Putting the brakes on online hate: cooling off
In this column, Merien ten Houten argues: a simple cooling-off period can stop most threatening messages without resorting to censorship.
Published on September 19, 2026
Merien co-founded E52 in 2015 and envisioned AI in journalism, leading to Laio. He writes bold columns on hydrogen and mobility—often with a sharp edge.
The courtroom is filled with people who regret what they’ve done. “It was a spur-of-the-moment thing,” “I didn’t really mean it”—excuses that come too late. Online threats are often impulsive, but the consequences for the victim can range from severe to disastrous. Still, we don’t want censorship. So how do we stop that impulse without infringing on freedom of speech? A sentiment model—an AI model that detects threatening messages—offers a solution. And it aligns with the Digital Services Act (DSA).
Impulsive hate with real consequences
Online hate is often a matter of seconds. An outburst of anger, a threatening comment, a message that later brings regret. Yet the consequences can linger for years: fear for the victim, legal repercussions for the perpetrator. The question isn’t whether we should do something, but how we do it without falling into the trap of mass surveillance or censorship.
The proposal is simple: analyze every message locally using a sentiment model. As soon as the score enters the threatening range, the message is held locally. The user can only send it after 30 to 60 minutes, following a second confirmation. This approach takes the sting out of impulsive threats without blocking the content.
It’s the digital equivalent of counting to ten before saying something you’ll regret—but enforced by technology. Because let’s be honest: anyone who, after 45 minutes, a walk to the coffee machine, and a deep sigh, still thinks that mayor “should be hung from the highest tree” is a deliberate extremist who knowingly breaks the law. But in most cases, common sense will prevail over adrenaline. The author will delete the post themselves, ashamed of their own rashness. And no one saw a thing—the whole process was local.
Legal framework: the DSA as a basis
The Digital Services Act (DSA) provides the necessary framework to implement this safeguard. Article 34 of the DSA requires very large platforms to identify systemic societal risks, including the dissemination of illegal content, hate speech, and threats to public debate. Article 35 then requires platforms to take targeted, proportionate measures to mitigate those risks. A cooling-off period is one way to fulfill this obligation, as long as it is not discriminatory or disproportionate. Because the process is local and delays rather than prohibits, many legal hurdles related to privacy are avoided.
Technically speaking, implementation is feasible. Sentiment analysis models, such as those from Meta and Google, already rate posts based on emotional intensity. Adding a cooling-off period is a small step. Moreover, these AI models are small—often only 15 to 50 megabytes—and can easily run on a modern smartphone. Precisely because the technology remains local and delays rather than censors, the intervention is strictly proportionate.
TikTok has already experimented with cooling-off periods for repeat offenders, which led to a 22% decrease in reports of hate speech in 2025.
Privacy and proportionality
The biggest challenge is balancing security and privacy. Under the ePrivacy Directive, scanning messages is only permitted if there is a legal exception, such as combating illegal content. The DSA provides this exception by requiring platforms to remove illegal and harmful content. Sentiment analysis is therefore permitted, provided it targets threatening or hateful content and not random content. In addition, platforms must comply with the GDPR: data minimization, purpose limitation, and transparency. Users must know how the system works and how they can appeal. The cooling-off period is proportionate because it does not remove content but merely enforces a pause.
The impact of this measure is twofold. First, it reduces the number of impulsive threats by giving users a moment to think. Second, it reinforces users’ responsibility. Currently, we often see people appearing in court claiming it was “a spur-of-the-moment decision.” A cooling-off period forces them to consciously choose their words.
This represents a shared responsibility between the user and the platform: the user chooses their words, and the platform provides the brake. But because tech companies dislike friction—which, after all, hinders interaction and ad revenue—external pressure is needed to get platforms to implement such a system.