18 July 2026 · BotShark Team
AI-written replies aren't a novelty anymore. Here's what's actually changed.
The LLM-assisted account used to be the rare, cutting-edge bot pattern. It's not rare anymore, and the honest update is less about detection getting smarter and more about the sheer volume tipping over.
We wrote about the LLM-assisted account as one pattern among twelve — a reply that's a little too smooth, a little too consistent, probably generated rather than typed. Worth a standalone update, because the honest thing that's changed since isn't the sophistication of the pattern. It's the sheer volume of it.
The shift, in one number
The clearest evidence of the shift isn't a clever new detection technique — it's straightforward traffic measurement. The 2025 Imperva Bad Bot Report found that automated traffic surpassed human traffic on the internet for the first time in over a decade, with bots now accounting for 51% of all web activity, and specifically flagged AI and large language models as the reason simple bot development suddenly got accessible to people with no real technical background (Imperva, 2025 Bad Bot Report).
What used to require effort now requires none
The original post on this pattern was written when generating a plausible, on-topic reply at scale genuinely required either a person typing it or a much cruder template system that was comparatively easy to spot. That ceiling is functionally gone. Researchers tracking machine-generated text across social platforms have found it's become a measurable, non-trivial share of what people actually read day to day (arXiv, "Machines in the Crowd? Measuring the Footprint of Machine-Generated Text on Reddit"). Investigators tracking modern propaganda operations describe the same shift from the operator's side: campaigns increasingly run on what's been nicknamed "AI slop" — high-volume, low-effort generated content that's simply cheaper to produce than a human-written reply ever was (NBC News, "Some of the largest online propaganda campaigns are using 'AI slop'").
Why detection hasn't caught up at the same pace
Single-post AI-text detectors are a genuinely hard, unreliable problem, and that hasn't fundamentally changed just because generation got cheaper. What has gotten easier is exactly what the original post already pointed at — not judging one reply in isolation, but looking at the pattern across an account's entire reply history: is the tone suspiciously uniform, is the length and complexity oddly consistent, does it actually track the specific thing it's replying to.
What this means practically
The old advice still holds and is arguably more relevant now: treat unusually smooth, on-topic-but-generic replies as decoration rather than genuine opinion. What's changed is the base rate — given how much of total web traffic is now automated, it's worth adjusting your prior upward for how much of what you're reading in a busy reply section is generated.
Sources: