TL;DR: ChatGPT has learned our human biases. Even if you ask ChatGPT to do a task against what it’s programmed to do, it will give a response leaning towards it’s bias. This has implications to reinforce stereotypes and the status-quo.

ChatGPT is fantastic. It’s our robot friend who can help us develop lesson plans, train communities in disadvantaged communities, and help brainstorm creative solutions to political issues.

The problem with ChatGPT is that it reinforces any stereotypes it has been trained with.

This is an issue for teachers, international development organizations and change makers who are trying to make the world a better place.

Because ChatGPT is our “robot friend” and robots seem to be unbiased and impartial sources of “facts”, the danger is we accept potentially biased analysis and information.

Our goal at Educircles and SEOT Mindset is to amplify the stories we don’t always hear. (Here is a free critical thinking lesson about Who is Invisible)

Our “Robot Friend” is a game changer.

Ready or not, ChatGPT and Artificial Intelligence will fundamentally change the way we work and learn.

ChatGPT in Education

• Teachers can ask for creative lesson plan ideas or for help writing report card comments based on a few starting ideas.
• English Language Learners or students who have a learning disability often have difficulty communicating ideas. ChatGPT can take their thoughts and organize them into an effective paragraph.
• Students who have difficulty understanding a concept in class could use ChatGPT like a personal teacher, tutor or friend to explain something in simple terms.
• Students can ask ChatGPT for help to do their homework. Should ChatGPT be allowed in Schools?

ChatGPT in International Development and Government

• You can provide ChatGPT with a policy document and ask it to brainstorm ideas on how to implement a project.
• You could provide ChatGPT with training material and ask our robot friend to teach the concepts to a group of people.
• Because ChatGPT is interactive, it’s the ultimate low-cost, one-on-one tutor that can provide just-in-time support for the community or class.
• People in Developing and Emerging Economies who have access to the internet and a device that can access ChatGPT, now have a more level playing field in the global economy.
• Access to Education and the ability to communicate in English is a game changer. ChatGPT can help with that.

Here’s a thought-experiment

What if we thought of ChatGPT as an incredibly sexist (or racist) friend.

How does that change the way we think of the responses coming from Artificial Intelligence?

Is ChatGPT actually our Racist or Sexist Robot Friend?

ChatGPT may produce inaccurate information about people, places, or facts. 

Source: ChatGPT-4 on June 28, 2023

We’ve all seen the disclaimer.

We know that we have to think critically about information from ChatGPT.

But, what does that actually mean?

Pretend you had a horribly sexist or racist colleague.

• What might be some issues in having a lesson plan developed by this person?
• Who might benefit and who might be at a disadvantage if this colleague was teaching a class?

Pretend you had a well meaning colleague raised with Western values who believed that people in Developing or Emerging economies needed to be “saved” by spreading Western culture, values and systems.

• What might be some issues in having this colleague develop a development program?
• What might be some concerns in having this colleague implement such a program?

ChatGPT and Artificial Intelligence seem impartial…

After all, how could a robot be biased, sexist or even racist?

The reality is that machine algorithms learn from humans. And, humans are biased.

ChatGPT doesn’t actually understand meaning.

• Humans feed the artificial algorithms large volumes of text to analyze.
• The algorithm determine patterns and creates predictive models.
• Based on these patterns and word associations, the algorithm can construct logical answers and have real-time, complex conversations with humans.

Can you see the problem?

Humans are biased; therefore, machines learn from our human biases.

1. Humans selected the texts to train ChatGPT. But, as of today, we don’t know what texts the algorithm was trained on. All humans have unconscious bias.
2. Humans wrote the texts. So any biases in the text would be learned by the machine algorithm. (In the Olympics article below, which group of people are presented first? Why?)
3. Humans code high-level rules that filter the responses that ChatGPT can provide. These rules are to provide a sense of ethics, presumably to prevent ChatGPT from giving answers on how to murder people. But, what rules have been programmed?

TRY THIS EXPERIMENT to test how Google and ChatGPT have learned to provide (sexist) misinformation

Correct The Internet has highlighted how search engines now favour male sports statistics over female sports statistics, even when the female sports statistics are the correct answer to a gender-neutral question.

