The Internet Built a Spam Machine. Now It’s Building the Antibody.

Reddit’s AI now blocks 23 million spam views a day. But the real story is what that number reveals about the fight underneath it.

On July 6, 2026, Reddit published a number that should stop anyone paying attention to the internet: its automated systems now block roughly 23 million spam views a day before a single human sees them. Every 24 hours, the platform catches about 25,000 new spammy posts and comments, revokes nearly 2 million fake votes, and — for hate and violent content — moves from detection to enforcement in under five seconds.

Read that as a security bulletin and it sounds reassuring. Read it as a symptom and it tells a different story: the volume of synthetic manipulation flooding the open web has gotten so large that the only thing fast enough to fight it is the same technology that created it.

The number behind the number

Reddit didn’t build this system to fight ordinary spam. It built it to fight generative engine optimization — brands and bad actors seeding fake posts and comments specifically so that ChatGPT, Gemini, and other AI assistants will surface them as if they were genuine opinion. Reddit’s human-written, community-voted discussions have become a prized data source for AI chatbots, which makes the platform a high-value target for anyone who wants their product, candidate, or narrative laundered into an AI-generated answer that reads as neutral, crowd-sourced truth.

To catch that kind of manipulation, Reddit turned to large language models themselves, specifically because they’re better than older rule-based filters at recognizing subtle, coordinated patterns of fake behavior and manufactured hype. The company also screens new accounts at the moment of creation and now asks suspected automated accounts to verify they’re human. Over the first quarter of 2026, these measures cut user exposure to spam by 20% year over year, while cutting false positives by more than 40% — meaning the system got both more aggressive and more precise at the same time.

That’s the encouraging half of the story.

Why the fire got this big in the first place

Reddit’s spam problem didn’t appear in a vacuum. It’s a local symptom of a web-wide condition. Independent crawls give a sense of scale: an Ahrefs analysis of nearly a million pages published in April 2025 found that 74.2% contained detectable AI-generated content, and separate research from Graphite found more than half of newly published English-language articles were AI-generated or AI-assisted. Cloudflare has reported that bot traffic now makes up the majority of web requests, overtaking human traffic for the first time.

Some of that is harmless automation. A lot of it isn’t. Estimates suggest that a meaningful share of YouTube’s video feed now consists of low-effort, AI-generated “slop” content optimized purely to farm ad revenue, and mentions of the term “AI slop” itself grew roughly ninefold in a single year, according to Meltwater’s media tracking — enough that both Merriam-Webster and the Australian National Dictionary named it their word of the year for 2025.

This is the environment any trust-and-safety team is now operating in: a web where the cost of producing plausible-sounding content has collapsed to nearly zero, and the old heuristics for spotting fakes — bad grammar, repeated phrasing, obvious keyword stuffing — no longer work, because the fakes are often written by the same class of model the detectors are running on.

The deeper problem Reddit can’t fix alone

Here’s the detail that makes this more than a routine product update: Reddit’s announcement landed almost exactly seven weeks after Cornell Tech researchers published a preprint documenting just how exposed AI systems are to this kind of manipulation. The study, led by researcher Tingwei Zhang, found that planted text as short as thirteen words could steer AI “deep research” agents into recommending fabricated products, with success rates reaching 100% in some tested configurations. Zhang told reporters that these research agents tend to treat a random forum comment with roughly the same credibility as a government website.

That’s the structural crack underneath the encouraging spam-reduction numbers. Retrieval-augmented AI systems are designed to pull in and trust live web content, because that’s what makes them useful for anything current — news, product research, recommendations. But that same design means any platform they index is a viable attack surface for the exact seeding tactics Reddit is now fighting. Reddit can clean its own forums. It can’t clean Wikipedia, Quora, YouTube comments, or the thousands of smaller forums that AI systems also crawl and cite. Closing one window doesn’t secure the building.

Put plainly: the actual vulnerability isn’t that spammers can post fake content. That has always been possible. It’s that modern AI assistants read the fake content for you, extract a pattern from it, and hand you a confident, cited-sounding answer — without showing you the manipulation underneath. You never see the campaign. You just see the recommendation.

It isn’t only Reddit

The same arms race is playing out across e-commerce. Amazon has said it uses large language models alongside machine learning and behavioral analysis to catch fake reviews, incentivized ratings, and manipulated seller accounts before shoppers ever see them, and has reported stopping hundreds of millions of suspected fake reviews. Academic researchers are pushing the detection side further still: a May 2026 study out of the University of East London combined language-pattern analysis with behavioral signals — review timing, rating-to-tone mismatches, reviewer history — and reported catching fake reviews with 93% accuracy on Amazon and 91% on Yelp, outperforming older, purely text-based detectors.

The pattern repeats everywhere synthetic content meets a marketplace of trust: email security vendors now use AI classifiers to catch AI-written phishing that slips past older spam filters; researchers are racing to detect AI-generated peer reviews in academic publishing before AI-written science starts grading AI-written science. Wherever generative models made deception cheaper, defenders have reached for the same class of model to make detection cheaper too.

An arms race, not a cure

It’s tempting to read all this as a clean success story — AI got out of hand, so we built better AI to contain it, problem solved. That framing undersells what’s actually happening. Every system described here is playing defense against an attacker who gets faster and cheaper with every model upgrade. Reddit’s 20% year-over-year reduction in spam exposure is a real, measurable win. It is not the same thing as winning. The company attributes the jump in detected spam mostly to better tooling rather than a shrinking problem — the raw volume being thrown at the platform hasn’t necessarily gone down at all.

That’s the honest shape of an arms race: each side’s improvements buy a temporary edge, not a resolution. LLM-based detectors get better at spotting coordinated inauthentic behavior; spam operations adjust their prompts, vary their phrasing, and route around the new pattern within weeks. The Cornell Tech finding that thirteen words of planted text can steer an AI agent isn’t a bug that gets patched once — it’s a structural property of how retrieval-based AI systems currently decide what to trust, and it will keep producing new attack variants as fast as researchers document the old ones.

What this actually means if you build or rely on these systems

For platforms, the lesson is that spam defense can no longer be bolted on after the fact — Reddit’s gains came from screening accounts at creation and analyzing behavioral patterns in real time, not from filtering finished posts. For anyone building or deploying AI agents that browse and cite the open web, the lesson is sharper: source credibility can’t be an afterthought. An assistant that weighs a forum comment the same as a government filing isn’t a minor design quirk; it’s an open invitation. And for everyone reading AI-generated answers, it’s worth holding a little more skepticism than the confident tone of the response suggests is warranted — the fluency of an AI answer says nothing about whether the material underneath it was planted.

The optimistic version of this story is that AI is getting genuinely good at catching AI. The more useful version is quieter: AI is getting good at catching yesterday’s AI-generated spam, while today’s is already adapting. Reddit’s numbers are worth celebrating. They’re also worth reading as a scoreboard in a game that hasn’t ended — because the moment defenders stop improving, the attackers who never stopped will show up in the numbers again.

What’s your read — is AI-versus-AI detection actually closing the gap, or just raising the price of admission for spammers who can afford better models too? Curious what you’re seeing in your own corner of the internet.