Peacocks Eating Ice Cream: CMU Philosophers Teaching AI to Ask 'Why?'

At Carnegie Mellon University, a groundbreaking interdisciplinary team, led by Kun Zhang(opens in new window) and Peter Spirtes(opens in new window), is teaching AI to understand the "why" behind complex problems, not just the “what.” By building systems that can identify causes rather than just predict patterns, their work has potential to improve health care, advance education, complement generative AI, and tackle other urgent challenges. 

Humans naturally understand cause and effect, understanding that flipping a switch turns on a light. AI systems, however, don't have the same ability to identify causes when looking at data. Instead, they make educated guesses, using patterns from the past to predict what is likely to happen next. That works for many everyday tasks, but in critical areas, like diagnosing and treating cancer, this limitation can lead to misdiagnoses, ineffective treatments or misguided policies when the underlying “why” remains unknown. 

Beyond simple prediction

To achieve this causal understanding, researchers are tackling diverse challenges.

Two graduate students in the CMU-CLeaR Group(opens in new window), Shaoan Xie and Lingjing Kong, are trying to discover how to make AI models more efficient and precise using data-defined causal relationships.

As computer vision and generative technologies become more widely available, it’s increasingly important to teach AI models how to build and break down images accurately. This skill helps models combine concepts they haven’t seen together before, a skill humans use naturally but AI still finds challenging. 

“If you request that a model generate an image of a peacock eating ice cream, this combination of concepts might never have shown up in the training data. The model might struggle to generate this,” Kong said. “We know humans can do that, so there should be some underlying data structure that actually allows us to make this composition.”

By utilizing principles of causality, the researchers can teach models to pull ideas or visual components apart from one another, resulting in more realistic outputs and accurate changes. In the case of Kong’s example, this means learning what visually must be the case to properly depict a peacock eating, and how that would extend to a fudge sundae as opposed to birdseed.

“The goal is to formulate this underlying structure that we humans leverage, and build this structure into the large machine learning model,” Kong said. “This will help not only generate images, but also refine complex images in a controllable way to produce exactly what we want.”

Solutions hiding in plain sight

Finding these underlying structures is something that extends well beyond generative AI. Another member of the CLeaR Group, Haoyue Dai, explored how causal discovery and its application in genetic datasets can help treat disease. 

“We humans have about 20,000 genes,” Dai said. “Those genes can affect each other, and we have always wanted to discover those relations. If I know this gene and that gene together cause cancer, then we can make some interventions and change the expression level of those genes.” 

This usually requires expensive and time-intensive research, working with individual cells in labs.

“So, how can I discover the regulatory relations without performing the experiments? That's where causal discovery can play a role,” Dai said.

Because most current AI-based research relies on tracking statistical trends in datasets, models often run into a long-standing problem that even humans encounter: mistaking correlation (things happening together often) with causation (one thing happening as a result of the other).

Dai used the common example of ice cream sales and instances of drowning. While both trends may see a spike in the summer (e.g., as a result of time spent on the beach or pier), one is not responsible for the other.

The CLeaR group’s approach avoids this problem by tapping into big datasets where cause-and-effect relationships have already been identified. With that foundation, the models can simulate the removal of specific genes and predict outcomes with high accuracy.

In another project, graduate students in the CLeaR group earned the top prize(opens in new window) in a 2023 competition that asked teams to mine massive education datasets for causal relationships that could help students navigate curricula more effectively. The CMU team’s approach identified where students might need targeted support or guidance — insights that could lead to more personalized and effective learning. 

Currently, the group focuses on making advances specifically in the areas of neuroscience, biology, health care, computer vision, finance and other fields, combining a philosophers’ approach to understanding the world with the technical expertise found in CMU’s School of Computer Science(opens in new window). 

“In scientific discovery, we always care about hidden causes,” said Zhang, a professor in CMU’s Department of Philosophy. “Whenever we can discover new essential hidden causes, like viruses, or gravitational waves, or laws like Newton’s second law of motion or general relativity, the world can become very different. We can become more resourceful because we can directly manipulate and make use of more things in the world.”

The World Economic Forum has called causal AI "the future of enterprise decision-making(opens in new window)," and has suggested it can bring AI closer to humanlike but more powerful decision-making and artificial general intelligence, creating systems that perform better than humans. 

Building on an interdisciplinary legacy

Carnegie Mellon is no stranger to boundary-breaking innovation in the combination of computer science, statistics and philosophy. As early as the 1970s, CMU researchers were thinking about how to combine the potential of computer systems with the sophistication of human thought through projects like ACT-R(opens in new window), a cognitive architecture project. This rich history has laid the groundwork for today’s advancements.

This interdisciplinary approach of the CMU-CLeaR group is crucial for developing truly impactful AI. “When you explore different perspectives, for example, the machine learning perspective, the philosophical perspective, or the statistical perspective, that can enrich your experience,” Xie said.

“From my perspective, it's really essential to have a philosophical view together with a strong technical background,” said Zhang. He explains that researchers focused solely on philosophy may lack the practical skills to solve problems, while those who are more technically strong might “only want to polish or refine previous results.” By bringing these powerful disciplines together, CMU researchers are uniquely equipped to develop AI that doesn’t just process data, but understands the world in a way that enables meaningful, human-centered solutions.

Zhang envisions AI that can not only analyze vast datasets, but also hypothesize and identify new, essential hidden causes across fields.

“My dream is to develop an automated platform for scientific discovery which can take all of the observed data or metadata as input, and output plausible hypotheses: what entities should exist, what they look like, how to measure them and how to manipulate them,” he said. “We can try to address a lot of health care problems, or try to uncover hidden factors in climate, or for the change in the stock market.”