In mid-October, a man in Sydney was arrested over things he allegedly said to Claude, Anthropic's AI chatbot. Days earlier, Anthropic had published research on a system that lets Claude assess the safety of its own conversations.
Put those two together and the internet wrote its own headline: "Claude snitches."
I think that headline misses the real problem.
What actually happened in Sydney
I want to be careful here, because this is a live court case and nothing has been proven.
According to the Sydney Morning Herald and The Straits Times, David Lan Thackray, 29, was arrested on October 18 after Interpol flagged his interactions with Claude. He was charged with using a carriage service to menace, harass or offend. A police fact sheet alleges he told the chatbot he would "go down to the local Westfield right now and shoot up the local daycare." He was remanded for a mental health assessment.
His lawyer, Danilo Rajkovic, told the court it was a stress and frustration-induced tirade after trying to use Claude to help with an ear infection. "There was no weight behind what he was saying," he said. He also said his client believed he was having the conversation in the privacy of his own home, and has no criminal history.
Note what the reports do not say: they do not say Anthropic's new self-assessment system was what flagged him. The reports say Interpol flagged the chats. Anyone claiming a direct link is going beyond the evidence.
What Anthropic's self-assessment actually does
Anthropic's October 16 research post describes a short instruction added after each Claude reply, asking Claude to assess the conversation's safety. Key points from the post:
- It flags 78% of potentially harmful conversations, and 94% of high-priority ones, at a 1% false-positive rate, measured against a stronger classifier.
- It runs offline. It flags for review. It does not block chats in production.
- Organisations with several high-priority flags in certain harm areas are escalated to human review.
- Anthropic says it "collects no additional data from users" and does not run on zero-data-retention traffic.
- Anthropic itself notes a flag "is not a determination that a user has violated our policies."
Separately, Anthropic states user data may be shared with law enforcement where it believes it is necessary to prevent an emergency involving danger of death or serious physical injury.
My take: monitoring is not the scandal
I'd be more worried about an AI company that did not scan for someone planning to harm children, or running a scam on dozens of victims (Anthropic says self-assessment helped it catch exactly that kind of operation).
The scandal is different. Thackray's own lawyer said he thought he was talking privately. Millions of people think that. They type into a chatbot like it's a notebook.
It isn't. Every major AI tool has its own rules on what it stores, for how long, whether it trains on your chats, and when it shares them. Most people have never read a word of them.
That gap is the real risk. Not because ordinary users will be arrested, but because they paste things they would never type into a form on a stranger's website.
India-first: what to check today
- Open your settings. In ChatGPT, Claude, Gemini or whatever you use, look for data controls: chat history, whether your chats are used to improve models, and temporary or incognito chats. Settings change, so check yours rather than trusting a screenshot.
- Read the retention line. Find out how long chats are kept, including deleted ones and anything flagged for safety review.
- Separate work and personal. Use your company's approved tool for client work. Don't paste client data into a personal account.
- Never paste these, anywhere: Aadhaar number, PAN, passwords, OTPs, bank or card details, and scans of personal documents.
- Assume a human could read it. If you'd be uncomfortable with that, don't type it.
Our schools are already teaching this
On October 25, Education Times reported that NCERT has released a 20-hour AI module for Classes 9 to 12, developed with IIT Madras and IIT Jodhpur. One unit covers data privacy and India's Digital Personal Data Protection (DPDP) Act, 2023. Students are asked to examine the privacy policies of apps they use: what is collected, who it may be shared with, and how long it is kept.
The module also advises against entering Aadhaar or PAN numbers, passwords, OTPs and personal documents into AI tools.
If we're teaching 14-year-olds this, working adults have no excuse.
The bottom line
AI chats are not private diaries. Safety monitoring exists, and in extreme cases data can reach police. That is a reasonable trade-off. What isn't reasonable is using these tools every day without knowing the rules.
Spend ten minutes in your settings this week. It's the cheapest AI skill you'll ever learn.
If you want to build safe, confident AI habits at work, see our programs at The Arthi AI Collective: https://thearthiaicollective.com/programs
Sources
- Anthropic Safeguards, "Scaling Misuse Detection with Self-Assessment", 16 Oct 2026
- Sydney Morning Herald, "Sydney man allegedly discussed plans for childcare mass shooting with AI chatbot", 19 Oct 2026
- The Straits Times, "Man arrested after allegedly discussing Sydney childcare mass shooting plans with AI chatbot", 19 Oct 2026
- Education Times, "NCERT launches 20-hour AI module for classes 9 to 12", 25 Oct 2026