As Michael would surely agree, there are no real days off in crypto. I was reminded myself when I recently spent a long weekend at the fantastic Readercon fiction convention. Inevitably, I missed some important crypto stories, but I also got some up-close insight into another looming novelty: the existential threat that automated large language models (LLMs) like GPT3 pose to the entire internet.
That might sound hyperbolic. But at Readercon, I met Neil Clarke, founder and editor of the top-tier science fiction magazine Clarkesworld, which along with other fiction publications has become a canary in the coal mine of A.I. run amok. The rise of ChatGPT has inundated these journals with a flood of fake GPT-generated story submissions, a plague so severe Clarkesworld was forced to temporarily pause submissions this February, threatening the work and livelihoods of real authors.
“I’ve been calling it spam,” says Clarke, “Because that’s what it is. I sometimes refuse to even call it ‘artificial intelligence.’ You can’t humanize these things. It’s not like the science fiction of movies where it’s aware. It’s a statistical [language] model.”
The mention of spam should raise the antenna of longtime cryptocurrency watchers: the same problem lay at the very origins of Bitcoin.
Between 1998 and 2002, computer scientist Adam Back developed the concept of “Hashcash,” primarily intended to combat e-mail spam by requiring a tiny payment to send one. Back and his ideas became foundational to the development of Bitcoin, and he’s now CEO of crypto developer Blockstream.
Two decades later, with robotic barbarian hordes poised to swamp human communication systems, it might be time to revisit the Hashcash concept.
Large Language Hustlers
“ChatGPT came out in late November,” Clarke says, “And we immediately started seeing submissions using it. The first people to adopt it were the ones already submitting plagiarized works. It was readily embraced by people who were trying to make a quick buck off other people’s work.”
As they faced down the spam problem, Clarke says he and his team quickly realized the attack was coordinated. YouTube and TikTok channels focused on get-rich-quick schemes were promising viewers they could make thousands of dollars by submitting GPT-generated stories to fiction magazines like Clarkesworld. Clarkesworld pays a few hundred dollars per story, depending on length – not much more than beer money in some parts of the world, but extremely meaningful in others.
Those fraudulent promises from online grifters seem to have spread fast. Clarke says he received 54 AI-generated submissions in December. In January, he got 117 fake stories. In February, the number hit 514 before Clarke closed submissions midday on February 20.
“And that morning alone,” he says, “we had 50.”
Clarkesworld has a small staff, who normally review about 1,100 submissions a month. So the accelerating flood of trash threatened to overwhelm them, and solutions weren’t obvious.
“We have an open submission process, specifically designed to welcome in new writers and new voices,” says Clarke. “So we could close submissions from certain locations [to fight spam], but we also have legitimate authors coming in from those countries. And we’ve been told things like, ‘The payment for this story will cover my bills for a month.’”
“Authors like that are getting buried. The A.I. submissions hurt new authors, and authors who might not be from communities that are well-connected.” This is one clear way auto-generated content threatens to make the internet worse for human beings – particularly those at the margins.
“If you go back 15 or 20 years when we took submissions on paper,” Clarke says, “just the cost of postage was enough to decrease submissions from outside the U.S., Canada, and U.K. substantially. And as soon as you have digital submissions, we had this flood of international submissions.” That has led to a huge diversification of the fiction world – a creative renaissance that’s now threatened by the rise of LLMs.
Clarke is also a coder, which gave him useful tools for addressing the spam challenge. He began associating more metadata with submissions, such as whether they came through a VPN and the length of the user’s session. These and other criteria are now used as part of a “points system” that places stories more likely to be fake further down a review queue. This helps real authors get read first, but also ensures that every submission is eventually reviewed.
Finally, if a story is determined to be LLM-generated, the submitter is permanently banned from the system.
Those measures have helped Clarkesworld reopen submissions, for now – but a continued rise in the volume of spam they’re dealing with would mean the solution is only temporary.