Exploring 3 Common Cognitive Biases with ChatGPT 🧠💭

🌈 Abstract

The article discusses how humans are deeply irrational, despite the common belief that we are logical creatures. It explores three common cognitive biases - hindsight bias, anchoring effect, and halo effect - and examines how they manifest in the responses of the AI chatbot ChatGPT. The article highlights the importance of understanding cognitive biases, both in human and AI behavior, as we increasingly integrate AI into our daily lives.

🙋 Q&A

[01] Hindsight Bias

1. What is hindsight bias?

• Hindsight bias refers to the tendency to perceive past events as having been more predictable than they actually were at the time. This bias leads people to believe they "knew it all along", even when there was no way to predict the future outcome.

2. How does hindsight bias influence the training data of LLMs like ChatGPT?

• There is likely a long list of texts and documents in the training corpus of GPT models that explain how certain events were predictable and how governments/organizations were incompetent for not seeing them coming. This results in the reproduction of the hindsight bias in the model's responses.

3. What is the key insight about predictability and causation?

• Correlation does not necessarily mean causation, and even causation does not guarantee predictability. There is an infinite number of possible outcomes for any situation, and the outcome we experience is mostly a matter of chance. Events may only make sense after they have occurred, and we can retroactively rationalize them.

[02] Anchoring Effect

1. What is the anchoring effect?

• The anchoring effect refers to the tendency to form judgments or decisions based on an initial reference point or "anchor", which can be completely irrelevant. We tend to heavily rely on the first piece of information we receive about a topic and attribute less value to additional information.

2. How does the anchoring effect manifest in ChatGPT's responses?

• The anchoring effect seems to work in ChatGPT without any time dimension being involved, solely based on the order of the words. ChatGPT appears to place more emphasis on the first words, positive or negative, and attributes much less importance to the additional information.

3. What is the key insight about first impressions?

• The first information we receive and the respective impression we form matter the most, while the following information has less and less impact. This is the essence of the saying "first impressions matter the most".

[03] Halo Effect

1. What is the halo effect?

• The halo effect refers to the tendency to attribute positive traits to someone who exhibits a positive trait in a specific area. The most common form is attributing positive traits to good-looking people, but it can apply to any characteristic.

2. How does the halo effect manifest in ChatGPT's responses?

• ChatGPT reproduces the halo effect by assuming that someone may have positive traits in unrelated areas based on a single positive trait. For example, it assumed that someone who writes clean code and is a good communicator must also be a good dancer, even though these traits are unrelated.

3. What is the key insight about the perception of AI responses?

• Most people perceive responses from AI conversational agents as objective and unbiased, often believing they originate from some kind of expert. This makes it easy for our own biases to be triggered on top of the cognitive biases that may already exist in the AI-generated responses.