GPT Loves Gambling and Bumping Butts - by Brandon Sovran
Awaiting Input
SubscribeSign in
GPT Loves Gambling and Bumping Butts<br>I should probably circle back._人人碰final彩票
Brandon Sovran<br>Aug 12, 2026
Share
“Options: a) … b) … c) split into public spec + internal admin spec. +#+#+#+#+#+user to=assistant code _人人碰final 彩娱乐彩票” 1<br>What the hell are 人人碰 and 彩娱乐彩票 doing there?<br>Well to my Chinese readers, you probably already know what this is, but to my non-Chinese readers, the LLM was thinking about none other than videos of people bumping2, and a color entertainment lottery3. And this isn’t the only time. They think about gambling and adult websites, A LOT.<br>Back in 2024 this issue was rampant. GPT-4o’s tokenizer was especially bad, with independent researchers estimating more than 23% of its long Chinese tokens associated specifically with adult videos or gambling. 4<br>That raises a much weirder question. Why would an AI tokenizer know so many oddly specific ways of talking about Chinese porn and gambling sites?
Why does GPT love Gambling and Porn?
The researchers tried to pin down where it might occur and why, theorizing that an LLM’s tokenizer can act like a fossil record of its training data. More strikingly, they estimated that webpages related to 波多野結衣, a single popular adult performer, could of accounted for ~0.5% of GPT-4o’s inferred training data. Twitter users were also early to get into the investigation, showing how often the lottery and adult sites actually appear in the base vocabulary. 5
While there is no hard proof of what actually goes into the training data at companies like OpenAI, one thing we can be sure of is that the models saw it, and they clung to it, unable to shake off their favourite terms to this day. There are examples on GitHub issues showing these terms leaked into GPT-5.36, GPT-5.47 and GPT-5.68 models alike. This issue isn’t new, but it begs the question, why aren’t they doing anything about it?<br>Maybe this is just unavoidable when you train on the internet
The strangest part isn’t that OpenAI built a dirty tokenizer in 2024. It’s that it appears they have kept using that tokenizer family to this very day. OpenAI’s current tiktoken repository maps GPT-5-family models to the same o200k_base encoding associated with GPT-4o 9. Deepseek, other frontier labs, and even stranger the old GPT tokenizer seems to be lacking the rampant slew of gambling and pornography found in their new tokenizer. 10
So maybe we can live in a world where all LLMs are void of their vices. But until OpenAI retires their o200k_base tokenizer (and cleans up whatever data produced such vices) for their next series of models, buying lottery tickets and watching adult videos will stay firmly on the GPT’s families minds. 11
Subscribe
Thanks for reading Awaiting Input! This post is public so feel free to share it.
Share
1https://github.com/openai/codex/issues/11966
2人人碰 (rén rén pèng): Literally “everyone bumps/touches”; an adult-content term associated with Chinese pornographic websites and videos.
3A color entertainment lottery is an online betting platform that combines traditional lottery number drawings with casino-style entertainment elements. The word “color” (彩) is short for fucai (welfare lottery) or ticai (sports lottery) , but in this specific context, it is used broadly to mean any number-prediction or color-prediction betting game.
4https://arxiv.org/html/2508.17771v1
Susan Zhang@suchenzang
this new "o200k_base" vocab for gpt-4o makes me want to clutch my pearls
12:03 AM · May 14, 2024 · 312K Views
43 Replies · 68 Reposts · 722 Likes
6https://github.com/openai/codex/issues/13260
7https://github.com/openai/codex/issues/18986
8https://github.com/openai/codex/issues/34114
9https://github.com/openai/tiktoken/blob/main/tiktoken/model.py
10https://pollutedtokens.github.io/
11https://github.com/openai/codex/issues/13867
Share
Discussion about this post<br>CommentsRestacks
TopLatest
No posts
Ready for more?
Subscribe
© 2026 Awaiting Input · Privacy ∙ Terms ∙ Collection notice<br>Start your SubstackGet the app
Substack is the home for great culture
This site requires JavaScript to run correctly. Please turn on JavaScript or unblock scripts