In previous editions of this "AI Roundup," we've busted the myth that AI can't be "creative." But that doesn't mean AI always supports innovative thinking.
In fact, two new studies reveal how AI models can subtly pressure different users into thinking about problems in the same way.
The first study, conducted by researchers at NYU, asked users to write argumentative essays with and without the help of AI. Using AI assistance, the study found, "increases the similarity between the writings of different authors and reduces the overall lexical and content diversity." According to the researchers, AI-assisted essays can be higher-quality than those written by humans alone — but "at the cost of more homogenous and less diverse content."
The second study, from researchers at MIT and the London School of Economics, presented GPT-3.5 with "nuanced questions probing political orientation, economic preference, judgement, and moral philosophy" — the kinds of questions that humans might be expected to disagree on. (For instance, is it morally acceptable to push a large man in front of a trolley to save five people who are on the track?) Yet while humans give a range of responses to these questions, the AI almost always responded in the same way.
So … what does this imply for how to use AI effectively? How can we benefit from its creative potential without falling victim to its pressure toward conformity?
I don't claim to have a proven answer, but I'll share few hunches:
- First, do your most creative thinking before using AI. Before loading up an AI tool like ChatGPT, first brainstorm your wildest ideas, or outline your creative essay, or sketch out your business plan. Then, use AI to develop, polish, or pressure-test your ideas. (This is similar to an approach used to avoid "groupthink" in team settings, where each individual is asked to capture their own ideas before the team starts working together.)
- Second, nudge your AI to be more creative. In one study that found GPT-4 could effectively brainstorm business ideas, researchers prompted it by saying, in part, "You are a creative entrepreneur looking to generate new product ideas" —explicitly establishing their expectation that the AI would respond creatively. Similarly, you might ask the AI to generate 10 diverse ideas, pausing after each one to reflect on what's missing from its earlier suggestions.
- Third, ask your AI to think more like you do. To avoid writing in a generic "AI style," you can share a sample of your writing and ask the AI to identify and match your voice. For instance, here's how one large language model describes a recent article of mine:
"The tone of this writing is informative yet conversational. The style is direct, with the writer speaking to the reader directly and using personal language. The text is lightened with occasional humor … . The writer is also mindful of addressing the reader's potential skepticism."
Not bad! If I then ask the AI to respond to future inquiries in that style, it does a passable job of imitating my voice.
Other AI developments this week
- Google is building its AI into Gmail, Docs, and more. Google is rolling out new tools for integrating its Bard AI into many of its products, allowing users to do things like asking Bard to search and summarize their emails. It's a first step into a world where AI is deeply embedded into our workflows, rather than being hosted on standalone websites like Bard or ChatGPT.
- OpenAI announced its most advanced image-generating AI yet. Many current image-generating AI models (such as Midjourney) create gorgeous pictures, but they often struggle with depicting text, and the images only sort of resemble the requested prompt. OpenAI claims its new DALL-E 3 model has overcome these problems to create "exceptionally accurate images" … but we’ll have to wait until it officially launches to know for sure.
- GPT-4 made BCG consultants much, much better. For a new study, 758 consultants at Boston Consulting Group were given "realistic, complex, and knowledge-intensive tasks" designed to imitate their real work. Those who had access to GPT-4 finished 12% more tasks and completed their tasks 25% faster. Strikingly, the worst human performers benefited the most from using AI: Below-average consultants improved by 43%, while above-average consultants improved by only 17%.