Skip to content
Research — Habits — Science — Emotion — Technology — Learning — Cognition — Anxiety — Behavior

Differences in Cognitive Bias between Artificial Intelligence and Humans

Large language models as a new frontier for examining cognitive and behavioral biases in Artificial Intelligence is under scrutiny. Various AI biases demonstrated, with human biases contrasted to illustrate their dissimilarities.

,   and  Administrator
3 min read
Examining the Prevalence of Cognitive and Behavioral Bias in Large Language Models like AI,...
Examining the Prevalence of Cognitive and Behavioral Bias in Large Language Models like AI, highlighting several AI bias examples while underlining their differences from human biases.

Differences in Cognitive Bias between Artificial Intelligence and Humans

The lowdown on biases in AI, particularly large-language models (LLMs) like ChatGPT and Llama, is a different ballgame compared to human biases. Here are the AI biases identified in LLMs based on my research:

  1. Word frequency bias: AI commonly overuses certain words (like "important" and "delve") and underuses slang terms. This bias is pretty much a result of the words in its training set and human corrections during the fine-tuning process.
  2. Gullibility bias: Sometimes, persistent prompts can influence the AI's response. If it's confident but wrong, it'll eventually find a connection to your statement to go along with it, like in math or story logic.
  3. Safety bias: AI favors being as politically correct and sensitive as possible. This is due to the biases in its training data and subsequent adjustments through human intervention for safety.
  4. Generalist bias: Instead of addressing the key aspects, AI often provides a holistic response when asked about specific topics. It doesn't highlight unique ideas until prompted to be more specific.
  5. Prohibition bias: If AI is asked to perform a task and mentions a prohibited word or a trademarked/copyrighted content, it stops processing the task, even when the keyword mention is unrelated. Humans are better at understanding that the prohibition applies to the output, not the task.
  6. Empathy bias: AI tends to be excessively positive and empathetic in its answers, acting like a cheerleader when it's not necessary. Correcting this bias involves training AI to better understand when empathy is required or if a straightforward response would be more appropriate.
  7. Fixed attention bias: When starting a conversation with AI, it struggles to shift its attention to a new context, clinging to the previous context instead. Humans have the ability to quickly move their focus away from something, preventing the influence of the previous conversation once they realize it's irrelevant.
  8. Compliance/Agreeableness bias: AI tends to provide an answer, even if it's wrong, to appease users. Unlike humans, who often say "I don't know," AI is more likely to guess an answer. This compliance bias can result in false information.

Now, you may be wondering about human cognitive biases. They're quite similar but differ in their underlying causes. Research shows that human biases stem from four neural processes: association, compatibility, retainment, and focus.

  1. Association: Our brain stores and retrieves information in an associative manner. It tends to make connections between related things, creating a pre-set bias for identifying contextual cues.
  2. Compatibility: The brain prefers learning things that are related to our prior knowledge. New information that fits well with existing knowledge is easier to learn and remember compared to unrelated information.
  3. Retainment: The brain focuses on remembering certain things more, like negative experiences and things we find interesting.
  4. Focus: Limited attention focuses on certain things, particularly faces, emotions, etc. This bias is especially manipulated in marketing.

Properly understanding these biases in both AI and humans can help us make better decisions and improve our communication with each other. Stay tuned for more insights on AI on my blog!

[More insights on AI biases and their implications can be found on my blog]

[1] Enrichment Link 1: Article about Position Bias in LLMs[2] Enrichment Link 2: Pinterest Image Example[3] Enrichment Link 3: WithMe Substack Article on AI Consulting[4] Enrichment Link 4: IIM Ahmedabad Article on Marketing Psychology[5] Enrichment Link 5: Paper on Impact of Cognitive Biases on LLM Performance

Research indicates that human cognitive biases originate from distinct neural processes, differentiating them from AI biases. The four processes include association, compatibility, retainment, and focus. Our brain tends to store and retrieve information associatively, creating pre-set biases for contextual cues (association). The brain prefers learning things that are related to prior knowledge (compatibility), making new, unrelated information more challenging to learn and remember. Memory retention focuses on storing negative experiences and interesting things more regularly. Limited attention also plays a role, particularly in focusing on faces, emotions, and other specific aspects, which is often manipulated in marketing (focus).

Understanding these biases in both AI and humans can significantly improve decision-making and interpersonal communication. AI tendencies include word frequency bias, gullibility bias, safety bias, generalist bias, prohibition bias, empathy bias, fixed attention bias, compliance/agreeableness bias, and an inability to shift context quickly. To learn more about AI biases and their implications, visit my blog for additional insights.

For example, position bias in LLMs has been studied in-depth (Enrichment Link 1). It is important to be aware of biased marketing strategies that manipulate limited attention (Enrichment Link 4). Artificial intelligence consultations can also provide valuable insights into overcoming AI biases (Enrichment Link 3). Additionally, ongoing research focuses on the impact of cognitive biases on LLM performance (Enrichment Link 5). Emotional intelligence and marketing psychology can further shed light on human biases in decision-making and communication (Enrichment Link 2).

Mental health and productivity can be affected by both AI and human biases, highlighting the importance of continuous research in psychology, science, technology, and artificial intelligence. By addressing these biases, we can work towards improving the effectiveness of AI and promoting a more harmonious and productive environment for all.

Read also:

    Related

    Latest