[Technology Saw] – A new study reveals that ChatGPT-3 exhibits human bias.
Highlights:
- This study explains how ChatGPT-3 mirrors human biases by favoring specific information types, akin to human behavior.
- The study explores if AI systems like ChatGPT-3, trained on human data inherit biases seen in human communication.
- Researchers use the ‘transmission chain methodology,’ like a telephone game experiment to analyze ChatGPT-3 and uncover biases that might be challenging to detect using other methods.
- The study found that ChatGPT-3 replicates human biases in summarizing stories, preferring negative and stereotype-conforming information.
- Ways to make AI, such as ChatGPT to be objective.
In this article, we will look at a recent study on how AI, particularly chatbots like ChatGPT-3, can exhibit biases similar to those of humans.
Bias means having a preference for certain types of information over others, and this research aims to understand if AI systems replicate the biases found in human communication.
The reason for this study is the widespread use of large language models like ChatGPT-3 in various fields. These AI systems impact fields like academia, journalism and copywriting, among others.
The study explore whether these AI models, being trained on human-generated data, inherit human biases.
The researchers use a method called the ‘transmission chain methodology’ to study ChatGPT-3. This is like an experimental version of the telephone game. It helps reveal biases that might be hard to notice with other methods.
In their study, ChatGPT-3 is given a story and asked to summarize it.
Also, the stories used in the study were designed to identify different biases in human subjects.
The results show that ChatGPT-3 exhibits biases similar to those of humans in various scenarios. For instance:
- When a story contains both gender-stereotype-consistent and inconsistent information, ChatGPT tends to favor the stereotype-consistent details, just like humans.
- In a story about a girl’s trip to Australia with both positive and negative details, ChatGPT shows a preference for retaining the negative aspects, mirroring human bias.
- When a story involves social versus nonsocial elements, ChatGPT, like humans, favors social information.
- In a consumer report scenario, the AI is more likely to remember and pass on threat-related details over neutral or mildly negative information.
- In narratives resembling creation myths with various biases, ChatGPT tends to preferentially transmit negative, social, and biologically counterintuitive information.
The study found that ChatGPT-3 reproduces human biases in all the experiments. It tends to favor negative information—information conforming to gender stereotypes—and gives importance to threat-related information.
The researchers emphasize that when using ChatGPT-3, it’s essential to be aware that its responses are not neutral and may magnify pre-existing human tendencies.
ChatGPT
ChatGPT is a cutting-edge AI created by OpenAI. Its main job is to understand what you are saying and respond in a way that sounds like a human would. This is regardless of what topic you are talking about.
Under the hood, ChatGPT uses a type of machine learning called deep learning, with a structure called a transformer.
This setup allows ChatGPT to analyze loads of data during training and learn how to predict what words should come next in a sentence. So when you give it an input, it uses what it’s learned to generate a response that fits the context.
One cool thing about ChatGPT is its flexibility in understanding different styles of conversation. It’s been trained on tons of conversations from all sorts of places, like social media, forums and books.
This exposure helps it adapt to different writing styles and languages, making it feel more like you’re talking to a real person.
By customizing the model for specific tasks, developers can make it more accurate and useful for those particular jobs.
But, like all AI, ChatGPT isn’t perfect. It can pick up biases from its training data, just like people can. These biases might show up as stereotypes, cultural preferences or language quirks, which could affect the responses they give.
That’s why it’s important to be aware of these potential biases and think critically about the information ChatGPT provides.
Ways to make AI such as ChatGPT to be Objective
Achieving fairness and impartiality in AI systems, such as ChatGPT, involves employing various strategies to reduce biases and promote neutrality. This is a breakdown of these methods in simpler terms:
Diverse Data Selection: Think of AI as a smart learner. To make sure it learns fairly, expose it to a broad range of information during training.
This variety includes different perspectives, like looking at a problem from various angles to prevent the AI from leaning towards one side.
Bias Identification and Removal: Imagine the AI as a student taking a test. It is important to regularly check its answers to see if there are any unfair preferences or mistakes.
If you find any, help the AI learn better by providing corrected information, like a teacher guiding a student to understand a concept correctly.
Ethical Guidelines: Creating rules for AI is like setting ground rules for a game. You need to establish clear guidelines to make sure the AI behaves ethically.
This means avoiding topics that could lead to unfair or discriminatory results, similar to how we avoid unfair moves in a game.
Algorithm Transparency: AI is like a problem-solving friend. So, make sure it explains its solutions so that we can understand how it makes decisions.
This transparency helps to identify and fix any biases, just like discussing and improving a friend’s reasoning.
Continuous Evaluation and Improvement: Imagine the AI as a student who keeps learning. So, regularly check its performance and tweak its learning methods to make it smarter over time.
This ongoing learning process helps the AI adapt to new information and become fairer.
Diverse Development Teams: Creating AI is like working on a group project. Consequently, encourage a mix of people with different backgrounds and perspectives to contribute.
Balanced Training: Training AI is similar to learning from different textbooks. As a result, expose it to a balanced mix of information from various sources and viewpoints.
This prevents the AI from becoming biased toward one type of information, just like studying different books gives a more comprehensive understanding.
User Feedback Integration: Imagine AI as a social media profile that gets comments. So, create ways for users to tell us if they notice any unfairness.
This feedback helps us identify and fix issues, like friends helping each other improve.
Regular Audits and Assessments: Think of AI like a company that gets checked by independent auditors. You need to regularly have independent parties review the AI to ensure it’s fair and follows ethical standards.
Clear Objective Definition: Defining AI’s goals is like setting a clear mission for a team. Also, make sure everyone knows what the AI is supposed to achieve and ensure these goals align with fairness and unbiased decision-making.
By using these strategies, you are like a coach guiding AI to play a fair game. This helps to reduce biases and making it better at helping everyone.
However, this study raises awareness about the biases in AI and emphasizes the need for caution when relying on these systems, recognizing that they may inherit and propagate certain biases present in human communication.
The study accentuates the significance of acknowledging AI biases, cautioning users that responses from ChatGPT-3 might amplify existing human tendencies.
Nonetheless, it acknowledges limitations and highlights the need for future research on diverse AI systems to understand how they present biases based on different information inputs.