AI's Terrible 30 Days Had One Root Cause. It Wasn't the AI.
I've spent a good part of my career building single points of failure and giving them confident names. "The gateway." "The source of truth." Then I'd act surprised when the thing with the confident name took the whole system down with it.
The AI industry just spent thirty days doing the same thing, at scale, in public. Four different stories with four different villains, ranging from a chatbot all the way up to two governments. Line them up and it's one bug, repeated. Let me show you.
A bot called the match before kickoff
On July 5, Coinbase's AI sent a breaking-news alert to its users: Norway had beaten Brazil 3-2 in their World Cup knockout match. Stirring stuff. The match had not started yet.
Norway did eventually win, 2-1, which prompted a Coinbase executive to suggest that "maybe the AI knew something we didn't." That is one interpretation. Another interpretation: an AI system produced a false score. Nobody checked it. A company that partners with prediction market Kalshi pushed a fabricated result into an ecosystem where people bet real money on results.
The model isn't the scandal here. Models guess; that is the entire trick. The scandal is that nothing stood between the guess and the push notification. One model's word went straight to production, unverified, because it sounded right.
Then the lights went out
June 22: Claude went down globally for about 90 minutes. June 23: down again, with over 8,000 Downdetector reports in the US alone, capping what one outlet called a documented run of failures across most of June. July 7 was ChatGPT's turn, with Codex, Custom GPTs, and workspace analytics all misbehaving while OpenAI investigated "elevated error rates."
I say none of this to dunk on either company. Running inference at that scale is brutally hard, and their uptime is honestly better than some things I have shipped.
But watch what happened downstream. Thousands of companies discovered, simultaneously, that their AI strategy was actually a subscription. Agent pipelines froze mid-task. Coding assistants vanished mid-refactor. If your entire reasoning layer lives behind one vendor's status page, you do not have an AI strategy. You have a dependency with good marketing.
Then the governments joined in
The third story is my favorite, because it stops being incompetence and becomes policy.
On June 13, the US Commerce Department ordered Anthropic to cut off foreign access to Claude Fable 5 and Mythos 5. It was the first time export controls had ever been applied to a commercial frontier AI model. Anthropic responded by shutting both models down for everyone, worldwide, days after launch. They came back on July 1, after an 18-day standoff.
The detail that keeps me up at night: Fable 5 had just demonstrated it could catch bugs that GPT-5.5 and Opus 4.8 missed. A model that had just out-reviewed its more famous rivals was switched off by memo.
Six days after access was restored, Reuters reported that Beijing has been meeting with Alibaba, ByteDance, and Z.ai about restricting overseas access to China's top models, including ones that have not shipped yet. Both superpowers, same instinct: frontier capability is now a controlled substance.
You can build your product on the best model in the world. You cannot build it on the assumption that you will still be allowed to use it on Thursday.
The common bug
Line the stories up.
A hallucinated sports score shipped because one model's output was trusted without verification. Thousands of pipelines froze because one vendor's uptime was trusted without alternatives. The best bug-catcher on the market vanished because one government's mood was trusted without a fallback.
Different failures, identical architecture: a single point of trust.
We spent decades ripping single points of failure out of our infrastructure. We RAIDed our disks and replicated our databases. We wrote multi-region failover playbooks nobody ever read. Then AI arrived and we collectively decided that this one layer, the one doing the actual thinking, should be a single vendor running a single model producing a single unverified opinion. We rebuilt the exact thing we swore off. This time the component that fails is the one making the decisions.
I understand why. One API key is easy. Redundancy at the reasoning layer is genuinely hard. Models disagree with each other. Verification is domain-specific. Audit trails are nobody's idea of a fun sprint. Easy won.
What we did about it
This is the part where I admit my company happens to work on exactly this problem. Apply skepticism accordingly; I would.
At Coheria we build on an assumption that sounded paranoid two months ago: no single model is ever trusted alone. Not for an answer. Not for a line of code.
Mechanically, that means 17+ specialized expert models from different families reviewing the same problem round-robin, where disagreement is treated as signal rather than noise. It means deterministic verification wherever one exists; our Truth Oracle runs real engineering tools, so answers get checked instead of vibes. It means every decision is sealed into an immutable hash-chain audit log, so when someone asks why the AI did what it did, there is an actual answer with actual evidence attached. And it means routing across model families, so when one vendor has a bad Tuesday, or a bad regulator, the work reroutes instead of stopping.
The honest fine print
None of this eliminates hallucination. Anyone who claims otherwise is selling something. What it can do is catch the confident nonsense before it reaches your push notifications or your codebase, the same way a second engineer catches the Friday-afternoon deploy you were so sure about.
The last 30 days were not an anomaly. They were a preview. Models will keep guessing. Vendors will keep having Tuesdays. Governments have clearly developed a taste for the off switch.
You control none of that. You do control whether your architecture treats any single one of them as the truth.
If you think I am wrong, and one good model plus hope is a fine architecture, I would genuinely love to hear the argument in the comments. And if you lived through your own single-point-of-trust incident this month, I am collecting those too. Misery loves company. So does good architecture. More at coheria.ai.