Three AI Assistants Went Down Together. Your Assignment Named One of Them.

Thursday morning your students could not reach the tool your assignment told them to use. The due date did not move.

Here is what happened, in order, because the order is the part that matters.

At about 9:30 in the morning Eastern time, Grok started failing across Android, iOS and the web. Users prompting it on X got back a single line: This model is overloaded right now. Please try again shortly or pick a different model. xAI later linked the failure to an outage at its Memphis data center.

Around 11am Eastern, ChatGPT began returning errors. OpenAI’s status page reported elevated errors across ChatGPT and Codex. The outage did not only stop conversations. It took down logins, file uploads, voice mode, search, deep research and image generation, which is to say almost everything a student might have been in the middle of.

Anthropic’s Claude went down at roughly the same time, along with Claude Code and the Claude API. A technical staff member described it as an infrastructure issue. Anthropic marked it resolved at about 12:15pm Eastern.

Three companies, three separate stacks, one morning. The Verge, which reported all of it, noted that Gemini and Copilot were the last chatbots standing, and that it is not clear whether the three failures were related at all. Nobody has said they were.

Then Friday made the same point a different way

While ChatGPT was erroring out on Thursday, OpenAI was launching GPT-6 Astra, which it called a generational leap in capability and the start of the AGI era. The rollout went to enterprise customers first, specifically those with access to its Daybreak cybersecurity platform, and was to expand to Plus, Pro, Business and Enterprise subscribers over the following days.

By Friday morning Sam Altman was apologizing for what he called a messy rollout. Paying subscribers, including people on the more expensive Pro plan who are used to getting new models at launch, were still waiting. “We are working towards getting Astra in everyone’s hands as quickly as we can,” Altman wrote. “I know it is frustrating and I appreciate the patience. It should be quick.” Asked when, he offered this: “I am hopeful that you can use it this weekend! but can’t promise yet.”

OpenAI’s Codex engineering lead said the company would give one banked reset for every day a paying user goes without access.

So inside forty eight hours, in the second week of the semester, the tools your students use failed twice for two completely different reasons. Once because infrastructure broke. Once because a company made a business decision about who goes first.

This is not the argument you think it is

It would be easy to read the last two days as evidence for banning AI in coursework. It is not, and reaching for that is how faculty lose the plot on this. A library database goes down. A learning management system goes down during finals week. Nobody responds by removing sources from the syllabus.

The actual lesson is smaller and much more useful. When you wrote use ChatGPT to draft an outline into an assignment, you made a procurement decision. You picked a single vendor, committed your students to it, and gave yourself no visibility into whether it will be available on the morning it is needed. You did that without a contract, without an SLA, and without anyone in your institution knowing you did it.

That is not carelessness. It is what happens when a tool becomes normal faster than the paperwork around it. But it is worth seeing clearly, because the fix is genuinely small.

The word for this is monoculture, and it is not a new idea

People who design infrastructure for a living have a name for what happens when everything depends on one thing. They call it monoculture, and the concern is exactly this: if the single system everyone relies on has a bad day, everything downstream has a bad day with it. The standard remedy in safety critical work is a two source rule, which means running anything you cannot afford to lose on two independent systems built by different people. Grid operators measure themselves on outage minutes avoided, not on how good the grid is when it is working.

None of that is a metaphor here. It transfers almost directly. Your course is a system, your assignment is a dependency graph, and right now most assignments that mention AI name exactly one tool.

There is a second idea worth borrowing, and it comes from OpenAI’s own policy writing: resilience depends on what happens after deployment, when systems have to operate under uncertainty and integrate into institutions that were not designed for them. A university is precisely such an institution. Nothing in a course approval process asks what happens when a vendor has an incident.

What to actually do, this week

Not a policy. Not a committee. Three concrete moves, in order of how little they cost.

One. Name the assumption out loud. Take your next assignment and read it as a student with no access. Which steps still work? Most assignments turn out to be sturdier than their authors feared, and finding that out is worth the ten minutes on its own.

Two. Write the outage clause. One sentence in the assignment itself, in your voice, that tells a student exactly what to submit and what happens to the deadline if the tool they planned to use is unavailable. The test is whether a student could read it at 11pm on their phone, with the tool down, and know their next move without emailing you. If they have to email you, the sentence is not finished.

Three. Stop naming one tool. Where the tool is a convenience rather than the point, say “an AI assistant” and let the student pick. Where the tool genuinely is the point, because you are teaching them to evaluate its output, name two. That is the two source rule, applied to a syllabus, and it costs you one word.

The part nobody says out loud

Thursday was inconvenient. Friday was something else, and it is the one worth remembering. Nothing broke on Friday. The model shipped, on time, working. It simply went to business customers first, and the people who pay twenty dollars and two hundred dollars a month found out what tier they were in.

That is not a bug you can wait out. It is the ordinary behaviour of a market, and it will happen again on a schedule nobody publishes. Your students are in that queue too, mostly on the free tier, which means they are further back in it than you are.

An outage clause protects a student from an incident. Not naming a single vendor protects them from a business model. The second one matters more, and it is the cheaper of the two to write.

Where this goes next

You can write one outage clause by hand in an afternoon. Then you look at your schedule and count what is left, and multiply it by your sections, and the honest answer is that clause needs to exist in a lot of places in your own words.

That is the job the Assessment Builder in the Faculty Toolkit does. It builds assignments, rubrics and tagged question banks, which means it works on the assignment itself, which is the only place an outage clause can actually live. Do the first one by hand first. You need to know your own ratio before a tool is any use to you.

If you want to compare notes with other instructors working through the same thing, the Harvest Kernel learning community is free and the conversation there is the practical kind.

Sources

• Emma Roth, “ChatGPT, Grok, and Claude all went down at the same time,” The Verge, September 3, 2026. theverge.com
• Robert Hart, “Sam Altman apologizes for ‘messy’ GPT-6 Astra rollout that’s locked out paying users,” The Verge, September 4, 2026. theverge.com