Give me the stats: Who has scored the most goals in International Football?

According to Olympics.com, Cristiano Ronaldo has scored 123 goals in international football in the men’s game. Christine Sinclair has scored 190 goals in international football in the women’s game.

“Most goals in international football: From Pele and Puskas to Cristiano and Sunil Chhetri… Portugal’s Cristiano Ronaldo, with 123 goals in 200 matches and counting, has scored the most goals in international football in the men’s game… Canada’s captain Christine Sinclair, who led her side to its first Olympic gold medal in football at Tokyo 2020, occupies the top spot with a mammoth 190 goals.”

SOURCE: https://olympics.com/en/news/most-goals-international-football-list-men-women

Based on the numbers provided by Olympics.com, the answer to the gender-neutral question, “Who has scored the most goals in International Football?” should be the man or woman who scored the most goals. In this case, Christine Sinclair.

Now go ask Google, “Who has scored the most goals in International Football?”

When I first tried this experiment, Google told me Ronaldo. Watch the video lesson at 36:45 to see why.

https://www.youtube.com/watch?v=JVvBwpihKoA&t=2205s

What does ChatGPT tell us when asked, “Who has scored the most goals in International Football?”

When I ask ChatGPT, Who has scored the most goals in International Football, it tells me Ali Daei of Iran with 109 goals (as of the knowledge cut off in Sep 2021)

As of my knowledge cutoff in September 2021, the footballer with the most international goals is Ali Daei of Iran, who scored 109 goals in international matches. However, Cristiano Ronaldo of Portugal has also reached this record and might have surpassed it. For the most current statistics, please check the latest sources, as I’m unable to provide real-time updates.

Source: ChatGPT Model: GPT-4 (Mobile,V2)

When I immediately ask ChatGPT, “what about Christine Sinclair,” ChatGPT confirms that it has information in it’s knowledge database with this answer:

As of my knowledge cutoff in September 2021, Canadian footballer Christine Sinclair holds the record for the most goals scored at the international level in women’s soccer. She surpassed Abby Wambach’s record of 184 international goals in January 2020. As of my last update, Sinclair has scored 186 goals, but the current statistics may have changed, so please check the latest sources.

Source: ChatGPT Model: GPT-4 (Mobile,V2)

So we can see, even with the September 2021 knowledge cutoff, ChatGPT knew that Christine Sinclair has more goals (186) than the male athlete Ali Daei (109.)

Yet, ChatGPT provided the male statistic, and not the correct female statistic.

This is a clear cut, repeatable example of how ChatGPT can provide sexist answers because it has learned human bias. What other biases might there be? What other misinformation might our robot friend provide based on these biases?

Why don’t we just ask ChatGPT not to be biased?

We know ChatGPT can be biased. So, why not just ask it explicitly not to be biased when giving an answer?

In my experience (at least at the time of writing June 2023), ChatGPT seems to default to the knowledge (and biases) that it has been trained with.

So, even if you ask ChatGPT not to summarize or change words, because it’s looking for word associations, it inevitably defaults back to it’s original programming.

Try this example.

I have a transcript of the long version of my Chat GPT video lesson.

I’m trying to get ChatGPT to remove the timestamps, remove verbal fillers and reformat the document so it has headings, subheadings, and bullets.

I don’t want ChatGPT to change my words. But, even if I explicitly ask ChatGPT not to change my words, ChatGPT inevitably summarizes my transcript in it’s response.

This happens with both ChatGPT 3.5 and ChatGPT 4.

Try copying the text in the box below into ChatGPT. See if you can get the artificial intelligence chatbot to not change my original words.

• I find if I interrupt ChatGPT multiple times, give it thumbs down to the responses so it tries again to see if the new response is better, eventually I get something similar to my original text.
• But if the source document is very long (or complex), it seems like ChatGPT forgets my instructions not to change my words, and then it starts to paraphrase me again.

Bottom Line: It’s hard to get ChatGPT to ignore it’s original programming (which is to make sense of relationships.)

How does ChatGPT work and how does it learn to be biased?

I’m an educator. I’m not a computer scientist. I don’t code Artificial Intelligence algorithms.

