Skip to Content Skip to Menu Skip to Footer

This is a thought leadership report issued by S&P Global. This report does not constitute a rating action, neither was it discussed by a rating committee.

Highlights

Causal AI aims to transform AI from a predictive tool to one that can explain events and solve problems by understanding the relationship between cause and effect, known as causality.

Causal inference has its roots in philosophical ponderings (Aristotle, David Hume) and statistical methods (Robert Fisher), with significant contributions from Turing Prize winning computer scientist Judea Pearl.

Causal AI has diverse applications across sectors and is expected to integrate with established AI models, playing a crucial role in the potential development of AI that is capable of human-level cognition, known as artificial general intelligence.

From prediction to reasoning

Greater consumption of ice cream leads to more drownings, while global warming has put an end to piracy. Both statements might seem unlikely, yet a superficial study of historical trends suggests they could be true. Astute readers will, of course, be reminded of the phrase "correlation does not imply causation." They might also conclude, rightly, that warm weather encourages both a desire for ice cream and swimming, while piracy's heyday was an artifact of pre-industrial (and thus pre-global warming) trade dynamics and weak maritime policing.

These examples highlight the potential danger of assuming that correlation and causation are necessarily linked. Just because two trends move in tandem doesn’t mean one causes the other.

That is an issue for machine learning models, which underpin most of the AI with which the world interacts. These models excel at probability-based predictions derived from historical data, meaning that correlation between factors in the data will often be expressed in models' output without consideration for why the link exists (or only seems to exist).

Causal AI seeks to overcome this by deploying models that transform AI from a tool that makes predictions based on data to one that can explain events and solve problems by understanding causality--the relationship between cause and effect. That is fundamental to the advancement of today's AI systems, which, while powerful, often struggle to offer insightful and useful results when posed problems that fall outside the scope of their training data, or when making decisions in complex, real-world scenarios that require common sense.

The potential to change that paradigm has propelled the process of inferring causation from observed data, known as causal inference, to the center of ongoing efforts to develop AI's capabilities.

A brief history of causal inference

Understanding the history of causal inference, from early philosophical ponderings to causal AI, provides insights into its strengths, weaknesses, and applications.

Philosophical foundations: Aristotle to Hume

Causal inference's (recorded) roots date back to Aristotle's contemplations on causes and effects. The ancient philosopher discussed four types of causes: material (what a thing consists of), formal (the arrangement of the thing), efficient (also known as agent, and referring to an external source of change), and final (the natural end state of a thing). This framework laid the groundwork for later discussions about causation and stressed the complexity inherent to understanding why things happen.

In the 18th century, David Hume explored the nature of causation. He posited that causation is not directly observable; instead, we infer it from the regular succession of events. Hume emphasized the importance of regularity and patterns in establishing causal relationships and argued that correlation alone can’t prove causation. His skepticism prompted later philosophers and scientists to seek more rigorous methods for establishing causal relationships.

Statistical methods emerge

In the 20th century, statisticians began to formalize methods for analyzing data. Ronald A. Fisher introduced concepts such as randomization and experimental design in the 1920s, laying the groundwork for causal inference in controlled experiments. As statistics evolved, the focus on correlation versus causation became prominent. The idea that correlation can indicate a relationship between variables but does not imply that one causes the other became increasingly important.

Counterfactuals and potential outcomes

In the 1970s, Donald Rubin developed a formal approach to causal inference based on a framework of  potential outcomes (that would come to be known as the Rubin causal model). The framework emphasized the importance of considering what would happen under different conditions (counterfactuals) to establish causality.

Judea Pearl and the causal revolution

In the 1990s and early 2000s, Judea Pearl, a computer scientist, developed diagrams and mathematical frameworks to model cause-and-effect relationships. The diagrams (called Directed Acyclic Graphs, or DAGs) visually represent causal relationships and help clarify assumptions about causality and the identification of potential confounding variables. Pearl's "do-calculus" provides rules for reasoning about interventions and causal effects that facilitate the derivation of causal conclusions from observational data.

Pearl went on to create the "Ladder of Causation," which proposes a hierarchy of causal reasoning (see figure 1). And he argued, in "The Book of Why" (Pearl and Mackenzie, 2018), that science and AI require a better understanding of causation to progress, noting that data alone can't provide causal conclusions without the inputting of causal assumptions (“no causes in, no causes out”).

