TikTok is cutting hundreds of trust and safety jobs across Singapore, Indonesia, and Dublin, and handing the work to AI. The reaction has been predictable. Moderators and unions say the models make too many mistakes. TikTok points to its numbers: 97 percent of harmful content already taken down by AI this year, 99 percent of it before any user saw it.
Both sides are arguing about who should run the filter. Nobody is asking why the filter is the whole safety strategy.
The philosophy behind the statistic
Look at TikTok's own statistic again. Before a single layoff, machines were already catching 97 percent of the garbage. The humans losing their jobs this month were the residue of an automation that mostly happened years ago. This isn't a change in philosophy. It's the completion of one, and the philosophy is this: accept everything from everyone, then judge it on the way out.
That's reactive filtering, and every major platform runs on it. Accounts are free. Identity is disposable. Posting costs nothing and banning costs the banned party nothing, because the next account is thirty seconds away. When your inputs are unbounded and consequence-free, the only lever you have left is inspection at the output. So you build a firehose and then you build a bigger nozzle, forever.
Two races, one technology
For a while, that arms race was at least asymmetric in the platform's favor. Producing garbage at scale took effort. Not anymore. The same class of models now generates content and moderates it. The filter and the flood are the same technology, racing itself, and the flood side has no false-positive problem to worry about. A moderation model that wrongly removes real speech creates legal and PR damage. A spam model that wrongly produces a dud post creates nothing. One side of this race pays for its mistakes. Guess which one improves faster.
One side of this race pays for its mistakes. Guess which one improves faster.
A dead architecture, not a strained one
This is why I'd call the architecture dead rather than merely strained. Filtering output never changes the incentive at the input. Every removed video, whether a human or a model removed it, teaches the producer nothing except which patterns to vary. The account that posted it was worthless before the ban and worthless after. There is no memory in the system, so there is no cost, so the volume never goes down. TikTok can hit 99 percent and the remaining 1 percent of an infinite number is still infinite.
TikTok can hit 99 percent and the remaining 1 percent of an infinite number is still infinite.
Consequence, not a better filter
The alternative layer is not a better filter. It's consequence.
That's the premise behind Chirpper, an invite-only network I'm building, so weigh my argument with that in mind. The design is simple to state: every account is vouched in by an existing member, and reputation flows both ways through that chain.
That structure changes what posting costs in the way that matters. Garbage costs the poster standing they can't recreate with a new email address, and it costs something to the people who brought them in. Coordinated abuse gets worse odds still, because the accounts are related through their lineage, and when one surfaces, the cluster surfaces with it. The economics that make industrial-scale garbage rational simply don't survive contact with persistent, propagating accountability.
What this doesn't fix
I'll state the limits just as plainly: consequence at the input does not eliminate judgment at the output. Real people in good standing post things that need a human call. What it changes is the volume arriving at the filter and the incentive to flood it, which is the difference between moderation as a judgment function and moderation as a war of attrition.
The difference between moderation as a judgment function and moderation as a war of attrition.
Where this ends
TikTok's models will probably become very good filters. It won't matter. They are a better nozzle on a firehose the architecture itself guarantees. The platforms that eventually feel safe won't be the ones that judged the flood fastest. They'll be the ones that made flooding a bad idea.