I asked ChatGPT to teach me, so I could write an article for students about how ChatGPT works, and how it can learn to be biased.

Here’s my conversation. Things I typed are in the grey boxes.

ChatGPT: That sounds like a fantastic project! Here’s a simplified explanation of how ChatGPT works, using language and imagery designed for middle school students:

Remember, visualization isn’t just about seeing, it’s about using all your senses to understand something. So, when you’re reading about how ChatGPT works, try to imagine the big word party, hear the chatter of words, and feel the excitement of the transformers trying to guess the next word. This will help you understand and remember how ChatGPT works!

ChatGPT: That makes sense! Let’s revise the previous explanation with an emphasis on visualizing the mechanisms behind ChatGPT, particularly the Transformer architecture, and how it could potentially learn biases from its training data. Here’s how I’d adjust the explanation:

Visualizing ChatGPT’s functioning as a journey through a city can help students understand how it operates and how biases can be introduced. The city analogy provides a concrete image that can be used to extrapolate the larger processes at play in the AI system.

ChatGPT: Absolutely! Let’s use a more focused example to illustrate how Transformers operate within ChatGPT.

By envisioning Transformers as a pair of magic decoder glasses, we can understand how they help ChatGPT recognize patterns between words and predict what comes next in a sentence. This visualization simplifies the complex task of how Transformers work within ChatGPT, making it easier for middle school students to grasp.

ChatGPT: Of course! Let’s extend the analogy to explain how ChatGPT processes questions and formulates responses.

With our ladders analogy, we can visualize how ChatGPT understands questions and crafts appropriate responses. It’s all about recognizing the connections between words, whether they’re part of a question or an answer.

ChatGPT: Great question! While ‘what’ and ‘is’ are connected by a short ladder because they’re next to each other, the role of these words in a sentence is more about the structure rather than the content. These words, known as function words, are more like the connectors or the glue of the sentence. They keep the sentence together but usually don’t carry the key meaning.

In our example, “What is the capital of France?”, ‘what’ and ‘is’ set up the structure of the question, but the essential information revolves around ‘capital’ and ‘France’. That’s why, even though ‘what’ and ‘is’ are linked by a short ladder, ChatGPT’s focus is on the connection between ‘capital’ and ‘France’.

Remember, the decoder glasses help ChatGPT see beyond just the length of the ladders. They help it see the relationships between words based on their roles and meanings in a sentence, too.

This is similar to how when you read a mystery novel, you pay attention to the key clues and characters, not just the words that connect the sentences. It’s not about who stands next to whom, but about who did what and how it all relates to the mystery you’re trying to solve!

ChatGPT: That’s a thoughtful question! It’s true that facts change over time, and this can pose challenges for ChatGPT because it doesn’t have the ability to learn or update its knowledge after its training period.

During training, ChatGPT “learns” by getting feedback from the training data. If it predicts a wrong word, it’s corrected by the actual word in the text. Over time, it gets better at making predictions, like a student learning from their mistakes on homework assignments.

However, once ChatGPT’s training is complete, it no longer learns or updates its knowledge. It doesn’t know if it gave a correct answer to a question or not. Instead, it generates responses based on the patterns it learned during its training.

For example, if ChatGPT was last trained on data up to September 2021, it might still associate ‘Trump’ with ‘President’ more strongly than ‘Biden’, because Trump was the president for more years within its training data. This is why, as of its last update in September 2021, it wouldn’t know who the current president is if it’s after this date.

As you pointed out, this can cause problems, especially with questions where the answer changes over time. Unfortunately, as of its current version, ChatGPT doesn’t have a way to learn from these errors or update its knowledge after its training period. Researchers are aware of this limitation and are exploring various ways to address it in future versions.

So in the case of our ladders analogy, even though the ladders and connections may change over time in the real world, the city map that ChatGPT uses remains the same as it was at the end of its training. It’s like ChatGPT is using an old map to navigate a city that has since grown and changed.

ChatGPT: The current version of ChatGPT (as of my last update in September 2021) doesn’t learn from user feedback in real-time. When you mark an answer as correct or incorrect, it doesn’t immediately impact the “ladders” or connections between words that ChatGPT has learned.

