughi, Deb Roy and Sinan Aral in Science in 2018 examined about 126,000 cascades of fact-checked stories on Twitter between 2006 and 2017, shared by about three million people. False news was about 70% more likely to be retweeted than true news and spread more widely. The researchers explained this by novelty and emotion, and found that humans spread it more than bots. This reminds us that responsibility does not rest on the machine alone: we are partners in what is rewarded.

Memory shifts as well. In 2011 Betsy Sparrow and colleagues found that people who expect information to stay available remember where it is better than what it says. That is a shift in what we store, not necessarily a loss of understanding, but it obliges us to distinguish what we want to retain from what we are content merely to reach.

Not all tools are alike

In a randomised trial published in JMIR Mental Health in 2017, 34 university students received cognitive behavioural therapy from the chatbot Woebot for two weeks, against 36 students who read an information booklet; depression scores fell in the chatbot group alone, and no difference between the groups appeared in anxiety. But the sample is small, the period short, and one of the authors founded the company. It is a promising signal, not a final verdict.

And in our own country

The "Digital 2026" report (DataReportal) recorded about 38.6 million social-media user identities in Saudi Arabia in October 2025, equal to 111% of the population, and the identities may not represent distinct people. The figure shows breadth of presence, not the amount of harm, but it makes digital literacy a public matter.

What do we do?

• Arrange your environment, not your reflexes: disable unnecessary notifications and set times for browsing.
• Ask before engaging with what provokes you: why did this appear to me now? Who benefits from my anger?
• Diversify your sources deliberately and keep hours for reading and walking without a screen.
• Teach the young how recommendation works, as we teach them safety rules.

Moderation here is part of preserving a blessing; time and attention are trusts we shall be asked about, and this does not mean abandoning technology but using it well.