Can AI systems be trusted to make unbiased decisions? Scientists found out

AI is often praised for being fast, efficient, and objective. But when it comes to non-biased judgment and decision-making, can AI really outperform a human?

Not always, says a new study. In fact, AI can fall into the same traps that we do – overconfidence, risk avoidance, and other mental shortcuts – especially in situations that aren’t purely logical.

A multidisciplinary team of scientists came together and tested ChatGPT, a popular AI developed by OpenAI, and found that while the tool excels at math and logic, it shows human-like biases in subjective scenarios.

Understanding human bias – the basics

Human bias comes from the brain’s need to make quick decisions. Our brains are wired to take mental shortcuts – called heuristics – to save time and energy. These shortcuts help us make sense of the world fast, but they’re not always accurate.

For example, if you’ve had one bad experience with something, your brain might unfairly generalize that feeling to everything similar. It’s like your brain goes, “Hey, last time this happened, it sucked – let’s avoid it altogether,” even if the current situation is totally different.

Bias also shows up because of how we’re shaped by our environment – family, culture, media, even past experiences. Over time, we build mental filters that influence how we see people, situations, or information.

Confirmation bias, for instance, makes us favor info that supports what we already believe and ignore what doesn’t. The science behind bias isn’t just about flaws – it’s about survival, speed, and pattern recognition.

Same mistakes, different mind

The researchers ran ChatGPT through 18 classic tests designed to catch biased thinking.

These weren’t simple problems of logic – they were tricky decision-making scenarios where people tend to make predictable errors. The results were surprising.

In nearly half the tests, ChatGPT made mistakes similar to those made by humans. It showed signs of overconfidence, ambiguity aversion, and even the gambler’s fallacy.

Yet, in other areas, it behaved differently. For example, it didn’t fall for the sunk cost fallacy or ignore base rates like people often do. This shows that while AI can mimic human flaws, it doesn’t replicate them across the board.

Interestingly, the newer GPT-4 model, while more accurate on math-based problems, showed even stronger biases in scenarios requiring judgment. This suggests that as AI improves in one area, it may regress in others.

“As AI learns from human data, it may also think like a human – biases and all,” said Yang Chen, lead author and assistant professor at Western University.

“Our research shows when AI is used to make judgment calls, it sometimes employs the same mental shortcuts as people.”

Where AI falls short

One clear takeaway from the study is that ChatGPT tends to play it safe. It often avoids risky options, even when taking a calculated risk could lead to a better outcome.

This cautious approach might make AI seem more reliable, but it can also limit potential gains in complex decision-making scenarios.

The AI also has a tendency to overestimate its own accuracy. It presents its conclusions with a level of confidence that can be misleading, especially when the correct answer isn’t clear-cut.

Additionally, ChatGPT tends to seek information that confirms what it already “believes,” a phenomenon that is often seen when humans gather evidence, and known as confirmation bias.

Another trait observed was the AI’s preference for certainty. When faced with ambiguous options, it gravitated toward the choice with more clear-cut information, even if the ambiguous one held greater potential benefits. These behaviors reflect a very human-like discomfort with uncertainty.

“When a decision has a clear right answer, AI nails it – it is better at finding the right formula than most people are,” says Anton Ovchinnikov of Queen’s University. “But when judgment is involved, AI may fall into the same cognitive traps as people.”

Why does any of this matter?

These findings come at a time when AI tools are increasingly being used to make unbiased, high-stakes decisions – like hiring employees, approving loans, or setting insurance rates.

If these tools bring the same flaws into the mix as humans do, we might be scaling bias rather than removing it.

“AI isn’t a neutral referee,” said Samuel Kirshner of the UNSW Business School. “If left unchecked, it might not fix decision-making problems – it could actually make them worse.”

That’s why the researchers emphasize the need for accountability from biased AI systems.

“AI should be treated like an employee who makes important decisions – it needs oversight and ethical guidelines,” explained Meena Andiappan of McMaster University. “Otherwise, we risk automating flawed thinking instead of improving it.”

Smarter AI needs unbiased monitoring

As AI evolves, so should our approach to using it. That means not only designing better models but also setting up systems to check regularly how well they’re working.

“The evolution from GPT-3.5 to 4.0 suggests the latest models are becoming more human in some areas, yet less human but more accurate in others,” commented Tracy Jenkin of Queen’s University.

“Managers must evaluate how different models perform on their decision-making use cases and regularly re-evaluate to avoid surprises. Some use cases will need significant model refinement.”

The message is clear: AI is powerful, but it’s not perfect. And if we want it to truly help us make better decisions, we’ll need to keep a close eye on how it thinks.

The full study was published in the journal Manufacturing & Service Operations Management.

—–

Like what you read? Subscribe to our newsletter for engaging articles, exclusive content, and the latest updates. 

Check us out on EarthSnap, a free app brought to you by Eric Ralls and Earth.com.

—–