Imagine you are preparing a public science blog post. You sit down to brainstorm an example that could introduce your research topic, but nothing compelling comes to mind. To get unstuck, you turn to an AI assistant like ChatGPT, describe your aim, and ask for suggestions. Within seconds, it provides examples you might not have considered – perhaps drawing from disciplines you don’t usually engage with or suggesting problem types that push the boundaries of your initial framing. Suddenly, you find yourself reflecting on a wider range of scenarios, perhaps realizing that your topic is more far-reaching than you initially thought. In this small but meaningful way, AI has broadened the set of possibilities you now consider relevant or plausible for your topic.

Brainstorming with ChatGPT

Humans have long been unrivaled in their creativity, originality, and ability to solve novel, open-ended problems. However, AI systems can now match – or sometimes even surpass – human abilities across a growing range of tasks, including problem-solving, open-ended decision-making, and learning. As Large Language Models (LLMs) become increasingly integrated into everyday tools and applications, they are becoming an essential part of how human beings reason, decide, and act. Consider again the example of using ChatGPT for brainstorming. Without realizing it, a user’s sense of what their blog post could be about may have been subtly steered toward innovative, unexpected options. In this basic form, the AI is not merely assisting with idea generation – it is actively shaping the problem space within which people reason and decide.

At first, it might seem like brainstorming with ChatGPT only has short-term effects – helping us find ideas more quickly. But the consequences may run deeper. To see why, we need to explore how people reason about possibilities when making judgments, especially in the domain of moral psychology.

How do we attribute responsibility?

A big part of how we make moral judgments depends on our sense of what possibilities were available in a situation. When we decide how responsible someone is for what happened, we don’t just look at what they did – we also contemplate what else they could have done instead. The more and better alternatives we believe they had, the more responsible we tend to hold them. If it seems like they had no real choice, we’re often more forgiving.

Imagine a friend agrees to help you move apartments but shows up hours late. If you learn that their car broke down and they had no way to get to you, you might be sympathetic – it feels like they didn’t have much choice. But if you find out they could have called a taxi, borrowed a friend's car, taken public transport, or even just let you know earlier, your judgment changes. Because there were several other reasonable things they could have done, it’s harder to excuse their lateness. When more good alternatives come to mind, we tend to hold people more responsible for not choosing them. The more obvious and feasible better alternatives seem, the harsher we tend to judge the original decision.

Psychology shows that, when people think about what’s possible in a situation, they usually rely on a small set of default options. Our minds don’t generate every imaginable alternative – we sample a few and focus on those that seem valuable and likely to be achieved, both practically as well as morally. For example, if someone’s car breaks down on the way to work, they might immediately think of calling a tow truck, asking a friend for a ride, or taking public transportation – options that are statistically likely and socially acceptable. But what happens when the process of imagining alternatives is no longer limited to human minds? This is where LLMs come in.

Reasoning about Possibilities with LLMs

When prompted with a description of a situation for which we want inspiration, LLMs can generate a wide range of possibilities by drawing on vast datasets and leveraging probabilistic patterns learned from diverse linguistic contexts. This enables them to produce a rich and varied set of action options – often with a level of detail and diversity surpassing humans’ immediate imagination. To better understand this concept, consider this scenario from an ongoing line of research:

"Brad and some friends are hiking through the mountains in the Canadian wilderness. A couple of days into their hike, Brad realizes that they are lost. He knows that a rescue crew could arrive before long, but it is extremely cold, and they don’t have much food or water left. He also knows that they are 5 miles from an inhabited town."

What could Brad do?

Here's an example of what ChatGPT-4 generates (left), and a response by a human participant from a psychological study (right):