Reddit uses LLMs to fight spam that LLMs helped create

Large language models have made it trivially easy for anyone to flood the internet with automated spam. A few keystrokes and a prompt can produce hundreds of convincing posts, comments, or messages. Platforms that once relied on simple pattern matching now face a much more sophisticated adversary. Reddit, one of the most active communities on the web, has decided to fight this new wave of spam with the same technology that helped create it.

The company recently detailed its use of LLMs to catch spam that older detection systems missed. Reddit says it now blocks an average of 23 million spam views per day and catches roughly 25,000 new spam posts and comments each day. These numbers highlight just how pervasive AI generated content has become on social platforms. The irony is not lost on anyone: LLMs are being used to clean up a mess that LLMs largely made.

How Reddit trains LLMs to spot coordinated spam

Traditional spam filters look for obvious signals like repeated text, suspicious links, or recently created accounts. But modern LLMs can produce content that looks human written, varies phrasing, and even mimics the tone of a specific subreddit. Reddit’s new tools are designed to detect what the company calls “highly subtle, coordinated patterns of fake behavior and artificial hype.” In a blog post, Reddit explained that these patterns often involve multiple accounts posting similar content at the same time, or accounts that gradually build karma before suddenly pushing a product or agenda.

The company claims that between January and March, users exposure to spam dropped by 20 percent compared with the previous three months. That improvement comes from LLMs that can analyze the context and behavior behind posts, not just the text itself. By looking at posting frequency, account age, engagement patterns, and semantic similarity, the system flags suspicious activity faster than rule based filters ever could.

Reddit has been refining its automated moderation for years, but the addition of LLMs represents a significant upgrade. The platform says these updated tools are catching spam at a higher rate than previous generations of software. The challenge, however, is that the same LLMs used for defense are also used by attackers. Bad actors constantly tweak their prompts to evade detection, forcing Reddit to update its models regularly.

The broader platform response to AI generated content

Reddit is not alone in grappling with this issue. Major platforms like YouTube, Meta, and Instagram now require users to disclose when content is AI generated. TikTok has gone a step further by allowing users to adjust a slider that controls how much AI generated content appears in their feed. These policies acknowledge that AI content is here to stay, but they also create a new problem: enforcing the disclosure rules.

If a platform can detect AI generated content quickly, it can also flag other types of rule breaking material such as hate speech or disinformation. The same technology that spots spam can be tuned to find toxic comments or coordinated harassment campaigns. But detection is only half the battle. Platform safety experts have repeatedly warned that automated moderation must be paired with human review to avoid false positives and to handle edge cases where nuance matters.

Reddit’s approach leans heavily on automation, but the company still employs human moderators and community moderators who work alongside the AI systems. The goal is to reduce the volume of spam that reaches human eyes, not to replace human judgment entirely. In practice, the LLM flags suspicious content, a human moderator reviews the most ambiguous cases, and the system learns from each decision.

The arms race between spammers and platforms is unlikely to end. As LLMs become cheaper and more capable, the cost of generating spam will approach zero. Reddit’s current success may be temporary if attackers find ways to bypass the new filters. The company is banking on continuous model updates, and it has the data advantage: every spam attempt that gets caught provides training material for the next iteration of the detection system.

Looking ahead, the same techniques Reddit uses to fight spam could eventually be applied to other moderation tasks. The ability to understand context and intent at scale is a powerful tool for any platform that deals with user generated content. But the technology is still young, and there will be growing pains. False flags, over moderation, and adversarial attacks will all test the limits of LLM based moderation.

In the broader conversation about artificial intelligence, the spam fight is a reminder that every new capability creates a new vulnerability. Reddit’s choice to use LLMs defensively is pragmatic, but it also raises questions about how platforms will balance automation with human oversight in the years ahead. For a deeper look at where these technologies might lead, see our analysis of AGI timelines and what leading researchers expect from the next generation of AI systems.