1.2 Million School AI Chats: The Real Risk Isn't Cheating
An analysis of 1.2 million US student AI chats found safety-flagged prompts about self-harm and violence far outpace unsafe web searches, while most school policy still fixates on cheating.
What the data actually shows
Most school debate about student AI use still circles the same question: are they cheating? A new dataset suggests that framing is missing the bigger risk.
Securly, a US school safety and filtering vendor, analysed nearly 1.2 million student AI conversations logged on school-issued devices across 1,312 districts in 39 states between 1 December 2025 and 20 February 2026, according to reporting in K-12 Dive and EdWeek. Around 117,000 individual students generated that traffic. This is not a survey asking students to self-report. It is telemetry from actual school devices, which makes it one of the first real-time pictures of what pupils do with AI chatbots during the school day rather than what they say they do.
ChatGPT accounted for 42 percent of the traffic, Google's Gemini 21 percent, with the remainder split across a growing field of AI features embedded inside other classroom tools. In districts that had actually set usage guardrails, 80 percent of conversations stayed within policy.
Cheating flags dominate, but they are not the alarming part
Roughly 20 percent of conversations triggered a content flag. Of those flagged conversations, 94.6 percent involved students trying to get the AI to produce finished work outright, asking it to write an essay, solve a maths problem set, or answer a quiz for them. That is the figure that will grab headlines, and it confirms what teachers already suspect: a meaningful slice of AI use is students outsourcing tasks rather than learning from them.
But it is also the least surprising number in the report. Academic integrity concerns are well understood, heavily discussed, and increasingly built into how teachers set assignments.
The number that should worry safeguarding leads more
Separately, around 2 percent of all prompts in the dataset, not just the flagged ones, showed signs of self-harm, bullying or violence. Securly compared this with unsafe content rates in ordinary student web searches, which sat at roughly 0.4 percent. That means potentially unsafe material is turning up in AI chat at close to five times the rate it turns up in a web search.
The comparison matters because web filtering is a mature, well-resourced layer of school IT. AI chat monitoring, for most districts and schools, barely exists yet.
Traditional web filters were built to block URLs and categorise search terms. A chatbot conversation is different: a student can describe a plan, a feeling, or an intent in natural language, in a private back-and-forth that never resembles a flagged search query. Existing filtering infrastructure was not designed to catch that, which is precisely the "visibility gap" Securly says pushed it to build dedicated AI monitoring into its platform this year.
Why this lands differently for UK schools
The dataset is American, and UK schools operate under different device management and filtering norms, often through providers such as Smoothwall or Impero rather than Securly. But the underlying dynamic, natural-language chat traffic that existing filtering tools cannot meaningfully see, applies wherever pupils use AI on school networks or personal devices during the day.
England's KCSIE 2026 safeguarding guidance already requires schools to have monitoring arrangements that account for AI and deepfake risks from September, but it does not specify how a school is meant to gain visibility into what a pupil actually typed into a chatbot. Most UK filtering contracts were negotiated before generative AI chat was in daily use, and few have been audited since to check whether they can see it at all.
That gap is where this US data is genuinely instructive. It suggests the safeguarding risk in AI chat is not hypothetical or rare, it is running at several times the rate schools already treat as significant when it shows up in web searches.
What to do
Ask your filtering or safeguarding provider directly whether their system logs and flags AI chatbot conversations, not just websites and search terms, and ask for evidence rather than a feature list. Treat AI chat monitoring as a distinct line item in the safeguarding review due before September, not an assumed extension of existing web filtering. Make sure designated safeguarding leads know that AI-related disclosures may arrive through a chatbot transcript rather than a flagged search or a reported conversation.
What to watch
Whether UK filtering vendors publish comparable usage data for British school networks, whether Ofsted inspection frameworks start asking specifically about AI chat visibility rather than general online safety provision, and whether any UK district-level pilot of AI monitoring software produces numbers that confirm or complicate the US pattern.