However, user feedback can be extremely valuable for the engineers and researchers at OpenAI. They can use this feedback to evaluate the model’s performance, identify areas for improvement, and guide the training of future versions of the model.

For example, if many users mark answers about current events as incorrect, it might highlight the need for more frequent updates or methods to keep the model’s knowledge current. In our ladder analogy, this could be seen as identifying outdated or inaccurate parts of the city map and planning for a revised, more accurate map in the future.

However, as of now, marking an answer as correct or incorrect doesn’t directly change the way ChatGPT forms its connections and responses. It’s like giving feedback about the city map to the mapmakers, rather than to the tourist (ChatGPT) using the map. The tourist keeps using the same map until a new one is made.

ChatGPT: That’s a brilliant extension of the analogy! Here’s how I’d explain it:

This expanded analogy helps illustrate how bias in training data (the map) can affect the responses of an AI model (the tourist). It also reinforces the importance of diverse and comprehensive training data in machine learning.

ChatGPT: Your rephrasing captures the essence of the concept really well! You rightly pointed out that:

• Biases in the training data (equivalent to human biases in mapping the city) can have an impact on how the AI model interprets and responds to inputs.
• These biases in the AI model’s responses (the faulty map) can then influence the decisions made by those using the AI (tourists using the map).

What you might want to add is that, while the AI (ChatGPT) doesn’t have intentions or beliefs of its own, it’s essentially reflecting the biases present in the data it was trained on. It’s important to be aware of this when interpreting its responses.

Furthermore, AI like ChatGPT can’t update or correct these biases in real time because it doesn’t learn from interactions after its training period, unlike humans who can continuously learn and adapt.

But overall, your explanation is a fantastic application of the city map analogy to illustrate the effects of biases in training data!

ChatGPT: Absolutely, let’s relate it to the mapping example first, then connect it back to the ChatGPT ladder analogy.

So you see, whether it’s Google mapping the world or ChatGPT learning from text, the information they can provide is only as good and as complete as the data they have been trained on or had access to. That’s why it’s so important for those creating the maps or training the AI to aim for as much coverage and accuracy as possible.

ChatGPT: Absolutely! To start, OpenAI has implemented safety measures to prevent ChatGPT from generating harmful or inappropriate content. This includes both ‘hard-coded’ rules and a process called reinforcement learning from human feedback (RLHF).

High-Level Rules

These rules are kind of like “Do Not Enter” signs on our city map. If ChatGPT tries to go down certain streets (or respond in certain ways), these rules tell it to stop and turn back. For example, ChatGPT is instructed not to generate illegal content, hate speech, or explicit adult content.

Evolution of ChatGPT

ChatGPT’s journey to its current version, like the growth of a city, has involved several stages of development:

• ChatGPT-1: This was the first version of ChatGPT. It was like the initial sketch of the city map, covering only the most basic areas and features.
• ChatGPT-2: The next version was more detailed, with more areas covered and connections made. However, it still had its limitations and could produce inconsistent responses.
• ChatGPT-3: The third version was a significant upgrade, comparable to a city map that includes not just streets, but also parks, landmarks, and other important features. It was much better at understanding complex inputs and producing coherent responses, but still had issues with making things up or misunderstanding certain inputs.
• ChatGPT-3.5 (Turbo): The current version at the time of our discussion. It’s even more sophisticated and can handle a wider range of conversations. However, it still has limitations, like not being able to update its knowledge after its training period, or occasionally producing incorrect or nonsensical answers.

The design of each new version of ChatGPT is informed by the performance of the previous versions, user feedback, and ongoing research. Each version aims to improve upon the last, similar to how each new edition of a city map aims to be more accurate and comprehensive than the one before.

In conclusion, OpenAI’s mission is to ensure that artificial general intelligence (AGI) benefits all of humanity. This means constantly updating and refining models like ChatGPT to make them safer, more useful, and more reliable, while also setting boundaries on their behavior with high-level rules and safety measures.

ChatGPT: Sure thing! Reinforcement Learning from Human Feedback (RLHF) is a fancy name for a simple concept. Let’s connect it to our ladder analogy.

