Reddit is turning large language models against the very problem those models helped create. The platform says it built new LLM-powered tools to catch spam and bot content, much of which is now generated by LLMs in the first place, according to TechCrunch AI. It’s a strange loop, but in the current AI landscape, platforms are fighting fire with fire.
Here’s what Reddit is claiming. The company now blocks 23 million spam views per day and catches roughly 25,000 new spam posts and comments daily. It says the upgraded tools cut users’ exposure to spam by 20% from January to March compared with the prior three months. TechCrunch AI reports that Reddit credits its models with spotting patterns older systems missed.
“We leverage LLMs to catch the highly subtle, coordinated patterns of fake behavior and artificial hype that older systems once missed,” the company wrote in a blog post.
Why this matters
Spam isn’t new, and neither are automated filters. Social platforms have been building them for years. What’s changed is the economics on the other side. Cheap, accessible LLMs mean bad actors can spin up convincing fake posts, coordinated hype, and comment floods at a scale and quality that keyword filters and older rule-based systems can’t keep up with.
That’s the real shift here. The old defenses were tuned to catch clumsy spam: repeated links, obvious templates, bursts from a single account. AI-generated spam doesn’t look like that. It varies its wording, mimics real conversation, and coordinates across accounts in ways that stay under the radar of pattern-matching built for a pre-LLM internet. Reddit is essentially admitting that to detect machine-made manipulation, you need a machine that understands language the same way.
The bigger play
What stands out to me is the second-order benefit Reddit hints at. If a platform can detect AI-generated content faster, it can also flag other violative content faster. Hate speech, coordinated harassment, and manipulation campaigns all share the same detection challenge: subtle, context-dependent signals that simple filters miss. Better language understanding on the spam side spills over into moderation broadly.
This also lines up with where the rest of the industry is heading, just from a different angle. TechCrunch AI notes that YouTube, Meta, and Instagram let users post AI-generated content as long as they disclose it. TikTok goes further, letting users toggle how much AI content they want to see. Those platforms are managing AI content out in the open. Reddit’s focus is the hidden version: the stuff that pretends to be human and isn’t disclosed at all.
The catch
Don’t read this as a solved problem. Platform experts keep making the same point, and it’s worth repeating: AI moderation works best paired with human moderation. Models are good at flagging suspicious patterns at scale. They’re worse at judgment calls, context, and the edge cases where a false positive silences a real user. A 20% reduction in spam exposure is meaningful, but it’s a percentage, not a finish line. The spammers have the same tools, and they’ll adapt.
That’s the uncomfortable truth of this arms race. Both sides are now armed with the same technology. Reddit gets to use LLMs to defend the platform, and the people flooding it with junk get to use LLMs to attack it. Whoever iterates faster wins the quarter, not the war.
What to watch
A few things worth tracking if you run a community, moderate content, or just care about platform health:
- Detection rates as a moving target. A 20% drop is a snapshot. Watch whether Reddit sustains or grows that number as spammers adjust.
- Spillover into moderation. If the same models start catching hate speech and coordinated harassment faster, that’s the bigger story.
- False positives. More aggressive automated detection means more legitimate posts caught in the net. How Reddit handles appeals will tell you how much it trusts its own models.
The short version: the internet’s spam problem got worse because AI got cheap, and now the same tech is the frontline defense. Reddit’s early numbers suggest it’s working, for now. You can find the full details at the original TechCrunch AI report.