Give TikTok a blank slate and a little time, and it doesn’t hand you a balanced sampler platter of the internet. It hands you a conveyor belt. And if you keep watching, that belt can drift, quietly, predictably, toward conspiracy junk and radical content.
That’s the takeaway from an experiment byOuest-France, a major regional French newspaper, which turned a simple question into a brutal stress test: what happens if you let an AI “user” scroll TikTok for 100 hours with a neutral profile?
In under four days of nonstop scrolling, the bot’s feed slid from harmless entertainment into a steady stream of extremist-adjacent material, exactly the kind of “how did I end up here?” rabbit hole real users describe all the time.
A “normal user” bot, and a feed that didn’t stay normal for long
Ouest-Francedidn’t build some cartoon villain account that only searches for propaganda. The bot was designed to behave like an average TikTok user: scrolling at a human pace, pausing now and then, and interacting lightly, enough to look real, not enough to “train” the system with obvious preferences.
At first, TikTok did what TikTok does: a grab bag of viral dances, comedy clips, and creative tutorials. That’s the platform’s standard “new user” routine, throw a bunch of categories at you and see what sticks.
Then came the turn. Around the20th hour, the bot started getting served videos with clickbait titles that cast doubt on established facts. And here’s the part that should make every parent, policymaker, and platform exec sweat: the bot didn’t have to like, share, or comment. It just had to watch a beat longer.
TikTok’s recommendation engine treated slightly longer watch time as a green light, an “engagement” signal, and kept pushing in that direction. That’s how the slide happens: not with a hard right turn, but with a thousand tiny nudges.
This isn’t some one-off French freakout, either. TheCenter for Countering Digital Hatereported similar dynamics in 2025 on other major platforms, including YouTube and Instagram: recommendation systems tend to escalate toward more polarizing content because polarizing content holds attention.
Europe’s TikTok problem: 150 million users and a firehose of data
TikTok says it has more than150 million monthly active users in Europe. A lot of them are teenagers, prime targets for misinformation, identity bait, and algorithmic manipulation dressed up as “content.”
And TikTok isn’t guessing. The app collects mountains of behavioral data every day: how long you watch each video, what you replay, what you skip, what you share, when you log on, and where you are (via location signals and device data). Feed that into machine learning, and the system gets better, relentlessly, at keeping you glued to the screen.
TikTok CEOShou Zi Chewtold the European Parliament in September 2025 that the company’s moderation is robust, citing40,000 moderators worldwideand billions invested in AI tools to detect harmful content.
But theOuest-Francebot experiment suggests moderation and recommendation are two different beasts. You can delete the worst stuff and still have an algorithm that “walks” users to the edge, serving borderline content that normalizes the next, more extreme step.
Europe is tightening the screws, and TikTok is in the crosshairs
The European Union’sDigital Services Act(fully in force since 2024) forces the biggest platforms to be more transparent about how their algorithms work and to publish quarterly reports on efforts to limit illegal or harmful content.
In October 2025,Thierry Breton, the EU’s internal market commissioner and one of Big Tech’s most aggressive regulators, announced a risk assessment aimed squarely at recommendation algorithms. Child-protection groups across multiple EU countries have been sounding alarms, and Brussels is listening.
The technical conflict is baked in. These systems are built to optimize for engagement, mostly measured by time spent. That goal doesn’t naturally align with “balanced information” or “verified facts.” It aligns with whatever keeps your thumb scrolling.
Meta, for its part, has experimented with “de-optimized” recommendation approaches on Instagram, favoring content variety over pure engagement. Meta reported in November 2025 that this led to a12% dropin average time spent, while user satisfaction (measured by surveys) improved. Translation: the addictive stuff works, and turning it down costs money.
The real fight is the business model: attention equals cash
TheOuest-Franceexperiment lands on the uncomfortable truth everyone in Silicon Valley knows and pretends not to: the ad model rewards whatever keeps people online longer. More minutes equals more ads equals more revenue. That incentive doesn’t just tolerate toxic spirals, it can quietly profit from them.
European lawmakers are pushing back. French lawmakerLaure de La Raudièrefiled an amendment in December 2025 that would ban recommendation systems for minors that are based solely on engagement optimization.
TikTok’s response to theOuest-Francereporting: a “balanced browsing mode” designed to limit recommendations of similar content. It’s still being tested in five European countries. And if regulators decide the current system is a public-health hazard for kids, “optional mode” may not cut it for long.
