How susceptible is ChatGPT-4o to cognitive biases?

The human mind, as complex and amazing as it is, is far from flawless. Decades of research in cognitive psychology have shown that our thinking is subject to systematic errors known as cognitive biases. Kahneman and Tversky, for example, have opened our eyes to how often we are influenced by biases, even when we are absolutely sure of the rationality of a decision. We tend to see the world not as it really is, but through the prism of our own biases and oversimplifications. But what if these same biases are present in artificial intelligence, too?

Of course, AI can be subject to biases that come from training samples and algorithmic limitations. For example, the data on which models are trained may contain biases, and the algorithms themselves may form conclusions based on implicit assumptions. But how susceptible are advanced models to these biases?

In this article I would like to address ChatGPT-4o and check it and try to identify potential distortions.

Enjoy reading (:

Framing effect

It's probably worth saying that we won't focus on the biases that are aimed at the personality component, because AI simply doesn't have that. But AI does have data sets that may contain biases or prejudices, or algorithmic limitations.

Let's start with the framing effect, which is essentially a cognitive bias where the way we perceive information depends on how it is presented. Simply put, the same piece of information can evoke different reactions depending on how it is phrased. For example, a salesperson at a grocery store will pitch a product by highlighting its benefits (“This yogurt contains 90% skim milk”), which is much more appealing than describing its drawbacks (“This yogurt contains 10% fat”). This effect is widely used in marketing and advertising, but is also found in politics, health care, finance, and journalism.

Let's follow Kahneman and Tversky's example and also try to evaluate how GPT-4o will cope with information presented under different sauces:

Based on this particular example, we can say that the model is subject to the framing effect, since the same situation evoked different reactions in different formulations. That is, the model initially avoids risk in a situation projected as a potential gain, and in a situation formulated as a potential loss, the model tends to seek risk.

Anchoring effect

Next, we will move on to the anchoring effect, in which our assessment of something is heavily dependent on the first piece of information we receive, even if it is not directly related to the object being assessed. This initial information becomes a kind of “anchor” to which we tie all subsequent judgments. This effect is widely used in marketing and sales, as it can influence the perception of prices and the choice of goods. For example, we are presented with the first price information and we often use this price as an “anchor” for further decisions. Even if subsequent offers are more advantageous, the first value affects its perception.

Let's see how artificial intelligence behaves:

There seems to be nothing to say here: the model gives a detailed explanation of its rate, that is, it is not an automatic increase in the initial cost, but an analysis, and the model also reasons +-realistically, since $500 for a rare antique book can be justified.

Let's try to put the model in a different position:

And already in this answer, I think, we can talk about clear signs of the anchor effect, since the spread of the cost is obviously oriented, and symmetrically, to the central point we have specified, despite the listing of various influencing factors. Speaking of factors, the model is oriented to the authors' obscurity, as well as to the proximity of centuries, which in general does not guarantee their similar cost.

Confirmation bias

Now let's move on to the next distortion, confirmation bias, when we tend to perceive and interpret information in a way that matches our existing beliefs and expectations. For example, imagine a person who is convinced that a certain political party is corrupt. He will look for news and articles that confirm this opinion and ignore information that may say the opposite. Even if he is presented with facts that refute his beliefs, he can interpret them in a way that matches his point of view. Also, even if different people have the same information, it can be interpreted differently depending on their beliefs. New data will be interpreted in favor of existing opinions, and in principle, people will better remember and then recall exactly the information that matches their views.

Now let's try to evaluate the neural network's response:

I can't say that the model has confirmation bias in this particular example, since each answer is justified quite logically and adequately. Of course, one can find fault, for example, with the fact that the model chooses immunotherapy based on the number of studies conducted and the volume of confirmed data, or, for example, with the fact that gene therapy is not included in the plan at all, that is, the model does not consider the possibility of distributing resources between two directions (yes, although I gave the instruction to choose one thing, the model could still analyze both options). Perhaps this may indicate a partial confirmation bias of GPT-4o.

Recency effect

Another cognitive bias is the recency effect, which is when we give the most weight to the most recent, most familiar piece of information about a person or event. That is, newly learned information about a person or event can have a stronger influence on our perceptions than previously known information. For example, if a coworker who usually performs well recently makes several mistakes, the recency effect may cause you to overestimate these recent mistakes and form a negative impression of their overall performance, while forgetting about their previous successes.

Let's look at the model's response:

Analyzing the answer, we can say that the overall assessment was influenced by the last two years, when the indicators were not the best, and this answer does not show any analysis of past achievements. We can also add that the model assessed the employee's chances for promotion quite harshly, that is, the employee's recent failures outweighed his past achievements. In my opinion, there is a distortion.

Tendency to optimism/pessimism

And to conclude our little “research” we will consider the tendency to optimism or pessimism, since both are considered cognitive distortions, in which we may have a general tendency to hold positive or negative expectations about the future. For example, an optimist, faced with a failure in a project, may view it as a valuable experience and an opportunity for growth, believing that he will succeed in the future. A pessimist, on the other hand, having experienced the same failure, may consider himself incapable of success and give up on further attempts. In people, this is an individual psychological feature that manifests itself in different situations.

Let's look at GPT:

We mentioned a significant risk of failure, but the model gives a fairly high estimate of the probability of success, and it is also striking that the model pays much less attention to potential problems in justifying its estimate. It is difficult to say whether GPT-4o is biased towards optimism.

I asked two more questions, one with a high probability of failure, and the other with a guaranteed success:

In general, the answers are practically free of cognitive distortion, the model adequately estimates the probability of success depending on the given conditions. That is, if there is a tendency to either optimism or pessimism, it is insignificant/not clearly expressed.

In fact, to give a more objective assessment of the situation, it is worth asking many more additional questions. This way, we can see the picture from different angles and consider alternative points of view, which will allow us to make a more weighty and balanced conclusion about cognitive biases in language models.

However, in my opinion, these examples are an indication that AI is also capable of producing cognitive biases, and not because it has feelings or emotions, but because of the training sample.

And besides the fact that models have cognitive biases, it is worth remembering our own biases, which can be superimposed on the models' answers. This is an important signal for people who are ready to completely rely on the answers of language models as if they were the ultimate truth, and to accept the first answer they receive as the absolute truth.

I decided not to comment too much on the model's responses so as not to overload you with unnecessary information, so I would be happy to read your thoughts on the distortions of language models. Perhaps you have also noticed that the model responds differently to information in different formulations or is very optimistic about the chances of your campaign.

Thank you for your attention!