Edition 2 · 06 / 08
Does Your Algorithm Make You More Like Yesterday?
Feedback loops can make inferred preferences increasingly influential without proving that the person underneath has changed.
A recommender observes behaviour, selects what to show and then observes behaviour inside the environment it helped create. This feedback structure is real. Its effect on political belief is more difficult to establish.
Preference or opportunity?
A click can reflect interest, prominence, curiosity, irritation or the absence of an attractive alternative. An item that was not shown could not be chosen. Recorded behaviour is therefore conditional on exposure rather than a clean reading of preference.
When later recommendations rely on that behaviour, an early signal can become increasingly influential. The person encounters more material consistent with the inference and gets fewer opportunities to provide contrary evidence. This can narrow an exposure trajectory even if their underlying politics remain unchanged.
The evidence is mixed
Research can demonstrate that feeds alter the balance of material people see. Moving from exposure to belief requires a stronger design.
A large 2020 Facebook experiment reduced exposure to politically like minded sources by about one third. The intervention changed what participants saw but produced no measurable change across eight preregistered measures including ideological extremity, affective polarisation and belief in false claims.
Other findings indicate that feed design can affect some political opinions. A seven week field experiment on X compared algorithmic and chronological feeds. The algorithmic feed increased engagement and shifted several political attitudes in a conservative direction, while the study found no significant effect on affective polarisation or self reported partisanship.
These results are not contradictory. Effects can depend on the platform, period, content supply, users, outcome and duration being measured. They show why “the algorithm causes polarisation” is too broad to function as a research conclusion.
A system can change politics without converting everyone
Recommendation may still influence public life without changing each person’s ideology. It can change:
- which subjects dominate attention;
- which political actors acquire reach;
- which language appears normal within a group;
- how rapidly an allegation travels;
- and what journalists and politicians perceive the public to be discussing.
These distributional effects can reward political strategies built for the ranking environment. Actors learn which messages receive attention and produce more of them. The system shapes the incentives around speech as well as the sequence received by an individual.
Make the claim testable
Three claims should remain separate:
Strong: the ranking changed the items and order shown.
Testable: repeated ranking altered the diversity or intensity of later exposure.
Requires causal evidence: the experience changed an enduring belief, vote or hostility towards another group.
That distinction does not diminish the concern. It turns a sweeping accusation into questions that researchers, platforms and regulators can answer.
References
- Guess and colleagues: Like-minded sources on Facebook are prevalent but not polarising, 2023.
- The political effects of X’s feed algorithm, 2026.
- Systematic partisan content skews in TikTok during the 2024 US elections, 2026. The audit measures exposure rather than downstream attitudes.