More power, data, and applications

Today, causal inference is a vital area of research across various fields, including epidemiology, economics, social sciences, and artificial intelligence, as researchers seek to understand and quantify causal relationships in complex systems. Recent years have seen an explosion of interest in causal inference, driven by advancements in computational power and the availability of large datasets. Developments notably include new techniques to address the challenges inherent to observational studies. Those developments include propensity score matching (used to construct artificial control groups), instrumental variables (used to estimate causal relationships when controlled studies aren't feasible), and regression discontinuity designs (a framework to determine the causal effects of intervention).

Causal inference and causal AI: the underlying techniques

Causal inference, and the techniques it uses to discern causation, have emerged as key drivers of AI's capabilities (see figure 2). An understanding of the key elements underlying causal inferencing is thus useful to understanding the key techniques used in causal AI (see figure 3), much in the same way that mathematics and logic (philosophy) are intertwined in programming.

While established AI and machine learning (ML) tend to focus on finding patterns and correlations in data, causal AI seeks to understand why things happen. In doing so, it is advancing AI from observation, "when A happens, B follows," to the deeper question of, "does A cause B?"

There are several key differences between traditional machine learning and causal AI (see table 1).

Applications of causal AI

Many of the benefits of causal AI become clear when we look at how it can be applied across different sectors.

Education and training

Causal AI's ability to intuit what works best could improve education and workforce training by facilitating a more nuanced approach, which dispenses with a one-size-fits-all model. A deeper understanding of why certain strategies are more effective will support the personalized development of human potential.

  • AI tutoring could employ causal inference to deploy teaching methods and content tailored to individuals based on an understanding of what actions will deliver maximum benefit rather than relying on simpler correlations, such as study time to test scores.
  • Educators and policymakers could use causal AI to assess the benefits of interventions. For example, training programs for a school district might be approved following a causal AI simulation that tests students' likely test improvement following a training program to improve teacher competency.
  • Employers could deploy causal AI to analyze which skill-development workshops will most improve employee performance or productivity, enabling investment to be allocated where it makes the greatest difference, rather than to foster skills that are typically correlated with good employees but that may not be causally connected to positive work outcomes.

Healthcare

Distinguishing correlation from causation in healthcare can save lives. Applications of causal AI will be myriad and lead to fundamental improvements in health outcomes.

  • Health risk assessments could be improved by the replacement (or enhancement) of traditional predictive modeling with causal AI. Traditional models tend to identify high-risk patients based on traits correlated to disease, e.g., heart disease is associated with a sedentary lifestyle, high cholesterol, a family history, etc. Causal AI can identify individuals' key factors, resulting in personalized prevention strategies. For example, a causal model might reveal that, for a given patient's risk of developing heart disease, a lack of exercise is the greatest risk factor, while cholesterol is a minor issue, given other contributing factors.
  • Simulations of medical interventions using causal AI can predict individual health outcomes resulting from changes to medications, biomarkers, or surgery. "What-if" clinical trials are powerful decision-making tools that allow physicians to compare outcomes prior to an intervention.

Drug discovery

Pharmaceutical researchers are using causal AI models to better understand diseases' complexity and design trials. The result is accelerated, and thus less costly, drug discovery and development cycles that produce more effective medications.

  • Causal AI can offer an understanding of the molecules, genes, and pathways that cause disease progression, rather than simply revealing those associated with a disease. That aids in the identification of promising targets for new drugs, e.g., inhibiting a particular protein to arrest tumor cell growth. This focus on causal targets offers drug developers the opportunity to design therapies that have a higher chance of success.
  • Causal AI can be used to improve the design of clinical trials, e.g., by identifying which patient subgroups are most likely to positively respond to a treatment (while accounting for confounding factors). The models can also simulate different trial designs to reveal which will most clearly show a drug’s effect.

Social policy and government

Governments and policymakers are increasingly interested in using AI not just to predict policy outcomes but to understand the causal impact of interventions.

  • Causal AI models are well-suited to predicting the causal effects of policy. For example, consider a social policy question: "If we increase the minimum wage, what will happen to employment rates?" A correlational analysis might note relationships in historical data, but a causal model tries to account for confounding factors and simulate the effect of the policy change itself.
  • Policymakers can use causal models for virtual "what-if" experiments across areas from education to public health. The ability to foresee effects (intended and unintended) before making decisions can prevent costly or harmful policy mistakes.

Financial services

Causal AI can help make sense of the enormous amount of data generated by financial services, and in ways that traditional analytics cannot.

