One million users and one ban, on the same Monday

Two posts landed on December 5, 2022, and they read like a conversation neither side meant to have.

The first was from Sam Altman, in full: "ChatGPT launched on wednesday. today it crossed 1 million users!" The archived tweet is timestamped Dec 5, 2022. In the replies, someone asked if it would be free forever. Altman answered that they'd have to monetize it at some point, because "the compute costs are eye-watering." Musk asked about the average cost per chat, and Altman guessed probably single-digit cents.

The second was on Meta Stack Overflow, posted by moderator Makyen at 05:34 UTC that day, which was still Sunday evening in the Americas. The title was blunt: "Temporary policy: ChatGPT is banned."

What the ban actually said

It's worth reading the reasoning, because it holds up. The post said the average rate of getting correct answers from ChatGPT was too low, so posting its answers was substantially harmful to the site. The core problem, in the moderators' framing, was that the answers had a high rate of being incorrect but typically looked like they might be good, and were very easy to produce.

The volume was the other half. The post described thousands of these answers, each needing a detailed read by someone with some subject knowledge to show it was bad. That had swamped the site's volunteer-based curation. So the moderators said they'd deal with users rather than individual posts, with sanctions for anyone believed to have used ChatGPT after the policy went up, even if a given post would otherwise be fine. The ban covered all content on the site except profile text.

Same week, two lessons

Put the two posts side by side. A million people tried a tool in five days. And one of the busiest Q&A sites for programmers found that a lot of those people were pasting its output somewhere public without checking it.

I read that as a story about skipping the step where a person confirms the answer works. Plenty of the answers were probably useful. The trouble, as the moderators described it, was that checking took expertise and posting took seconds.

Prompt for the check, not just the answer

You can ask the model to do some of the checking setup for you. It won't make the answer true, but it tells you exactly where to look.