In July, the European Commission issued a preliminary finding that Meta may have breached the Digital Services Act through the design of Facebook and Instagram. The investigation did not revolve around one illegal post. It focused on features such as infinite scroll, autoplay, push notifications and highly personalized recommendation systems, and on whether Meta had adequately assessed and mitigated risks associated with those features. Meta retains the right to respond, and the Commission’s finding is not a final ruling.
That distinction is becoming more important than the individual case. For much of the social-media era, the central regulatory argument was about content: which posts should be removed, how misinformation should be treated, when moderation becomes censorship and who should decide what users are permitted to say.
Regulators are now reaching further into the product itself. The question is increasingly not only what a platform contains, but how the platform decides what people see, how long they stay and how aggressively the service adapts to them.
A feed can shape behavior without containing illegal speech
A single video does not have to be unlawful for an algorithm to keep recommending increasingly similar material. A push notification contains no dangerous idea merely because it arrives late at night. Autoplay has no political viewpoint. Infinite scroll does not make a factual claim.
Yet those choices help determine what people encounter, how frequently they return and how easy it is to stop.
The European Commission’s July action is therefore significant because it treats product architecture as a potential source of systemic risk. California has moved in a similar direction through legislation signed in September that bars social-media platforms from offering users under sixteen features including autoplay and algorithmic feeds based on a user’s history and profile. Australia has chosen a more direct access rule: since December 2025, designated social-media platforms have been required to take reasonable steps to prevent Australians under sixteen from creating or keeping accounts.
These policies are not equivalent. One regulates features, another regulates access, and Europe’s DSA relies heavily on risk assessment and mitigation. What they share is a willingness to intervene before regulators can point to one prohibited piece of speech as the problem.
The object of regulation is moving from the message toward the mechanism.
Public opinion does not support a simple anti-social-media story
The political momentum can make it sound as if societies have reached a settled verdict that social media is harmful. A 37-country Pew Research Center survey published in October shows something more complicated.
Views have become significantly more negative in many countries since 2022 or 2023. In the United States, 64 percent of adults said social media had been bad for democracy; the corresponding figures were 70 percent in the Netherlands, 65 percent in France and 61 percent in Germany.
Across the 37 countries, however, the median view remained positive: 55 percent said social media had been good for democracy and 35 percent said it had been bad. The same survey captures the contradiction inside those judgments. A median of 77 percent said the internet and social media had made people more informed about current events, while 79 percent said they had made people easier to manipulate with false information and rumors.
This is a more useful starting point for regulation than either celebration or panic. Recommendation systems can help users discover information and communities they would never have found on their own. The same systems can also reinforce patterns of attention that regulators, researchers or users themselves regard as harmful.
Design is powerful partly because the same mechanism can produce very different outcomes.
AI companions make the design problem harder
Generative AI extends the argument beyond conventional social media because a chatbot does not merely distribute material posted by other users. It generates a new response during each interaction and can change that response according to what the user has said before.
California’s September legislation requires additional safeguards around companion chatbots used by children, including crisis protocols for suicidal ideation, parental controls, notifications when safety settings are disabled, independent child-safety audits and risk assessments. The state has therefore moved beyond deciding whether a specific chatbot response is permissible and toward regulating the conditions under which the relationship between user and system is designed.
That change matters. If a system repeatedly encourages dependence, adapts its tone to prolong conversation or carries sensitive personal history across interactions, the relevant risk may not be captured by identifying one prohibited sentence. Memory, personalization, anthropomorphic design and engagement incentives can become part of the regulatory object.
The content-moderation model was developed largely for a world in which users created the speech and platforms distributed it. AI makes the boundary between speaker, distributor and product much less stable.
The evidence for regulating design is stronger than some claims about harm
A move toward product regulation does not settle the scientific question of how much harm social media causes.
The National Academies’ 2024 consensus report on social media and adolescent health concluded that the available evidence did not support a population-level causal conclusion that social media was driving changes in adolescent health. The report documented risks for some users and benefits for others, and it emphasized the difficulty of separating cause from correlation. It also recommended more research access, greater transparency and closer attention to platform design.
That is an important limit. Governments should not treat a rise in adolescent distress and widespread social-media use as proof that every disputed feature caused the change.
The Australian age restriction illustrates the same problem from the policy side. The rule is now in force, but the long-term consequences for children’s wellbeing, social interaction and migration to other online services are still being evaluated. The fact that a government has acted is not evidence that the policy has already worked.
Design regulation is attractive because some mechanisms are easier to inspect than broad concepts such as misinformation or social harm. A regulator can determine whether autoplay is enabled, whether an account is private by default or whether a recommendation system offers users meaningful control. The causal case for changing those mechanisms still has to be made with evidence rather than assumed.
Product rules can influence speech even when they do not ban speech
There is also a civil-liberties problem that design regulation cannot escape.
A government does not have to order a platform to delete a political view in order to affect how widely that view travels. Rules governing recommendation systems, personalization and access can influence which users receive content and how frequently they encounter it.
Some interventions are narrow and comparatively transparent. Requiring a child’s account to be private by default is different from asking a platform to reduce distribution of a broad category of political material. Requiring an AI companion to disclose that it is artificial is different from telling it what political advice it may provide.
Putting all of those interventions under a single label such as “online safety” can hide important differences.
The challenge is to regulate mechanisms precisely enough that governments are not simply recreating the old speech dispute through a new technical vocabulary.
The next divide may be about mechanisms rather than messages
The first era of internet regulation was dominated by the question of what platforms should allow people to say. That question is not going away.
A second question is now becoming unavoidable: what kinds of systems should be allowed to decide how speech reaches people, how aggressively a service competes for attention and how deeply software adapts itself to an individual user?
Some design rules may prove more auditable than content judgments because they concern observable mechanisms. Regulators can test whether a minor’s account is public, whether an algorithmic feed is enabled by default, whether a company conducted a required risk assessment or whether a chatbot gives users the controls required by law.
That does not make design regulation neutral. It does give governments a possible path that does not begin with judging the truth or political value of every message.
Whether this becomes a better regulatory model will depend on discipline. If authorities identify specific mechanisms, produce evidence of plausible harm and impose proportionate requirements, product design may be a more tractable target than the endless policing of individual posts. If “design” becomes a broad justification for determining what information users should encounter, the old conflict over speech will return through a different door.
Sources
- European Commission — Preliminary DSA findings on the design of Instagram and Facebook
- California Governor — Child-safety chatbot and social-media laws, September 2026
- Australia eSafety Commissioner — Social media age restrictions
- Pew Research Center — Across the Globe, People Increasingly Say Social Media Is Harming Democracy
- National Academies — Social Media and Adolescent Health