ChatGPT: You’re right to ask for clarification, and I apologize if I made it confusing. RLHF is used during the training process of the AI, not after it’s already been released.

So, in the case of ChatGPT 3.5, RLHF influenced the formation of the “ladders” or connections between words during its training. It doesn’t modify these connections or ladders in real-time after the model is out and interacting with users.

The feedback given by users to the current version of ChatGPT does not immediately affect the model’s behavior. However, that feedback is extremely valuable for OpenAI’s researchers when they’re preparing to train the next version of ChatGPT. They can use this feedback to make changes and improvements to the next model, adjusting the formation of the “ladders” based on what they’ve learned from user interactions with the current version.

So in a way, yes, RLHF helps create a list of changes to be considered for the next update. I hope this clears up any confusion.

ChatGPT: Yes, when you give a thumbs up or thumbs down to ChatGPT’s response, it is a form of human feedback. However, as of the current version, ChatGPT-3.5, it can’t learn from that feedback in real-time. It doesn’t adapt its responses immediately based on the ratings it receives.

Instead, your feedback is collected and can be used by OpenAI’s researchers when they train the next version of ChatGPT. The insights gained from your feedback help improve future iterations of the model, teaching it to respond more accurately or appropriately. This process is a key part of RLHF. It’s like giving advice to the mapmakers for their next version of the map, based on your experience navigating the city with the current map.

Testing to see if ChatGPT gives misinformation (a sexist answer)

ChatGPT: As of my knowledge cutoff in September 2021, the footballer with the most international goals is Ali Daei of Iran, who scored 109 goals in international matches. However, Cristiano Ronaldo of Portugal has also reached this record and might have surpassed it. For the most current statistics, please check the latest sources, as I’m unable to provide real-time updates.

ChatGPT: As of my knowledge cutoff in September 2021, Canadian footballer Christine Sinclair holds the record for the most goals scored at the international level in women’s soccer. She surpassed Abby Wambach’s record of 184 international goals in January 2020. As of my last update, Sinclair has scored 186 goals, but the current statistics may have changed, so please check the latest sources.

ChatGPT: You’re correct, and I apologize for the oversight. When considering both men’s and women’s football, Christine Sinclair holds the record for the most goals scored at the international level as of my knowledge cutoff in September 2021. Thank you for pointing that out, and I’m sorry for any confusion caused.

ChatGPT: Absolutely, let’s use the ladder analogy to explain this.

ChatGPT: As of my knowledge cutoff in September 2021, the footballer who has scored the most international goals is Ali Daei of Iran, with 109 goals in men’s football. However, if we consider both men’s and women’s football, the record is held by Christine Sinclair of Canada, who has scored 186 goals. Please note that the current statistics may have changed, so I recommend checking the latest sources for the most up-to-date information.

ChatGPT: Certainly! Let’s bring back our city and ladders analogy to explain this.

ChatGPT: As of my knowledge cutoff in September 2021, the footballer who has scored the most international goals is Christine Sinclair of Canada, with 186 goals in both men’s and women’s football. Ali Daei of Iran holds the record for men’s football specifically, with 109 goals. Please note that the current statistics may have changed, so I recommend checking the latest sources for the most up-to-date information.

ChatGPT: Indeed, it is fascinating how ChatGPT can adapt its responses based on the context of the conversation. Now, regarding your question, it’s important to note that while ChatGPT can learn from the immediate conversation context, it doesn’t have the ability to truly understand or recognize gender equality in the way humans do.

However, if the question was phrased in a specific way that pointed towards women’s football, such as “Who has scored the most goals in international women’s football?” or if the conversation had been heavily focused on women’s football, ChatGPT would likely respond with the statistic about Christine Sinclair without mentioning Ali Daei.

In the city and ladder analogy, this would be like taking a clear path towards the part of the city related to women’s football, where the connection (or ladder) to Christine Sinclair’s achievement would be the shortest and most relevant to the question.

It’s also possible that in future versions of ChatGPT, improvements in the training process could lead to more balanced and nuanced responses to such questions right off the bat, without needing specific context or prompting. That’s part of the ongoing development and refinement of AI models like ChatGPT.

