ChatGPT “Something Went Wrong”: A Practical Fix

The generic ChatGPT error tells you almost nothing, so find the failing layer with a clean new chat, a short prompt and a second network before touching tenant settings.

Symptoms

You send a message and get a red bar or a small notice saying something went wrong, sometimes with "try again" or "please try reloading the conversation". The message you typed may vanish or sit unsent. It can happen on the first message of the day, on one long conversation only, or for every user in the office at once. Those three patterns point to different causes, which is why the first job is to notice which one you have.

Likely causes, most common first

1. A service incident. ChatGPT has partial outages fairly often. If several people see the error at the same time, check the status page before anything else.
2. Stale session or browser state. A tab that has been open for days, an expired login, or a cached script from before an update. Reloading usually fixes it; a private window confirms it.
3. An extension. Ad blockers, privacy tools, grammar checkers and "AI helper" extensions inject into the page and break requests. This is the most common cause on a single machine that used to work.
4. The conversation itself. Very long chats, large attachments or a request that hits a content or size limit can fail with the generic error while a new chat works fine.
5. The office network. Web filters, TLS inspection and proxies that block or rewrite traffic to chatgpt.com and OpenAI's content and static domains, or that break WebSocket and streaming connections. The tell is that it works on a phone hotspot.
6. Account or plan state. A seat that has just been removed from the workspace, a plan that has lapsed, or a sign-in change mid-session.
7. A VPN exit in an unexpected region. Some corporate VPNs exit through countries where the service behaves differently or is rate-limited.

Step-by-step fix

1. Copy the exact error text and note the time. You will need both if this goes to IT or OpenAI.
2. Open the OpenAI status page. If there is an active incident, stop here and wait.
3. Reload the page. If the error was a one-off, you are done.
4. Start a new chat and send a short message such as "reply with the word ok". If that works, the problem is the old conversation. Copy anything you need from it and carry on in the new one.
5. Open a private or incognito window, sign in and repeat step 4. If it now works, an extension or cached data is the cause. Disable extensions one at a time, or clear cookies and cache for chatgpt.com.
6. Try another network. Turn off the VPN, or use a phone hotspot for one message. If it works there, the office network is the cause and the fix belongs to IT, not to you.
7. Try another device on the office network. If it fails everywhere in the building, that also points to the network or to an incident that the status page has not caught up with.
8. Sign out fully and back in. If your organisation uses single sign-on, do this from the identity provider's portal so the session is really new.

Do not change firewall rules, reinstall browsers or reset the workspace as a first move. Each of those makes the picture harder to read.

What IT should check

• The web filter and proxy logs for blocked or rewritten requests to chatgpt.com, oaistatic.com and oaiusercontent.com during the time in your note. OpenAI publishes a list of domains and network recommendations in its Help Centre; match against the current version.
• Whether TLS inspection is applied to those domains. Streaming responses and WebSocket connections often fail under inspection even when normal page loads succeed.
• Whether a Defender or third-party web content filtering policy changed recently, including category-based blocks that quietly caught an AI domain.
• Whether the affected user's seat, group membership or conditional access policy changed that day.
• Whether the problem follows the user, the device or the network. Fix the one it follows.

If the network turns out to be the cause, the ChatGPT not working on a work network note has the allow-list approach, and the WebSocket network error note covers the streaming case.

When to escalate to OpenAI

Go to OpenAI support when the error reproduces in a private window, on a second network and on a second device, with no incident on the status page. Send the exact message, the time in UTC, the workspace and account email, the browser or app version, whether a VPN was in use, and the shortest prompt that triggers it. Redact anything from the conversation you would not put in an email to a supplier. A screenshot of the error with a client's document visible behind it is a new problem, not evidence.

Close the ticket only when the real task works again, not when the error message stops.

Source basis

• OpenAI: service status
• Google Chrome: clear cache and cookies
• Microsoft: web content filtering in Defender for Endpoint
• NCSC: using TLS to protect data