A handbook to Manitoba mushrooms sparks concerns about AI. But for technology, there's no field guide
Gavin Boutroy, Arturo Chang | CBC News | Posted: October 13, 2024 1:14 PM | Last Updated: October 13
No surefire way to tell whether a text was written using artificial intelligence: experts
Alexandre Brassard says culinary delight and a love of the outdoors aren't the only reasons to get interested in mushroom foraging.
"These are organisms that have been under-studied," he told Radio-Canada in French. "From time to time, these very mysterious organisms crop up and you want to learn more about them."
Brassard — dean of science at St. Boniface University — has been happy to share his fungi fascination with fellow Winnipeggers, teaching a course on how to identify them for amateur mycologists. But for years he's been frustrated by the lack of a single, beginner-friendly field guide to local Manitoba mushrooms.
Or at least that's what he thought.
"A student was telling me, 'You say there isn't a book about mushrooms in Manitoba, but no, there is. I saw it on Amazon yesterday,'" he said.
"It was $20. I told myself, 'Why not?' So I bought it. The book is very strange."
The item description for Mushrooms of Manitoba: A Fungal Odyssey through the Heart of Canada said the book "elucidates the vibrant hues, intricate contours and intriguing adaptations of the numerous mushroom species Indigenous to Manitoba."
But Brassard believes it's the work of artificial intelligence.
"There are some turns of phrase that are a bit weird," he said. "For example, at one point it says 'Armillaria mellea is among the most lethal imitations.' That makes no sense. 'Imitations' of what? And it isn't 'lethal.' It's edible."
He said the book goes over 15 poorly chosen species, some whose names aren't up-to-date or which don't even grow in the province.
But what's worse, Brassard said, is what the book fails to mention.
"There are many species of deadly Amanitas that aren't mentioned but should be in a book for beginners.… It doesn't talk about Lepiotas, which are deadly," he said.
"It doesn't give you a good sense of the danger and the species that one can find in Manitoba."
David Gerhard is the head of computer sciences at the University of Manitoba and has been researching AI for decades.
He said the book is a "dangerous" illustration of how chatbots like ChatGPT have led to a flood of AI-generated content pushed out by people eager to "make a quick buck" without much care as to what they're selling.
"The only person who gets hurt is the person who buys the book, either by wasting the money [or] by acting on the information in the book. That is, incorrect information," he said.
That includes online marketplaces like Amazon, which Gerhard says has been a boon for amateur writers, but which can be easily exploited by people generating book-length texts with the push of a button.
"When I buy a book, I have this expectation that a person wrote it, spent some time making sure the information was correct, an editor looked at it," Gerhard said. "None of those systems exist for this kind of one-off, self-publishing thing."
Amazon removed the listing after Radio-Canada reached out for comment.
"We have content guidelines governing which books can be listed for sale, and we have proactive and reactive methods to evaluate the content in our store, whether AI generated or not," Amazon spokesperson Tim Gillman said in a written statement.
CBC was unable to find the book's listed author, Jay O. Mark. CBC was unable to find any information about the book in various publicly accessible International Standard Book Number (ISBN) databases, and the book itself is self-published.
'The good Terminator'
An analysis of a sample of the book by Originality.ai — a company which advertises 98 per cent accuracy detecting AI-generated content — found the book was "Likely AI-generated" with 100 per cent confidence.
That prediction was also generated with AI.
"We kind of think about it a little bit like it's the good Terminator'," Originality.ai founder and CEO Jonathan Gillham said.
"What AI is really good at is identifying patterns in noise within large sets of data, and that's exactly what our AI detector does."
The tool has been trained to spot differences as it sifts through data sets of millions of human and AI texts in order to predict the likelihood that a machine wrote it.
Gillham said AI detectors are never going to be perfect, but that the tool has, itself, been thoroughly tested and that he's confident about what it can do.
Gerhard is skeptical.
"There have been companies that have tried to make AI-detection tools and then somebody will take the Declaration of Independence and put it in and the bot will say, 'Oh yeah, this is very clearly written by a bot,'" he said.
"There are a lot of circumstances where it's difficult or impossible to tell when a block of text is AI-generated."
Tech can 'get things wrong'
But he added there are some tells people have started to catch on to, with the caveat that even some slight tweaking can throw people off the scent.
The base text is "a little bit bland. It tries to say a lot without saying anything," Gerhard said.
"The end result is the average text of all of humanity, right? The kind of thing that the average person would say, which almost by definition isn't very interesting."
Not to say the technology is not without its purposes: Gerhard said he uses AI all the time, helping him do things like organizing files and brainstorming.
He's even used it in mushroom picking.
"I've used it to help me decide how to preserve mushrooms that I pick," said Gerhard, who got into the hobby a couple of years ago.
"I'm a strong proponent of these tools," he added.
"I think the big disconnect comes because as a society, we're used to computers getting things right, and we need to build a new mental model where we have a powerful computational tool that gets things wrong from time to time."