Researchers in the US have taught a computer how to make memes.
The Stanford University team says it can take "any image" and "produce a humorous and relevant caption".
But they admit it's not very good yet.
"We acknowledge that one of the greatest challenges in our project and other language modelling tasks is to capture humour, which varies across people and cultures," they write in a paper about their research, published online.
"At the end of the day, their purpose is to be funny."
They fed the neural network computer system 2600 memes, all of them consisting of text on an image. The memes it spat back out weren't quite as good as the real thing, with users able to tell a human-made meme from a computer-generated one 70 percent of the time.
"If indistinguishable this score would be near 50 percent," they said.
Another problem is the AI was unable to tell where the top half of the text ended and the second began, so the text output had to be split manually for the jokes to work.
"If the model could learn the breakpoints this would be a huge improvement and could fully automate the meme generation."
There was another problem too, but one less about AI and more about humanity.
"There was a bias in the dataset towards expletive, racist and sexist memes," the researchers admitted.