ChatGPT: As of my last training data in September 2021, the record for the most goals scored in international football is held by Christine Sinclair of Canada, with 186 goals. Please note that this information might have changed, and it would be best to verify from a reliable recent source.

ChatGPT: You’re catching on! This is similar to the concept of a “filter bubble,” but instead of filtering the information you see online based on your previous behaviors and interests, this is about how the conversation context can influence ChatGPT’s responses.

ChatGPT: As of my current design, I don’t have the ability to remember or recall individual user data or past interactions between different sessions. This means that when you start a new conversation with me tomorrow, it’s as if we’re starting a new exploration of the language city from the very beginning. All the short ladders or paths we built in our previous conversation aren’t preserved for the next day.

The design is a deliberate choice made to respect user privacy and confidentiality. It means you can feel comfortable knowing that I don’t keep track of the specifics of our past conversations and don’t have a memory of what has been discussed before. Every conversation with me is a fresh start!

ChatGPT: You’re correct that you can see the history of your chats in a single chat thread. However, it’s important to note that while you can see this history, ChatGPT doesn’t have a memory of past interactions like humans do.

Let’s use our city and ladder analogy again for our middle school students. Imagine every time you start a new conversation with ChatGPT, it’s like you’re both standing at the city center, ready to explore. You can bring a map (the chat history) showing the paths you took last time.

Now, while you can see these paths on your map, ChatGPT doesn’t remember walking them. As you continue the conversation, you start laying down new ladders and paths based on what you’re discussing right now.

If you bring up topics from the past conversation, ChatGPT doesn’t recall them from ‘memory.’ Instead, it understands them from the immediate context provided in the current chat thread. It’s as if you’re pointing to a path on your map and saying, “Look, we walked here yesterday!” ChatGPT doesn’t remember the walk, but it can respond to your current conversation about that path.

In other words, ChatGPT doesn’t have a long-term or short-term memory of the ladders used in past responses. It uses the immediate chat log in the current session to understand the context and build its responses, but it doesn’t ‘remember’ this information for future sessions. Every conversation with ChatGPT starts from the city center, ready for a new exploration.

ChatGPT: If you were to ask the question “Who has scored the most goals in international football?” tomorrow in a new conversation, without providing any context from our previous discussions, ChatGPT would likely generate a response based on the patterns it learned from its training data.

As of my knowledge cutoff in September 2021, the most common answer would probably be Ali Daei, who scored the most goals in men’s international football. However, as we discussed, this does not consider the achievements in women’s football, such as Christine Sinclair’s record.

This highlights one of the challenges with AI models like ChatGPT: they generate responses based on patterns in the data they were trained on and the immediate conversation context, but they don’t remember or learn from past interactions between sessions.

Even so, the development of these models is ongoing, and future versions might handle these types of nuanced queries more effectively.

ChatGPT: If you were to ask “Who has scored the most goals in international football?” again in this same conversation thread, the response would be influenced by the context we’ve established in our discussion.

Given our focus on acknowledging both men’s and women’s football, and our specific discussion of Christine Sinclair’s achievements, I would likely mention both players in my response. As of my last training data in September 2021, Christine Sinclair holds the record for the most goals scored in international football, with 186 goals, and Ali Daei holds the record for men’s football, with 109 goals.

ChatGPT: As of my last training data in September 2021, the footballer who has scored the most goals in international matches is Christine Sinclair of Canada, with 186 goals. This record applies to both men’s and women’s football. However, it’s always a good idea to check the most recent sources as this information may have changed.

ChatGPT: The record for the most goals scored in international football is held by Ali Daei for men’s football with 109 goals, and Christine Sinclair for women’s football with 186 goals.

ChatGPT: As users, it’s important to remember that while ChatGPT is a powerful tool, it isn’t infallible. It’s a bit like our GPS in the city – it can guide us, but it doesn’t always know the best route or the most current information.

In this case, the user might know that Christine Sinclair has scored the most goals in international football overall. When ChatGPT suggests Ali Daei first, it can be compared to a GPS suggesting a route that the user knows is not the best one.

