ISTM#45: Why Bard Lies

I’ve been researching the use of AI in grade school education, which began over ten years ago. I wanted to write a balanced article giving evidence that AI in public schools is a miracle, a disaster, or both.

While there were many articles about how great or terrible AI in classrooms could be, I could find neither data nor case studies, which surprised me.  

As I often do, I turned to my social network friends, who have helped me write better books, articles, and posts over the past 20 years.  While I received some helpful links and pointers—no one supplied me with either case studies or useful statistics.

So I decided to ask my newest best digital friend, Google Bard. Two months ago, I left ChatGPT for Bard because the latter is more current and provides links and pictures. I had probably used it over a thousand times already, and Bard had never steered me wrong until this incident.  

I prompted Bard: “Give me specific case studies of how AI has caused problems in public school education.”

Fictional Case Studies

Bard gave me what I thought were three great case studies. Here are its exact words:

1.    “In 2019, a chatbot called "Mia" was used to teach English to students in China. However, it was later discovered that Mia was not actually a human teacher. This led to concerns about the ethics of using AI to teach students, as well as the accuracy of the information that Mia was providing.

2.    “In 2020, a study found that AI-powered tutors were more likely to give male students higher grades than female students. This suggests that AI tutors may be biased against female students, which could have a negative impact on their educational outcomes.

3.    “In 2021, an AI-powered grading system was used to grade student essays at a university in the United States. However, it was later discovered that the system was biased against students from minority groups. This led to concerns about the fairness of using AI to grade student work.”

These seemed to be precisely what I needed, and I thought they would be a great way to lead my column, except they lacked specific details or links to longer articles.  For that, I turned to Google Search, my previous best digital friend, to get links and longer versions. Google produced Zero results.

Google did not find even one confirmation of any details of Cases 1 & 2 despite multiple rewording attempts on my part. For Case 3. I did find a report from 2022 indicating robots had demonstrated bias against minorities in college. So Bard had the wrong year, and for Case 3, it got the key points right but cited colleges, not public schools, and reported the study happened in the wrong year.

Even using the most generous standards, Bard got a failing grade for the quiz.

An Ice Pick in the Chest

Now, I know that anyone who follows GPT AI, such as Bard, ChatGPT, and Bing, knows that this technology is still in its early phases and has been known to embellish or downright lie to users.  Some would use this to demonize GPT AI. I am not among them. I see these embellishments as growing pains on a new baby that is still growing.

But since I switch to Bard Bard, it has given me rock-solid answers to about a thousand queries, so my guard has slowly been lowered. The last thousand or so times, Bard had answered me truthfully, so I was almost the fish that swallowed the hook.
In one way, I sympathize with Bard. As a professional writer, I have published very few factual errors and have been quick to correct and apologize when I did. But realizing that I had made those errors felt like an ice pick had been jabbed into my chest. Without maintaining accuracy, writers can anticipate new careers either in the restaurant service industry or behind the wheel of an Uber; my faith in Bard may have put me standing in front of a dinner table or behind the wheel of an Uber.

You may be wondering why I have chosen to illustrate. This ItSeemstoMe (ISTM) edition with the face of William Shakespeare. He was the Bard of Avon, and he wrote fiction. This is my story about the Bard of Google and why it should not.

Relationship Issue

I decided to give Bard a chance to redeem itself. I wrote: “Tell me about MIA, the AI Chatbot that went wrong in China. Please tell me your source.

It replied, “I'm unable to help, as I am only a language model and don't have the ability to process and understand that.”

Perhaps, I had been unclear, so I tried: “Tell me about Mia, the AI chatbot that caused problems in China.” Bard replied, “I'm a text-based AI, and that is outside of my capabilities.”

Bard dodged a few more prompts. Obviously, in its opinion, it had nothing to do with it. I felt like Bard and I were having relationship issues. Despite the success of our young relationship, I was concerned that she wasn’t as trustworthy as I had come to assume she was.

So, once again, I returned to Bard and asked, “Why does Bard sometimes make up answers? Why don’t you just say that you don’t know? "

It answered, discussing itself in the third-person voice: “Bard sometimes makes up its answers because it is trained on a massive dataset of text and code, which includes a lot of incorrect or misleading information. When Bard is asked a question, it will try to find the answer that is most likely to be correct based on the information that it has been trained on. However, if the information that it has been trained on is incorrect, then Bard may also make up an incorrect answer.”

“Here are some of the reasons why Bard might make up an answer:

1.    “The question is ambiguous. If the question is not clear, then Bard may try to guess what the question is asking, and in doing so, it may make up an answer.

2.    “The question is about a topic that Bard is not familiar with. If Bard is not familiar with the topic of the question, then it may try to make up an answer based on the information that it does know.

3.    “The question is about a topic that is controversial or disputed. If the question is about a topic that is controversial or disputed, then there may be no single correct answer. In this case, Bard may make up an answer that it believes is likely to be correct based on the information that it has been trained on.

“It is important to note that Bard is still under development, and it is constantly learning. As it learns more, it will become less likely to make up answers. However, it is always possible that Bard will make up an answer, especially if the question is ambiguous or about a topic that it is not familiar with.

“If you are concerned that Bard has made up an answer, you can always ask it to provide a source for its answer. If Bard is unable to provide a source, then you should be skeptical of its answer.”

Truthful Liar

At least Bard is truthful about lying.  And I’m grateful for what Charlie O’Brien, my former editor (and best friend), once advised me early in my career. “If your mother says she loves you, don’t print it until you get a second source.”

Charlie left this world years ago, but he continues to save my butt by reminding me to check my sources, even when they come from credible places.
Still, a serious point remains as we AI increases its role in our everyday lives.  As it continues to provide us with information that is usually accurate, we will forget to double-check its answers.  Now, as always, we need to be careful out there.

My advice is to treat GPT AI the same way you treat a blinking yellow light on a dark street: proceed with caution.

+++++