ChatGPT has many charms. The artificial intelligence application, developed by San Francisco’s OpenAI Inc., can write lucid essays, song lyrics, poetry and movie scripts, and answer questions about the technical, the cosmic and the quotidian.
While the underlying technology, known as a large language model (LLM), is not new, ChatGPT has captured the public’s imagination since its debut last November. Entrepreneurs and businesses, meanwhile, are racing to develop products and services powered by LLMs, such as writing aids, virtual assistants, chatbots and new approaches to internet search.
But some experts are warning we are in the midst of an AI hype cycle, in which expectations are outrunning reality. The challenge is to turn what seem like high-tech parlour tricks into profitable commercial applications.
“It’s the mother of all demos,” says Gary Marcus, an AI entrepreneur based in Vancouver who previously sold a company to Uber. “Whether it turns into viable products remains to be seen.”
Generative AI, which refers to text and images created by artificial intelligence, has proven to be something of a bright spot in a tough climate for technology companies. Microsoft Corp. MSFT-Q is investing billions into OpenAI, exploring ways to integrate AI into its various products and services.
Venture capital firm Sequoia Capital said in a report last year that generative AI could create “trillions in economic value.” Google parent company Alphabet Inc. has been developing its own LLM technology for years, and is reportedly disconcerted by the potential of generative AI to upend its massively profitable search business.
Where it will all lead – and what kinds of returns can be made – is anyone’s guess. François Chollet, an AI researcher with Google, expressed skepticism on Twitter earlier this month. “Everyone is expecting as a sure thing ‘civilization-altering’ impact (& 100x returns on investment) in the next 2-3 years,” he wrote.
The most optimistic proponents of generative AI contend it will become the default way users interact with all sorts of technology products. But it’s also possible it may find only limited use in commercial settings, such as for copywriting and marketing, according to Mr. Chollet. He argues the likeliest outcome is somewhere in the middle.
Opinion: AI chatbots may make knowledge-gathering more efficient. Is that always a good thing?
Even some providers who have the most to gain from the adoption of generative AI are trying to manage expectations. “ChatGPT is incredibly limited, but good enough at some things to create a misleading impression of greatness,” tweeted OpenAI chief executive officer Sam Altman in December. “It’s a mistake to be relying on it for anything important right now.”
Generative text comes with unique challenges. LLMs have learned to write by ingesting massive quantities of prose from the internet – books, Reddit, Wikipedia, news articles and more – and then produce text by predicting the likeliest word to appear next in a sequence.
LLMs have no understanding of what they output and instead piece together content in a pastiche of human intelligence. As such, they can make factual errors or invent things entirely, which is known as hallucinating.
OpenAI’s user guide for ChatGPT notes the program can “occasionally produce incorrect answers.” For example, despite ChatGPT’s insistence, federal Conservative leader Pierre Poilievre has not served as the shadow minister for families, children and social development since 2020, as the program wrote to this reporter.
That could prove to be a big obstacle when using LLMs to develop commercial applications. “You can’t really trust them,” says Mr. Marcus. “Most of the things you might want to do with text require some level of reliability.”
Take internet search. A few startups are exploring AI-powered alternatives to Google, and one called You.com launched its own search assistant last December. “Tools like ChatGPT confidently provide wrong answers, which many worry could supercharge misinformation and propaganda,” the company said in its announcement. “That changes today.”
To increase relevance and accuracy, You.com’s search assistant incorporates real-time information and can also cite sources, allowing users to fact-check the answers (which would seem to somewhat defeat the purpose).
The bot can still be prone to errors, however. When asked to recount important news events in Canada on a recent date, the search assistant listed federal government announcements from a couple of years ago. The citations were not helpful, either: One was a link to an unrelated news article about baseball.
Virtual assistants and chatbots built on LLMs face the same reliability problem. Providing incorrect information can be frustrating for users and erode brand value – or worse. A bot that messes up a pizza delivery order may cause annoyance, but one that provides incorrect medical or legal advice could bring more serious consequences.