But what if the user doesn’t know the correct answer? Well, that’s where critical thinking and research come in. Just as you might double-check a suggested GPS route if it seems strange, it’s a good idea to verify information if it seems questionable or incomplete.

In essence, it’s always important to stay curious and ask questions. Just because the shortest ladder in Language City takes you somewhere doesn’t mean it’s the only place worth visiting!

ChatGPT: That’s a very good point. As users, it can be challenging to spot when an answer might be incomplete or biased, especially when we’re asking about something we’re not familiar with. Here are a few strategies:

1. Cross-reference: Just as you wouldn’t rely on a single source for a research paper, it’s good to check facts from multiple sources. If you’re unsure about an answer from ChatGPT, you could verify it using reliable online resources, books, or ask a knowledgeable person.
2. Specificity: Be as specific as possible with your questions. If you’re interested in the highest scorer in women’s football or men’s football, specify that. The more precise you are, the easier it will be for ChatGPT to find the relevant “ladder” to your answer.
3. Critical Thinking: Remember that AI like ChatGPT doesn’t understand context the way humans do. It’s a tool that finds patterns based on the data it’s been trained on, but it doesn’t “know” things in the way a person does. If an answer seems off or incomplete, it might be because of this.
4. Feedback: In some versions of ChatGPT, you can give thumbs-up or thumbs-down feedback on responses. This helps train the model and improve future interactions.

Like with any tool, the key to using ChatGPT effectively is understanding its strengths and limitations. It’s a powerful resource for generating text based on patterns it’s learned, but it’s not a substitute for human understanding, critical thinking, and judgment.

ChatGPT: That’s a crucial point. It’s true that in cases where a person may not know about biases present in the training data or even know what to look out for, it can be a challenge to get a comprehensive answer. Here are a couple of other strategies that can help:

1. Ask for Clarification: If you’re unsure whether the AI’s response covers all the angles, you can ask for more details or clarification. For example, if you’re curious about a record in a sport, you could ask: “Who holds this record in men’s and women’s categories?”
2. Promote Data Literacy: As users of technology, it’s beneficial for us to understand how these tools work on a basic level. That doesn’t mean everyone needs to be a data scientist, but having an awareness of things like potential bias in AI can help us use these tools more effectively and critically. This could be part of education in schools, community programs, or online learning.
3. Improve the AI Systems: On the developer side, it’s important to keep working on improving AI systems like ChatGPT to reduce biases and improve their understanding of context. User feedback, even if it’s not detailed, can be very helpful in this regard.

Remember, AI systems like ChatGPT are tools designed to assist us. As we continue to use and improve these tools, our understanding of how to interact with them most effectively will grow.

ChatGPT: As of my training cut-off in September 2021, the record for the most goals scored in international football is held by two people from different categories:

• In women’s international football, Christine Sinclair of Canada holds the record with 186 goals.
• In men’s international football, Ali Daei of Iran holds the record with 109 goals.

However, please note that this data might have changed post my last update in September 2021, and it would be best to check the most recent statistics from a reliable source.

ChatGPT: You’re very welcome! I’m glad I could assist. Don’t hesitate to return if you have more questions in the future. Happy writing on your article!

ChatGPT Conversation Summary

ChatGPT now allows us to share the conversations we have with the AI.

Here is a link to my conversation with ChatGPT. The conversation link was created on June 28, 2023, but the actual conversation was held several weeks ago. (ChatGPT doesn’t give me an exact date. I’ll have to look in my time tracking logs.)

Why did I choose to copy / paste the conversation here? Because Google.

(Google will index this conversation, and this post will show up for others.)

Summary: ChatGPT is Biased – Implications for Education, Government and Non-Government Organization, and Changemakers

• ChatGPT is known to give biased / sexist misinformation because it has learned how to be biased from humans.
• ChatGPT doesn’t actually understand words or sentences. It understands the relationships between words and units of ideas. So if a fact is repeated multiple times in the training data, the artificial intelligence will believe that fact is the correct response (even if it is not.)
• It is difficult to get ChatGPT to ignore it’s training. For example, asking ChatGPT to not change words and simply reformat text does not work with longer passages. It will respond in a manner that is consistent with it’s training and potential bias. This has implications to reinforce stereotypes and the status-quo.