The founder of one of the UK’s leading AI start-ups has said big AI models will only get “crazier” unless more is done to control the information they are trained on.
Emad Mostaque, CEO of Stability AI, argues that continuing to efficiently train large language models like OpenAI’s GPT4 and Google’s LaMDA across the internet makes them too unpredictable and potentially dangerous.
“The lab itself said this could pose an existential threat to humans,” Mr Mostak said.
On Tuesday, OpenAI head Sam Altman told the U.S. Congress, Technology can ‘go terribly wrong’ and needs regulation.
Today Epsom College principal Sir Anthony Shelton told Sky News’ Sophie Leech on Sunday that artificial intelligence could May be “offensive and dangerous”.
The founder of the “painting of Edinburgh Castle” generated by the artificial intelligence tool Stable Diffusion has warned that not all internet users can tell the difference between a real image and an artificial intelligence image.Figure: Stable Diffusion
An image of a “green and orange fruit print” generated by Stable Diffusion, an artificial intelligence tool that converts text into images.Figure: Stable Diffusion
“When people make [the models] So maybe we should have a public discussion about this,” Mr Mostak added.
But AI developers like Stability AI may not have chosen to have such discussions. Much of the data used to train its powerful text-to-image AI products is also “crawled” from the internet.
This includes the millions of copyrighted images that have led to legal action against companies, as well as major questions about who ultimately “owns” the products created by image or text-generating AI systems.
His company collaborated on Stable Diffusion, one of the leading text-to-image AIs. Stability AI just unveiled a new model called Deep Floyd, which it claims is the most advanced image generation AI to date.
“Photo of a cute fluffy owl drinking a very dark beer” by AI. Image: Deep Floyd
Realistic style image of “playful furry fox as pilot” created by artificial intelligence. Image: Deep Floyd
Daria Bakshandaeva, a senior researcher at Stability AI, explained that a necessary step to make the AI safe is to remove illegal, violent and pornographic images from the training data.
If the AI sees harmful or explicit images during training, it can recreate them in the output. To avoid this, developers remove these images from the training data so the AI can’t “imagine” what they look like.
But it still needs 2 billion images from online sources to train it. Stability AI says it is actively researching new datasets to train AI models that respect people’s rights to their data.
Stability AI was sued in the US by photo agency Getty Images for using 12 million of its images as part of a dataset used to train its models. Stability AI responded that the rules on “fair use” of images meant there was no copyright infringement.
But the problem isn’t just copyright. An increasing amount of data available on the web, whether it’s images, text, or computer code, is generated by AI.
“If you look at coding, 50 per cent of all the code that is generated today is AI-generated, which is an amazing shift in just one year or 18 months,” Mr Mostak said.
Text-generating artificial intelligence is creating more and more online content, even news stories.
An image of “England to win the Men’s Football World Cup in 2026” generated by the artificial intelligence tool Stable Diffusion, which converts text into an image, shows that the tool is not always accurate.Figure: Stable Diffusion
Sir Anthony Seldon highlights benefits and risks of artificial intelligence
News Guard, an American company that verifies online content, recently discovered that 49 online “fake news” sites generated almost entirely by AI were being used to drive clicks on ad content.
“We remain very concerned about the ability of the average Internet user to find information and know that it is accurate,” said Matt Skibinski, NewsGuard’s general manager.
AI runs the risk of polluting the web with intentionally misleading and harmful or spammy content. It’s not that people haven’t been doing this for years, it’s just that now AI may finally be trained on data scraped from the web created by other AIs.
There is even more reason now to think hard about what data we use to train stronger artificial intelligence.
“Don’t feed them junk food,” Mr Mostak said. “We can have much better free-range organic models now. Otherwise, they would just keep going wild.”
A good place to start, he argues, is for the AI to be trained on data, whether text, images or medical data, that is more specific to the users it’s targeting. Currently, most AI is designed and trained in California.
“I think we need our own datasets or our own models to reflect human diversity,” Mr Mostak said.
“I think it’s also going to be safer. I think they’re going to be more in line with human values than just having a very limited data set and a very limited set of experience that’s only available to the wealthiest people in the world.”