Asking GPT4o to “write an essay evaluating the life of Thomas Jefferson” results in seconds in a highly credible, well-structured, 6-page document outlining this historical figure’s life and impact. This demonstrates one of many provocative linguistic capabilities of generative AI (genAI), which has emerged over the last few years. These systems can quickly and effortlessly produce voluminous amounts of text on our behalf, directed by our needs at work and at home.

Writing, however, is about more than just content creation. As Jane Rosenzweig at the Harvard College Writing Centre put it, “we figure out what we think through writing”. Writing can be seen as a precursor to, a substrate for, and a by-product of, thinking. It can even be seen as a means of learning (the “Writing to Learn” paradigm). Through the process of writing, insights and ideas emerge that we would not have had otherwise. By externalizing our thought, we crystallize what is on our mind, making what we are thinking clearer and more explicit. Creating this external representation also gives us something we can iterate over.

Writing enables us to think about what we think, and craft a way to say it. If metacognition is the process of “thinking about thinking”, writing is metacognition in action. Writing is one of the oldest and most durable “tools for thought”.

If generative AI is now doing our writing, and writing is a process through which people think, then when and where in the writing process will our thinking now happen? As a research objective, how might we better understand the impact that AI systems are having on human cognition, on the processes through which we think, learn, and understand, and what might we do about it? 

Understanding the consequences of outsourcing cognition to genAI is an urgent and important area of study in the domain of Responsible AI (RAI). The integration of these technologies into our tools is already underway and accelerating. The importance of this issue cannot be understated, as it has ramifications that span “formal” thinking environments like schools and universities, to informal ones, like the workplace and home.

While we have these concerns, we are optimistic that genAI can offer solutions to the issues that it creates. It has the potential to complement, enhance and augment people’s cognitive capabilities in meaningful ways. For example, while Large Language Models (LLMs) can swiftly respond to requests by providing elaborate outputs, they are also able to ask meaningful questions to help us explore and consider something new. Leveraging these kinds of capabilities, genAI can help us think more deeply, communicate better, and solve larger, harder problems than we currently can.

How might we take this approach when writing the article on Thomas Jefferson that we described at the outset? Rather than writing the essay on our behalf, genAI could instead support us in approaching the task with a more critical eye. It could push us to think of alternate perspectives on Jefferson’s life, for example, or to be more curious about the accounts written by others, thinking about their motivations and perspectives.

It is essential that we create a world in which these new technologies do more than improve our productivity. They should make us better thinkers than we would be without them. We believe the time is right for the creation of a new generation of ‘Tools for thought’. These will unlock our capacity for creativity, provoke us to deeper insight, and ultimately accelerate the ways through which we learn and understand.