AI Memes are novel, but will not remain so.

with the release of new open access AI image generation and text writing tools, lots of people have employed AI to generate new jokes. However, this does not yet herald an end to human authorship of memes. There are three key factors that make AI memes funny right now: 1. Uncanny valley of humor, 2. Human contextualization/framing, 3. knowledge of AI authorship.

The greatest example of this I can think of was a twitter thread that prompted GPT-3 to write news articles about trump dunking on the apostles from the New Testament. The prose was impressive but slightly stilted, adding to the humor. The AI expertly blended Trump's rhetorical style with domain knowledge of the apostles from the bible. But if we believed someone earnestly wrote these parodies, then it would seem trite, overdone, cringe, etc. It was only the novelty that AI is capable of writing like this that made it funny.

It is clear to me that AI has a long way to go before creating truly insightful memetic innovations. AI memes for now add novelty to existing tropes, and will eventually lose traction until something changes.


Never heard of AI generating jokes like that, but that's hilarious.
I do wonder too about AI generating memes, that the reason AI can generate memes like that, is because people have already created memes/jokes like that previously that gave AI the database that those kinds of words/sentences/phrases/structures are what give people entertainment. I guess in that sense, if we consider that people's perspective of humor will evolve and change, AI will need to change its database of what is funny as well, which means perhaps people need to feed AI on what is funny, and not AI giving new funny to people.
If we think about that, then maybe AI will never end human authorship of memes, because humans' perspective on what is funny is what make memes good memes.


When I started playing with GPT-3, I was totally blown away by the quality of the text, and was convinced that it represents a form of "proto-consciousness". There was novelty there, and a level of unpredictability that made it totally amazing. Now, i'm not as impressed by it, both because I've gotten used to its response style, and also because i'm even more amazed at what the next language models are and will be capable of (LaMDA, etc). GPT-3 was amazing when i first interacted with it, but it has become quite predictable for me.

Given that we live in a society of growth and acceleration (for better or worse), novelty will inevitably wear off as we continue to have these technologies around us.

1 Like

I dunno. These ones fell flat, but Bots of New York is still hilarious. I think part of the issue might be that GPT-3 is too good - these examples read like bad attempts at writing Onion-esque parody articles, they've got the theme but no real jokes, and they also lack the sure absurdity that makes Bots of New York so funny. It's really too much on the prompt instead of the bizzaro world GPT-2 ended up hallucinating on it's own. Maybe Bots of New York too will one day loose it's charm, but this just doesn't seem to hit the mark.

The best AI generated jokes are not actually trained on jokes, the reason they are so funny is because they fail to imitate what they were trained for, and the humor emerges in that gap. There is also a tendency for certain language models to create "puns" as a side effect of their loose sense of context and ease with handling polysemy. It's a classic case where something is funny only when it doesn't try to be.

1 Like