Summary: A new study finds that Chatgpt, although excellent in logic and mathematics, exhibits many of the same cognitive biases as humans by making subjective decisions. In common trial errors evidence, AI showed excess confidence, risk aversion and even the fallacy of the classic player, although he avoided other typical human errors such as base rate negligence.
Interestingly, the newest versions of AI were more precise analytically, but also showed stronger biases based on the trial in some cases. These findings raise concerns about trusting AI for high -risk decisions, since it may not eliminate human error, but automate it.
Key facts:
AI prone to bias: Chatgpt showed cognitive biases similar to humans in almost half of the proven scenarios. Judgment versus logical: AI stands out in objective tasks but struggles with subjective decision making.
Source: Inform
Can we really trust AI to make better decisions than humans? A new study says … not always.
Researchers have discovered that Openai chatgpt, one of the most advanced and popular AI models, makes the same types of decision -making errors as humans in some situations, which shows prejudices such as the excess of confidence of the fallacy (of the players), but acting inhumans in others (for example, it does not suffer from negligence of base base or Sunk falls).
Published in the Manufacture and Manufacture of Report Magazines management, the study reveals that Chatgpt not only cuts numbers, “thinks” in a disturbingly similar way to humans, including mental shortcuts and blind spots.
These biases remain quite stable in different commercial situations, but they can change as AI evolves from one version to the next.
Ai: an intelligent assistant with human defects
The study, “a manager and an AI enter a bar: chatgpt makes biased decisions like us?” Chatgpt put through 18 different bias tests. The results?
AI falls into human decision traps: Chatgpt showed prejudices such as excessive or ambiguity aversion, and conjunction fallacy (also known as the “pretty problem”), in almost half of the tests. Predecessor, sometimes he showed stronger biases in trials based on the trial.
Why does this matter
From employment hiring to loan approvals, AI is already shaping the main decisions in business and government. But if AI imitates human biases, could you reinforce bad decisions instead of fixing them?
“As AI learns from human data, it can also think as a human: prejudices and everything,” says Yang Chen, principal author and assistant professor at the Western University.
“Our research shows when AI is used to make calls of judgment, sometimes uses the same mental shortcuts as people.”
The study found that chatgpt tends to:
Play safe: AI avoids risk, even when the most risky elections could produce better results. Overestimes in itself – Chatgpt assumes that it is more precise than it really is. SEEK Confirmation – AI favors the information that supports existing assumptions, instead of challenging them. AI’s ambiguity prefers alternatives with more information and less ambiguity.
“When a decision has a clear correct answer, the AI nails it: it is better to find the correct formula than most people,” says Anton Ovchinnikov from the University of Queen. “But when it comes to trial, AI can fall into the same cognitive traps as people.”
So can we trust AI to make great decisions?
With governments worldwide working in AI regulations, the study raises an urgent question: should we trust AI to ask important calls when it can be as biased as humans?
“The AI is not a neutral referee,” says Samuel Kirshner by UNSW Business School. “If it is not controlled, it might not solve decision -making problems, it could actually get worse.”
Researchers say that is why companies and policy formulators need to monitor the decisions of the close as they would with a human decision maker.
“The AI should be treated as an employee who makes important decisions: he needs supervision and ethical guidelines,” says Meena Andiappan of McMaster University. “Otherwise, we run the risk of automating defective thinking instead of improving it.”
What follows?
The authors of the study recommend regular audits of decisions promoted by AI and the refining of AI systems to reduce biases. With the influence of growing AI, make sure that you improve decision making, instead of just replicating human defects, it will be key.
“The evolution of GPT-3.5 to 4.0 suggests that the latest models are becoming more human in some areas, but less human but more precise in others,” says Tracy Jenkin of the University of Queen.
“Managers should evaluate how different models perform in their cases of decision -making and reassess regularly to avoid surprises. Some use cases will need a significant model refinement.”
About this AI and cognition research news
Author: Ashley Smith
Source: Inform
Contact: Ashley Smith – Informa
Image: The image is accredited to Neuroscience News
Original research: open access.
“A manager and an AI enter a bar: chatgpt makes partial decisions like us?” By Tracy Jenkin et al. Manufacturing and Service Operations Management
Abstract
A manager and an AI enter a bar: Chatgpt makes biased decisions like us?
Problem definition: Large language models (LLM) are increasingly taking advantage of business and consumer decision -making processes.
Because the LLM learn from human data and feedback, which can be biased, determine if the LLMs exhibit behavioral decision biases similar to humans (for example, base rate negligence, risk aversion, confirmation bias, etc.) is crucial before implementing LLM in contexts and decision -making flows.
To understand this, we examine 18 common human biases that are important in the management of operations (OM) using the dominant LLM, chatgpt.
Methodology/Results: We carry out experiments where GPT-3.5 and GPT-4 act as participants to test these biases using adapted bullets of literature (“standard context”) and the variants reformulated in the inventory and the OM General contexts.
In almost half of the experiments, the previously trained generative transformer (GPT) reflects human biases, diving of prototypical human responses in the remaining experiments. We also observe that GPT models have a remarkable level of consistency between the standard and specific experiments of OM, as well as in the temporal versions of the GPT-3.5 model.
Our comparative analysis between GPT-3.5 and GPT-4 reveals a double-edged edge progression of GPT decision making, in which GPT-4 advances in the accuracy of decision making for problems with well-defined mathematical solutions, while simultaneously showing the greatest behavioral biases for preferences based on problems.
Management implications: First, our results emphasize that managers will obtain the greatest benefits when implementing GPT up to work flows taking advantage of the established formulas.
Second, that GPT showed a high level of response consistency in standard, inventory and non -inventory operating contexts provides optimism that LLMs can offer reliable support even when the details of the decision and problems of problems change.
Thirdly, although the selection between models, such as GPT-3.5 and GPT-4, represents compensation in cost and performance, our results suggest that managers should invest in higher performance models, particularly to solve problems with objective solutions.
Financing: This work was supported by the Social Sciences and Humanities Research Council of Canada (Grant SSHRC 430-2019-00505). The authors also thank Smith School of Business from the University of Queen for providing funds to support Y. Chen’s postdoctoral appointment.