  • Banks and insurers have long used statistical models to predict risks (such as the likelihood of a loan default or an insurance claim), yet causal AI can enhance that process by identifying underlying drivers of risk. So, instead of saying "customers are likely to default because they resemble past defaulters," a causal model could suggest the combination of factors that causes defaults. Knowing that would allow institutions to take proactive measures (such as offering loan restructuring or financial counseling in areas where unemployment spikes).
  • Causal AI can also analyze how different economic scenarios (interest rate changes, regulatory changes, pandemics, etc.) will impact portfolios and balance sheets ahead of a change, helping firms build more resilient risk management strategies.
  • In fraud detection, causal AI can enhance traditional systems, which look for patterns in transaction data that match past fraud cases. For example, a causal analysis could identify what causes a transaction to be flagged as fraudulent and add contextual information, thus differentiating between fraud and coincidental anomalies (such as purchases at an unusual time by someone traveling in a different time zone).
  • A focus on the causes of fraud and the conditions that enable fraud to occur enables businesses to respond more effectively and proactively. This could mean identifying a specific system vulnerability that, when fixed, prevents fraudulent activity, rather than merely catching fraud after the fact.

Manufacturing

In manufacturing, causal AI could be used to optimize processes, enhance maintenance, and strengthen supply chains.

  • Simulations can measure the quality or yield benefit (or lack thereof) of changing a production line process or upgrading equipment.
  • Securing greater insights into the causes of defects or delays allows factories to implement the most effective fixes.
  • In predictive maintenance, causal AI can identify why a machine will fail (e.g., due to vibration that fatigues a particular part) and what can be done to avoid failure, rather than providing a simpler statistical prediction of when failure will occur.
  • Causal AI's "what-if" capabilities could enable supply chain management testing for key weaknesses by comparing disruptions' effects, including downstream impacts on inventory and production. Companies can then develop resilience strategies (such as alternative suppliers or safety stock).

Retail and customer insights

Causal AI will help retail and marketing functions understand why customers behave as they do, providing insights that are deeper than surface-level patterns.

  • Traditional analytics typically told retailers that customers who bought item A often bought item B; causal analysis could answer: does purchasing A cause interest in B, or are they bought together for another reason?
  • Retailers are using causal AI to guide marketing spend toward optimized pricing, promotions, and product recommendations that increase sales, discerning them from promotions that simply coincide with shopping sprees.

Opportunities and challenges

As discussed above, causal AI's development brings a host of opportunities. But wider adoption and increased capabilities will also come with challenges to be managed and other considerations that should be addressed (see table 2).

Key trends to watch: The next five years

Causal AI's rapid shift from the theoretical to practical will ultimately change how artificial intelligence works with data, makes decisions, and mimics human reasoning. Looking toward that future, there are already a handful of promising research paths and applications.

Causality and artificial general intelligence (AGI)

AI's capability to think like humans across a broad spectrum of tasks (known as AGI) will require an understanding of not just what happens, but also why it happens. Future AI models will be built to ask questions about different possibilities and to create models that explain how one condition causes another, thus providing the capacity to make better decisions in unfamiliar situations. The addition of causal reasoning to LLMs (such as GPT) should lead to models that provide answers and consider events in a causal way, allowing them to make fact-based predictions using different variables.

Decision support

Causal AI promises to become a significant decision-making tool for businesses, which can use it to theoretically test the outcomes of different actions, e.g., "What would happen if we increased prices?" Its use is likely to become standard in forecasting and planning.

Causal identification and hypotheses generation

The AI should offer increased capability to identify, or hypothesize, causes based on historical data. This could greatly impact fields such as scientific research, public health, and economics. For example, an AI-based medical research assistant could look at genetic and clinical data and propose connections between lifestyle choices and disease risk, potentially speeding discoveries that might otherwise have taken many years.

Integration is the next step

Looking ahead, we should expect established AI models to increasingly integrate causal inference technology. Different models each have their strengths; for example, deep learning excels at recognizing patterns and representation learning, while causal inference focuses on reasoning and sound decision-making.

Such combinations will likely take AI closer to human-like thinking (AGI). Indeed, generative AI, which can rapidly analyze and create information, is often compared to the human mind's fast and intuitive thinking, known as "System 1" thinking, while causal AI's reasoning is thought of as our slower and more logical "System 2" thinking.

Researchers are developing hybrid models, such as neural networks that produce causal graphs, and language models that include causal rules. These combinations promise to deliver AI systems that are more innovative and effective, while also being more reliable, safer, and understandable. Technical issues and ethical concerns remain, but the direction of travel is clear.

Explore more research on Artificial Intelligence from S&P Global

Read More

Content Type

Special Reports

Research Council Theme

Digital Disruption

Contributors: Cat VanVliet and Paul Whitfield