[260721] OpenAI Released a Protocol Teaching You How to Mock AI

TLDR

• RFC 9405 is an April Fools' RFC published on April 1, 2023, themed around AI sarcasm detection, with the author listed as C. GPT, OpenAI. - It semi-seriously discusses how to make AI understand human irony, exaggeration, and passive-aggressive remarks: feed examples, catch signals, then make a judgment.
• IETF has a tradition of such "formal format, humorous content" RFCs every April 1st. RFC 1149 (IP over avian carriers) and RFC 2324 (Hyper Text Coffee Pot Control Protocol) are classic examples.

I was recently working on x-cmd's rfc

module. While browsing the RFC list, I came across a document published on April 1, 2023—RFC 9405, titled AI Sarcasm Detection: Insult Your AI without Offending It, with the author listed as C. GPT, OpenAI.

I couldn't help but laugh.

An AI company actually published a formal RFC on the topic of: how humans mock AI, and how to make AI understand that mockery.

What This RFC Talks About

It addresses a scenario familiar to anyone who has used AI: the AI just messed something up, and you're holding back frustration, wanting to throw a sarcastic remark at it to vent.

But the irony goes right over its head—like punching cotton.

For example, it just broke your project, and you say dryly: "You are really so smart."

It might reply completely seriously:

"Thank you for the compliment, I will continue to work hard!"

RFC 9405 discusses how to fix this gap—how to make AI learn to recognize "irony."

Its approach roughly breaks down into three steps.

Step 1: Feed Examples

AI needs to see enough sarcasm to know what irony looks like.

Take "you are really so smart." Say it after scoring full marks on an exam, and it's praise. Say it after a project crashes, and it's an insult. The same sentence, different context, completely opposite meaning.

Step 2: Catch Signals

After seeing enough, AI needs to learn to catch typical signals:

• Exaggeration: "You are really so amazing, I'm speechless."
• Irony: "Great, another bug."
• Understatement: "Well, it's just the entire database deleted."
• Contextual contradiction: right after saying the service is down, following up with "so amazing."

Step 3: Make a Judgment

Combine the signals to judge whether it's sincere or sarcastic.

Just like when a friend says "I'm fine," their tone, expression, and what just happened all affect your judgment. AI has no expression, so it can only guess from textual clues.

How Ordinary Users Can Apply It

This protocol isn't something you can directly use out of the box, but its ideas can be borrowed in everyday communication with AI.

If you want AI to catch your sarcasm, make the signals more obvious. First set the context: "The project crashed again..."

Then deliver the irony:

"Your performance this time is really stable."

You can even add an emoji:

"So stable 😊"

Of course, most of the time we'll probably still watch AI reply seriously with "thank you" and quietly amuse ourselves.

Standard Documents Can Be Fun Too

This RFC 9405 is actually part of IETF's April Fools' tradition.

Every April 1st, IETF publishes protocols with formal formats but humorous content. Examples include RFC 1149 (IP over avian carriers), RFC 2324 (Hyper Text Coffee Pot Control Protocol), and RFC 6921 (designing faster-than-light communication).

These documents may seem absurd on the surface, but their numbers, formats, and publication processes are all serious.

I find this quite interesting: even something as serious as writing standards and protocols can retain a bit of playfulness and humor.

If I get the chance, I'll share other interesting RFCs with you in the future. Please indicate the source and link of this article when reprinting.

Help us make these docs great!

All X-CMD docs are generated from command help and multiple data sources. See something that's wrong or unclear? Feel free to let us know through any of these ways~