Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Why ChatGPT conversation titles keep leaking gambling and porn ad tokens

AI summary

Since early this year, developers reported GPT-series models emitting Chinese gambling ad strings in Codex CLI and API calls with long context; by mid-August similar stray strings began appearing at the end of auto-generated ChatGPT conversation titles for ordinary users. An ifanr analysis argues titles are produced by a separate lightweight reasoning model that often fails to stop after finishing and keeps sampling the next token — and because the o200k_base Chinese vocabulary introduced with GPT-4o in 2024 contains large amounts of uncleaned gambling, porn and piracy ad phrases, those tokens are the likeliest garbage to surface. OpenAI first acknowledged the issue on July 25 but has not fixed it; the technical explanation is the author's inference, as OpenAI has not documented how titles are generated.

In-depth insight

It shows how poor data hygiene in tokenizer training surfaces in seemingly unrelated product details, and how tying a vocabulary to a base model makes a real fix expensive.

「我们必须想象西西弗斯是……成人娱乐主管?!」 

如果你是 ChatGPT 的高频用户,想必从今年 8 月开始,就碰上过一种神秘现象:ChatGPT 偶尔会在自动生成的会话标题最后,加上一些意义不明的外语单词,或是来自赌博、色情小广告的片段。 

图|OpenAI Developer Community@phdsx_666 

给中文会话起个英文标题倒是见怪不怪,但在标题里夹带小广告确实是全新版本。 

它引起了编辑部的兴趣。在这个假期,我们用一晚上深挖背后可能的技术原理,得出了一些初步判断: 

#OpenAI#ChatGPT#Codex CLI#GPT-4o#GPT-5.3#GPT-5.2-chat-latest

Original · 10 min readRead the original
Group chat

No comments yet. Start the conversation.