Reddit has always been the strange one among the big platforms. Facebook and YouTube run enormous paid trust and safety operations. Reddit runs on volunteers, user reports, and a bot called Automod that follows keyword rules many of which were written years ago by someone who has not logged in since. That arrangement is now getting its most significant change in a decade, and the new hire is a language model.
What Automod Could Never Do
To understand why Reddit built this, it helps to understand the specific way the old system fails.
Automod is a pattern matcher. You give it words, phrases, regular expressions, account age thresholds, karma minimums, and it acts when it sees a match. That works beautifully for a narrow class of problems: banned links, obvious slurs, brand new accounts posting affiliate spam. It works terribly for everything that requires reading a sentence and understanding what it is doing.
A rule like “no personal attacks” cannot be expressed as a keyword list. Neither can “no low effort posts,” “no reposts from the last 30 days,” “posts must include a flair and a source,” or the classic “be civil.” Moderators have spent years writing increasingly baroque Automod configurations to approximate rules that are fundamentally about meaning, and then spending their evenings cleaning up what slipped through.
Rules Hub is aimed squarely at that gap. Instead of asking whether a post contains a string, it asks a model whether the post matches the intent of a rule as written. That is a genuinely different capability, and for the thousands of small and mid sized subreddits with one or two active mods, it is the difference between rules that are enforced and rules that exist on paper.
How Much Power Does It Actually Have?
This is the part that determines whether the backlash is proportionate. The short version: mods decide, and the ceiling is high.
| Who or what | Decides |
|---|---|
| Moderators | Write the rules, choose which ones the AI enforces, and set the consequence |
| Rules Hub AI | Judges whether each post or comment matches the intent of an enabled rule |
| Outcome option 1 | Send to the mod queue for a human to review |
| Outcome option 2 | Filter the content pending review |
| Outcome option 3 | Remove it automatically, with no human in the loop |
| Safety rails | Mods can test a rule against old posts before enabling it, and inspect logs explaining each action |
Option three is where the argument lives. Reddit is technically correct that humans remain in charge, because a human chose the setting. But once that toggle is flipped, the first entity to read your post and rule on whether it was a joke or an insult is a model, and there is no guarantee anyone will ever look at the decision. On a busy subreddit, the practical difference between “a human approved this policy” and “a human reviewed this removal” is enormous.
The dry run feature is the most underrated part of the design. Being able to point a proposed rule at a few months of archived posts and see what it would have caught is exactly the check that Automod never offered, and it is the thing most likely to stop a badly worded rule from quietly nuking half a community’s content.
Why the Reaction Has Been Rough
The response from users has been almost uniformly hostile, and it is worth separating the reflex from the substance.
The reflex is straightforward AI fatigue. Reddit has licensed its content for AI training, models trained on Reddit conversations now answer questions that used to send people to Reddit, and the site’s own search and summary features increasingly put a model between users and the posts. Adding an AI that can delete your comment lands on an audience that already feels its contributions are being fed into a machine. One comment doing the rounds, asking whether this is “a bot to remove bot content,” captures the mood precisely.
The substance is more specific and harder to dismiss. Moderators of large subreddits have raised the false positive problem: a model that is 98 percent accurate sounds excellent until it is applied to a subreddit processing 50,000 comments a day, at which point it is wrongly removing a thousand of them. Reddit’s appeal path for an automated removal is thin, and the people who would handle appeals are the same overloaded volunteers the tool was built to relieve.
There is also a trust question that has nothing to do with accuracy. Moderation on Reddit derives its legitimacy from the fact that a person in your community made a call and can be argued with. Replace that with a model and the removal becomes an act of infrastructure. Communities tolerate a great deal from moderators they can shout at. They tolerate much less from a system that cannot be reasoned with.
Automod’s Long Goodbye
Reddit has not announced that Automod is being retired, but it has said Rules Hub could eventually take over many of the enforcement jobs Automod handles today, particularly when combined with tools like Post and Comment Guidance and Safety Filters. That is about as clear a signal of direction as a platform gives.
The company has been notably careful in its framing, and the caveat it keeps repeating is the honest one: this is not ready for large or complicated communities. Starting with brand new subreddits is a sensible way to test that. New communities have no established norms to violate, small volumes to get wrong, and moderators who never learned Automod’s syntax in the first place and have nothing to lose by trying something else.
| Stage | Status |
|---|---|
| Pilot | Ran for months with moderators from 700 plus communities, including the Mod Council Network |
| Now | Available to all newly created subreddits |
| Later in 2026 | Wider rollout to existing communities, opt in |
| Long term | Positioned to absorb much of what Automod does today |
The Part Nobody Is Talking About
Automated moderation is not new. YouTube has run it at scale for years, and creators have spent that decade complaining about opaque removals with no meaningful recourse. What is new here is that the enforcement is being handed to the volunteers, which means every subreddit becomes its own experiment with different settings, different tolerance for error, and different willingness to check the model’s work.
That decentralization cuts both ways. A well run community gets a tool that finally enforces the rules it already had. A badly run one gets an automated removal machine pointed at whatever its moderators dislike this month, running at a speed no human could match. Reddit has effectively distributed a moderation capability with real power and very little central oversight of how it is configured.
It also arrives while regulators are paying close attention to exactly this kind of automated decision making. The EU’s transparency rules are now enforceable, and systems that automatically act on user content sit uncomfortably close to the disclosure obligations those rules create. And the broader capability question is not theoretical either: recent government testing showed how convincingly AI agents can already construct false identities online, which is a reminder that the same technology is on both sides of the moderation problem.
What to Do If You Post on Reddit
For most users, nothing changes today. Rules Hub is only on new subreddits, and the communities you already read are still running the same Automod configuration they were last week.
What is worth watching is the mod log. If a post disappears without explanation in a newer community, the answer may now be a model rather than a moderator, and asking politely in modmail is still the only route to a human. If you moderate, the dry run tool is the single feature to use before anything else. Test a rule against a year of your own archive, look at what it would have removed, and only then decide whether you trust it enough to let it act on its own.
Reddit is betting that better enforcement is worth some wrongful removals. Its moderators, who will absorb the complaints either way, are not yet convinced. Both of those positions can be right at the same time, which is usually a sign that the rollout speed matters more than the technology.
