Most people, when they hear “AI safety,” are thinking about the same thing: machines going rogue and taking over the world.
I want to talk about a different kind of AI risk. One that’s already happening, and can cost you with permanent data loss.
In April 2026, Andrew Olsson — a technology commentator and YouTuber — posted a video documenting what happened when Anthropic closed his Claude account without warning. The video, titled “Anthropic Banned My Claude Account,” has since drawn attention from developers and IT professionals who recognise the situation as something that could easily happen to them.
The message Andrew received said he had breached Anthropic’s terms of service — something he strongly denies. He had no prior warning. No explanation of what the alleged breach actually was. Just a locked account and a wall of silence.
His chat history and project files — everything he’d built up inside Claude — were suddenly inaccessible. He lodged an appeal. Days passed. No response.
He went looking for a phone number, an email address, some way to reach a human at Anthropic. There wasn’t one.
He was completely stuck.
Now, I want to be clear about what actually happened to him. It wasn’t just a denial of service. It was data loss. Those two things together are the double whammy. He lost access to the tool and he lost everything he’d stored inside it. The account closure didn’t just lock the door — it locked the door with all his work still inside.
That’s what makes this different from a web service going down for an hour.
Here’s the twist I keep coming back to.
When people talk about AI risk, the assumption is always the same: the threat flows outward from AI toward us. The AI does something unexpected, something harmful, something beyond our control. That’s the direction the whole “AI safety” conversation points.
But this risk runs the other way. The threat isn’t the AI going rogue. It’s you becoming dependent on the AI — and then having access stripped away through no fault of your own.
Same phrase. Completely different direction.
Let’s think about what Andrew’s situation actually required of him. The AI provider unilaterally closed his account and did so on the presumption of guilt. A large corporation can brand you guilty, and then it’s up to you to try to prove yourself innocent.
This is the very opposite of one of the foundational principles of modern legal systems: the presumption of innocence. In any fair legal process, the burden of proof lies with the accuser, not the accused. You are innocent until proven guilty — and critically, you have the right to know the case against you and to contest it before an impartial party. None of that applied here.
So while you are presumed guilty, it means you are denied access to all of the existing data in your account that you may have spent hundreds or thousands of hours building up. After all, the correct way to use Claude is to build up your knowledge base and put them into projects. And all of a sudden, those files are inaccessible to you.
According to Andrew, he was stuck. And despite submitting the appeal form, he ended up losing hope that anything would happen.
When I heard of this case, I knew the story before. Goliath Corporation won’t take care of you as a customer. So you have to take care of yourself.
This isn’t a new lesson. The need to own your own data and keep your own backups predates AI by decades. Store your critical data on a third-party system and you are, at some level, at their mercy. Cloud storage taught us that. SaaS tools reinforced it.
What’s changed is the speed and the finality. When a traditional SaaS vendor terminates an account, there’s usually a wind-down period and a data export window. AI account closures, as Andrew found out, can happen overnight with no warning and no recovery path.
The lesson is the same as it’s always been. The stakes just got higher.
I’ll be honest about my own situation here.
I run a backup company. I have spent more than two decades telling people that data loss happens in ways you don’t anticipate, and that the only rational response is a backup that sits outside the system you’re trying to protect.
When I came across Andrew’s video, my first thought was: do I have a backup of my Claude project files?
I didn’t.
I went and did one manually this morning. There’s no automated way to do it — at least none I could find. So I exported what I could, saved it locally, and accepted that this was a gap I’d been ignoring.
A manual backup is better than nothing. But it’s telling that even someone in my position hadn’t thought about this until a stranger’s bad experience showed up in my feed.
The practical takeaway is straightforward. The data you generate inside AI tools — your project files, your curated context, your prompts — has real value. In some cases, it represents hundreds of hours of work. Treat it accordingly.
- Know where your data lives inside these tools
- Understand what export options exist, and use them
- Do a backup on a schedule that reflects how much you’d actually lose if access disappeared tonight
And ask the harder question: if your AI account was closed right now, with no warning and no appeal, what would you lose? If that answer makes you uncomfortable, that’s useful information.
The AI companies haven’t solved this problem. Until they do, we have to solve it ourselves.