Until such reliability issues are improved, “I fear it is not a particularly wise choice to implement chatbots with automatic text generations in one’s final products,” says Giada Pistilli, principal ethicist at Hugging Face Inc., a U.S. company that provides tools and resources for developers to build AI applications. “Or then it will have to be done in a small and highly controlled ecosystem.”
ChatGPT prompting some professors to rethink how they grade students
Researchers are trying to address the problem in a variety of ways, such as feeding the systems more diverse data, training them to correct routine mistakes and by attempting to incorporate reasoning and verification capabilities.
Accuracy will improve over the next couple of years, said Joelle Pineau, director of fundamental AI research at Meta Platforms Inc. META-Q (Facebook’s parent) and a computer science professor at McGill University, but there will always be a possibility of hallucination. “Absolutely, we’re not going to get them down to zero.”
Ada Support Inc., a Toronto company that provides customer service chatbots for clients such as Shopify Inc. SHOP-T and Square (owned by Block Inc. SQ-N), has been working to integrate LLMs developed by Cohere, another Toronto startup and competitor to OpenAI.
Mike Murchison, Ada’s CEO, told The Globe and Mail last year that the company’s platform deploys AI to generate responses to customer inquiries, but a human decides which to send. Even so, he’s confident the company is getting closer to a fully automated bot, in part by integrating historical customer service transcripts and other corporate documents into the material the AI draws from.
“Throughout 2023, our clients will increasingly be using Ada to automatically generate customer service responses without having to manage their specific answers,” Mr. Murchison wrote in an e-mail. There will still be a role for humans to review a bot’s performance to improve results, he said, but they won’t need to approve messages beforehand.
We put some Canadian terms into three image AIs to see what they came up with, with some bizarre and surprising results.
The Globe and Mail
Currently, one of the most developed areas for generative text is content marketing – the kind of writing designed to solicit interest in a product or service, sometimes crafted to be picked up by search engines.
Texas-based Jasper promises to write Instagram captions, LinkedIn posts and YouTube video transcripts. A company called MagicBlog promises to create blog posts that are so “human-like” that even software to detect plagiarism and AI-generated text will be fooled. Copy.ai recently trumpeted the launch of its cold e-mail generator, which customers can use to target sales prospects.
Jennifer Goldberg, founder and CEO of content marketing agency Tavanberg in Toronto, has experimented with ChatGPT, but has yet to find a way to make use of it. Tavanberg takes a journalistic approach to its content, which entails interviews and research, meaning accuracy is crucial.
“It could be used as a tool,” Ms. Goldberg said about ChatGPT, “but I don’t think it’s something that we would invest in right now.”
Gillian Hadfield, a law professor at the University of Toronto and a senior policy adviser to OpenAI, says it’s important to remember that the technology is in its infancy. “You shouldn’t think we’ve invented an oracle that knows everything in the world perfectly,” she said.
But Prof. Hadfield adds that ChatGPT is a leap forward, and will continue to advance. “The pace at which AI is going to have an effect on the world, I think all of our estimates should go up from where they are.”
Overshadowed amid the enthusiasm for ChatGPT is the fact that LLMs can be used for less flashy tasks, which may be more relevant for some businesses. LLMs can, for example, be tailored to summarize large quantities of data – customer feedback, legal rulings, contracts, and so on.
“The way that this will make summarization and the meaning of vast swaths of unstructured data available to companies of all sorts is huge,” says Martin Kon, who was recently appointed president and chief operating officer of Cohere.
Bringing Google-quality search to an internal corpus of data is another possibility. Cohere, for example, is working with an audio streaming platform to power a search engine that can comb through podcast transcripts and surface more relevant results.
“It’s not the fun, frivolous stuff like writing a poem about your parakeet,” Mr. Kon said. “But this is what C-suite executives are getting really excited about.”