Can AI decision-making emulate human reasoning?

One thing researchers know for sure is that AI technologies can emulate the decision-making capabilities of humans with increasing fidelity. A recent study in Nature¹ demonstrated that a large language model (LLM) can be fine-tuned to make decisions similar to most humans. After training the model on a set of historical data from 160 psychological studies (comprising over 10 million individual decisions), the researchers then exposed the model to new problems and found that it made the same decisions as humans more often than previous cognitive models. The study authors conclude that their model, which they named “Centaur” after the mythical hybrid beast, can be “used to improve our understanding of human decision-making.”

Understanding human decision-making is one thing; guiding, supplanting or augmenting it is another. A burgeoning field of research has taken that question up in recent years, stretching back to before the launch of ChatGPT and other generative AI tools led to widespread AI adoption. 

“AI decision making,” so-called, may be too broad a term to be useful. In most cases that have been trialed, AI systems have offered decision support, with human oversight and human judgment reigning supreme. In such cases, AI performs data analysis and provides a recommendation, helping to streamline the process of arriving at an informed, data-driven decision. 

The results of such real-world trials have been mixed. In one instance², researchers equipped a judge with an AI system that provided risk assessments on an offender’s likelihood of recidivism. The experienced judge rejected the AI’s judgement 30% of the time, and the researchers ultimately determined that the AI was overly harsh.

In other use cases, though—as in healthcare—the introduction of LLMs and machine learning has demonstrated a potential to save lives. For instance, a 2024 UCSD study³, showed that deploying a deep-learning sepsis alert was associated with an in-hospital mortality drop from about 11% to 9%. (The user experience built around the AI system also matters greatly; a previous, similar study⁴ illustrated the importance of automated real-time alerts attached to such a system.)

In the high-stakes worlds of medicine and criminal justice, AI algorithms are for now best restrained to the role of helping humans make informed decisions. In many applications, there is no substitute for human intelligence. But in lower-stakes industries, AI decision-making is already the norm. The field of predictive analytics, for instance, provides a low-stakes example of AI’s ability to incrementally improve the customer experience of a product: just consider how retailers like Amazon automate suggestions of which other products you might buy. Ultimately, a human being could manually suggest these other products, but that would hardly be a recipe for operational efficiency, and ultimately Amazon’s systems can parse large datasets for patterns faster than any human. AI systems, neural networks, and deep learning algorithms are expanding into many other fields such as supply chain management, pricing decisions and demand forecasting.

Will there come a time when AI can support—or supplant—the executives who make big, strategic business decisions? According to a recent study from IBM’s Institute for Business Value, over two-thirds of executives surveyed called “improved decision-making” the top benefit of agentic AI systems.

Still, it’s unlikely that these executives are relying on the outputs of AI systems to automate the highly strategic decisions made in the board room. A 2023 podcast from McKinsey outlines six ultimate stages of AI-for-strategy. These are, in increasing order of complexity:

1. Simple analytics (essentially, a form of data science)
2. Diagnostic intelligence
3. Predictive intelligence
4. AI advising actions to executives (problem-solving)
5. delegating decision-making to AI, with supervision (AI-assisted decision making)
6. Fully autonomous AI

The McKinsey conversation notes that at the moment, the first three are types of AI decision-making that are widely available and trustworthy, with the next three levels “taking time to develop.”

With the power of AI increasing every day, it may be coming sooner than most think.