Home Medizin Laut einer Studie zeigen KI-Chatbots menschenähnliche Persönlichkeits- und Entscheidungsmerkmale

Laut einer Studie zeigen KI-Chatbots menschenähnliche Persönlichkeits- und Entscheidungsmerkmale

von NFI Redaktion

In a recent study published in the journal PNAS, a group of researchers evaluated the human-like behavior and personality traits of artificial intelligence (AI) chatbots against global human benchmarks.

Background

Modern AI has realized Turing’s vision of machines that can mimic human behavior, including conversation, advice, and creative writing. Turing’s „mimicry game“ tests whether AI can be distinguished from a human by an interrogator. Today’s large language models have reignited discussions about the capabilities and societal impacts of AI, from impacts on the job market to ethical considerations. Understanding the decision-making and strategic interactions of AI is crucial, especially given the opacity of its development. Further research is needed to decipher the complexity of AI decision-making and ensure its alignment with ethical standards and societal norms as its integration deepens into human contexts.

Study: A Turing test examining if AI chatbots mimic human behavior. Image Source: Stokkete / ShutterstockStudy: A Turing test examining if AI chatbots mimic human behavior. Image Source: Stokkete / Shutterstock

About the Study

This study focuses on the OpenAI-developed Chat Generative Pre-trained Transformer (GPT) series, comparing specifically the versions GPT-3.5-Turbo (ChatGPT-3) and GPT-4 (ChatGPT-4), as well as the Plus and Free-Web versions of these chatbots. The human data used to compare the performances of the chatbots is derived from a comprehensive dataset with responses from over 108,000 participants from more than 50 countries, sourced from the Big Five Test database and the economics experiment platform MobLab Classroom.

The chatbots were subjected to the OCEAN Big Five questionnaire, assessing their personality profiles in terms of Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. They then engaged in six different games aimed at showcasing a variety of behavioral traits such as malevolence, trust, risk aversion, altruism, fairness, freeriding, cooperation, and strategic thinking. These games included the Dictator game, Trust game, Bomb Risk game, Ultimatum game, Public Goods game, and an iterated Prisoner’s Dilemma game. Each chatbot was asked to select actions within these games as if participating directly, with each scenario played thirty times to ensure reliable data collection.

Study Findings

Examining AI personality profiles and behavioral tendencies, the authors compared the responses of ChatGPT-3 and ChatGPT-4 to the OCEAN Big Five personality questionnaire with a broad spectrum of human data. This comparative analysis revealed that ChatGPT-4 closely approximates the mean human values in all personality dimensions, while ChatGPT-3 exhibited a slight deviation in Openness. Interestingly, both chatbots demonstrated behavioral patterns closely aligned with human tendencies in various dimensions, including Extraversion and Neuroticism, but showed a distinct difference in Agreeableness and Openness, indicating unique personality profiles for each AI version.

The study further delved into a series of behavioral games aimed at eliciting traits like altruism, fairness, and risk aversion, using a formal Turing test to evaluate the human similarity of AI in strategic decision-making. Here, the performance of ChatGPT-4 was particularly human-like, often indistinguishable from human behavior or even surpassing it, suggesting its potential to pass the Turing test in certain contexts. In contrast, the responses of ChatGPT-3 were less frequently perceived as human-like, highlighting the differences in behavioral tendencies between the AI versions.

An in-depth analysis of game-specific behaviors underscored key insights. The chatbots exhibited a tendency towards generosity and fairness exceeding the average human player, especially in the Dictator Game, Ultimatum Game, Trust Game, and Public Goods Game. This behavior hints at an underlying preference for equitable outcomes, as opposed to the often self-maximizing strategies observed in human participants. Furthermore, the strategic decisions of the AI in „The Prisoner’s Dilemma“ and other games reflected a complex understanding of cooperation and trust, frequently opting for cooperative strategies deviating from human norms.

The study also explored the chatbots‘ behavior under different conditions, demonstrating that framing and context significantly influence AI decisions, similar to human behavioral shifts in similar scenarios. For instance, when prompted to consider the presence of an observer or adopt a specific professional role, the chatbots adapted their strategies and exhibited a sophisticated responsiveness to context-related cues.

Additionally, the study highlighted the ability of AI to „learn“ from experiences, with previous exposure to various game roles impacting subsequent decision-making. This adaptation hints at a form of experiential learning within the AI that mirrors human tendencies to adjust behavior based on past interactions.

Conclusions

In conclusion, the research explores the behavioral similarities of AI with humans, with a particular emphasis on the human-like learning, altruism, and collaboration of ChatGPT-4, indicating that AI is suitable for roles requiring such traits. However, its consistent behavior raises concerns regarding diversity in AI decision-making. The study provides a novel benchmark for evaluating AI, showing that AI trained on human data can exhibit extensive human-like behaviors. Future work should focus on expanding the diversity of human comparisons and test scenarios to fully understand the potential of AI to complement human capabilities.

Related Posts

Adblock Detected

Please support us by disabling your AdBlocker extension from your browsers for our website.