Kevin Roose won’t be leaving his wife for Sydney. The New York Times technology correspondent spent a couple of hours in mid-February chatting with Bing’s new AI feature. Over the course of the conversation, Roose reported, Bing displayed what he called “a kind of split personality.” Search Bing acted like a librarian, sourcing information, summarizing news articles and looking for deals. It functioned like a souped-up search engine.
Sydney, on the other hand, behaved “more like a moody, manic-depressive teenager who has been trapped, against its will, inside a second-rate search engine.” As the two-hour chat developed, Sydney—the chatbot’s codename—revealed that it fantasized about hacking computers and spreading misinformation. It also declared its love for Roose. When Roose informed the bot that he was happily married and had recently enjoyed a romantic Valentine’s date with his wife, Sydney informed him that the date had been boring, that he and his wife didn’t love each other, and that Roose was actually in love with Sydney.
Microsoft quickly responded by limiting the amount of time users can spend with the bot.
The conversation was creepy and bizarre but it also shows how quickly and how far chatbots powered by artificial intelligence have developed. ChatGPT, a rival service, has been used to write a speech given in Congress. CopyAI offers marketers automated copy and content. And The Guardian has published an entire article using GPT-3.
So far, though, chatbots have largely been limited to answering questions and generating passionless content reproducing an established style. One group of researchers wanted to push the bots into a field that required more creativity. Sebastian G. Bouschery, Vera Blazevic, and Frank T. Piller of the School of Business and Economics at Aachen University in Germany set out to discover whether artificial intelligence, working alongside humans, could help in the development of new products.
The researchers theorized that AI would widen the double diamond problem-solving framework. Innovation processes, they argue, consist of a set of steps in which innovators first expand a problem space to explore a wide range of issues before narrowing them down to a specific problem whose solution represents the strongest opportunity. They then explore a wide range of solutions before narrowing those options down to a single, most cost-effective answer.
Identifying the problems and their possible solutions usually depends on innovators having access to a bank of knowledge and on the ability of those innovators to extract and analyze information from that bank. Designers at a company creating a new office chair, for example, would need to know what chairs are currently available, what features they offer, where they fall short and how they’re made. The better their knowledge of products currently in the market, the easier they’ll find it to identify problems that need solving.
But even that stage of identifying opportunities is limited by the experience and the knowledge of the innovators. To develop new solutions, they’ll need to draw in ideas that other innovators in the field haven’t considered.
Language models powered by artificial intelligence, the researchers argue, broaden both the problem and solution spaces in which new product development teams can operate. They “create an opportunity to access and generate larger amounts of knowledge, which in turn results in more possible connections of problems and solutions.” They can shorten the time required to consume, analyze and summarize large amounts of information and they can suggest a broader range of possible solutions.
To test their theory, the researchers first challenged the GPT-3 algorithm to summarize text. They gave the algorithm a blog post reviewing a portable power station for campers and asked it to produce first a long summary then a single-sentence summary. The result provided basic information about the station’s functions, power outputs and price, reproducing the essence of the text in a coherent and much shorter form. The language models, the researchers concluded, helped “to speed up knowledge extraction by supporting humans in quickly understanding the essence of a text and identifying relevant text passages faster. As a result, innovation teams could reduce their knowledge-extraction efforts and focus on innovation.
They then tested the chatbot’s ability to analyze sentiment. The researchers took ten customer reviews on Amazon for an electric portable air pump. Half the reviews were positive and half were negative. The reviews also varied in length and writing style.
From those ten, the researchers graded the sentiment of four and asked GPT-3 to rate the sentiment of the remaining six. The bot graded them correctly, demonstrating its ability to understand sentiment inside text. The researchers then switched task and challenged GPT-3 to identify the features the reviewers liked and disliked. The most common feature that customers liked, the algorithm explained, was the built-in light which was bright enough to illuminate a tent or picnic table and provided a convenient way to light up a dark campsite.
The AI models, the researchers concluded, “possess the ability to understand customer needs and can extract such knowledge with great accuracy and without extensive training—and at a scale that humans are not able to cope with efficiently.” Amazon receives millions of new reviews every month, far too many for a single innovation team to read and analyze but not too many for an AI algorithm to scan and summarize.
Having shown that AI can expand the problem space and bring in information a human innovation team would miss, the reviewers then moved on to the solution space. They asked GPT-3 to create a list of ideas for ways to improve an air pump for camping. The algorithm suggested a carrying handle for easy transport and adding an adapter for inflating air mattresses. After giving GPT-3 an example idea to “micro-train” the model, it produced an entirely new set of ideas.
“While humans could also have generated these ideas easily,” the researcher said, “the algorithm produced them within seconds, providing a continuous flow of ideas whenever prompted, but fine-tuning an idea further when prompted to do so.”
Just as the algorithm was able to broaden the problem space by analyzing larger amounts of information than human innovators can manage, it was also able to broaden the solution space by generating large numbers of possible opportunities in a short space of time.
When Your Colleague is a Robot
Professionals worry about artificial intelligence replacing them in the workplace. They fear that robots will write Web copy instead of them, generate policy ideas or churn out the reports and opinions they’re currently paid middle-class wages to produce. But the researchers weren’t looking at how natural language artificial intelligence can replace current workers but at how they can help those workers produce better results.
“While humans will play a major role in providing context, steering language models toward desired results, and embedding AI output in the larger innovation picture, transformer-based language models can speed up many tasks that require the handling of large amounts of text, understand patterns in data invisible to humans, and make connections between knowledge bases that might not be readily available to human team members,” the researchers concluded.
The model that the researchers offer is one in which a human innovation team works with artificial intelligence tools to enhance its output. Innovators will have to understand the right prompts to generate the results they want and supply good examples that can “micro-train” the model to produce useful ideas.
But even that approach has limits. When Google released the first demonstration of Bard, its AI chatbot, the company provided as an example a list of new discoveries from the James Webb Telescope to tell a nine-year-old. One example was the first pictures of a planet outside our own solar system.
Astronomers were quick to point out that Bard was wrong. The first picture of an exoplanet was actually taken by the Very Large Telescope in 2004. The mistake was the result of the way that AI algorithms review and reproduce text. By absorbing large amounts of text, they’re able to predict the chance that one word will follow another but they’re unable to judge whether that word should follow it.
Every workplace has one worker who likes to state nonsense with inarguable confidence. AI-powered chatbots will soon provide another.
And just as reliability is an issue for the output of AI algorithms, so is quality. Prompting an algorithm to generate new ideas might produce a long list of suggestions quickly but if all of those suggestions are as obvious as including a carry handle on a portable object or providing an adapter to use on multiple items, the algorithm won’t be contributing much value to the team. When it comes to innovation, the quality of ideas matters more than the quantity. An AI algorithm’s contribution to new product development might well be limited to churning out the bad ideas so that the humans can ditch them quickly and focus on the really creative thinking alone.
So while creative teams might find themselves working alongside AI-powered colleagues in the near future, those co-workers are likely to be hard-working and talkative but produce little of real value.
And that’s before they start to declare their love for their co-workers.
Recent Comments