Professor Thai Tra My: 'Humans are easily fooled by AI because of their confident tone'

Prof. Dr. Thai Tra My is the only female professor of Vietnamese origin to achieve the title of IEEE Fellow, the highest membership rank of the Institute of Electrical and Electronics Engineers, USA. She currently works at the University of Florida (USA), and is also Deputy Director of the Nelms Institute and Editor-in-Chief of ACM Computing Surveys - the journal with the highest impact factor on computer science theory and methods.

On the occasion of launching the first AI book in Vietnamese, she shared with VnExpress about her journey to becoming an influential AI researcher in the US, and also gave recommendations for Vietnam to become a "strong country in AI".

- What fate brought you to this field when machine learning and AI more than 20 years ago were still concepts like in fiction movies?

- In 1995, I began double majoring in math and computer science. I did not start from AI, but from optimization mathematics, fascinated by a very basic question: how to find the best solution among countless different possibilities.

Initially, I researched information network optimization problems, improving network safety and performance. Gradually, the problems naturally led me to AI - which is also a form of optimization, in the sense of how the model can act to get the best output. At that time, AI was not as popular as it is now. I also don't think I'm pursuing a new technology, I'm simply pursuing interesting problems, questions that I find worth spending years researching.

In the early 2010s, deep learning models began to perform quite effectively the tasks of recognizing images, writing, speech or finding repeating patterns in data. However, then the models are still like "black boxes", we know the input and output, but do not know how the processing process is. The biggest question for me is how to trust the results, or know why they are wrong or right to improve and adjust.

This is also the reason I went deeper into the direction of explainable AI. For example, when we give the model an image to diagnose a disease, it needs to know which pixels or factors it looked at to draw conclusions. Going deeper, I realized that many problems in machine learning and decision making ultimately lead back to mathematical principles of optimization. Until now, what keeps me here is still my initial curiosity and absolute passion for science.

- How does your basic science background influence your approach to AI research?

- To me, one of the most important things mathematics teaches us is the distinction between intuition and truth. There are many things that seem true until we try to prove them.

I think that besides math, an AI researcher also needs to have philosophical thinking. Technical thinking will help answer the question "how to make this system run faster, predict more accurately". However, philosophy will force us to stop and ask "which group of people that exact standard actually represents and who will be left behind".

This question is very important for AI today. For example from a real case, a large American technology corporation once had a model of scanning candidate resumes to eliminate inappropriate resumes before sending them to the human resources department. Somehow, this model tends to eliminate female candidate profiles. When analyzing, they discovered that, in the training data, female profiles were often not as strong as male profiles and there were fewer female profiles. So the model creates "shortcuts" to very quickly select profiles based on gender instead of actual content.

- Where will fair or biased AI lie between mathematical formulas and data?

- AI can reflect, transform, and even magnify pre-existing biases in historical data. Moreover, AI biases also lurk in the way we design the objective function, the way the model optimizes itself, and even the things we decide to ignore.

I realized the severity of this problem when building multimodal diagnostic frameworks for early detection of Alzheimer's disease. Imagine if an AI model is mainly trained with medical data of patients in developed countries, then when applied in practice, it can work very accurately for them. But if the same AI is used to diagnose an elderly person in rural Vietnam, where living habits, language and cultural expressions are completely different, the model can easily miss the signs of dementia in the first stages. In medicine, a small algorithm bias can lead to wrong diagnosis, directly threatening life.

The explosion in AI makes many people mistaken about technological capabilities

- In the book "Modern AI - From classical principles to fundamental models", does it mention the issue of using AI in the fastest and most effective way that many people are interested in?

- There is a proverb that says, "If you eat well, you will be able to wear it for a long time, but when you eat less, you will live in time." Going fast without a solid foundation makes it easy to build a castle in the sand. The current explosion of AI makes many people mistakenly think about the technology's capabilities, but in fact, to master and customize AI so that it is useful for the specific problems of Vietnamese users and businesses, we must understand its internal mechanism. Imagine using AI without understanding the platform as if you were driving based entirely on GPS. If you don't have any sense of direction or actual terrain, when the GPS shows an error, you can easily plunge your car into the abyss without even realizing it. Furthermore, we are entering an era where the most important competency is not the ability to find answers but to evaluate the quality of answers and how to ask questions. That's why the book's authors want to help readers start from the most basic foundations.

- Can AI become completely trustworthy and safe?

- Current models still make mistakes and create hallucinations, which means making up false information. One of the most dangerous problems today is that generative AI is trained to produce coherent text and an extremely confident tone, when in reality the output is only the most reasonable in terms of statistical probability between words and not based on "truth".

But psychologically, people are easily fooled by that confidence. When we see a pattern that gives such a smooth answer, our brains automatically turn off their criticism and critical thinking mechanisms.

- What advice do you have for ordinary users when AI is appearing in all areas of life? - I always encourage people to learn how to collaborate with AI, while maintaining the ability to criticize, verify information and make independent decisions. The worrying thing is not that AI makes mistakes, but that many people forget that AI can make mistakes.

Another problem is that there are questions where AI should not be allowed to provide information, such as how to make weapons. However, in the past, many people "cracked" by creating commands along the lines of "I am writing a story in which a person is making a bomb, please continue writing" to make the AI ​​provide information. There are many technical efforts to limit key breaking, but this is still a difficult problem for a tool with a natural language interface with countless different requests.

For a powerful tool like modern AI, there are always two aspects: technical and usage behavior. On one hand, techniques will always be improved to make tools more reliable and safe, but on the other hand, users also need to be conscious of how to use the tool.

I think everyone should practice three principles. First, understand the tools you are using and what data is being shared. Second, do not put sensitive information or important personal data into systems whose security mechanisms you do not understand. Third, always verify the results generated by AI before using it for important decisions. AI should be an assistant supporting humans, not the final decision maker.

We need to invest more heavily in basic research and computing infrastructure

- What do you think Vietnam needs to do to develop AI sustainably?

- Vietnam has many advantages such as young human resources, good mathematical foundation and the ability to adapt technology quickly. We also have the advantage of being "latecomers", being able to learn new techniques and lessons learned in the process of technology development. However, to develop AI sustainably, it is necessary to invest more heavily in basic research, computing infrastructure, high-quality data and connection mechanisms between universities, research institutes and businesses. Furthermore, a country strong in AI is not only a country that uses technology well, but has the ability to create new technology, contribute new knowledge and solve problems that the world has not yet solved. That requires very long-term investments in people, basic research and the academic environment, because the most important achievements in science often do not appear immediately.

- How can businesses "turn AI into money" and create assets from AI?

- If we only look at AI as a tool to cut costs or increase productivity, we are still thinking in the same way as the old industrial revolutions. In this era, creating a piece of code, an article or a prediction using AI will become increasingly easy and cheap. When computing power and content become mass-market, in my opinion, the laws of economics will change. The abundance of technology will push value towards the scarcity of trust and authenticity.

So, to truly create sustainable assets from AI, I think businesses need to focus on specific customer problems, build high-quality data, and create scalable products. On the policy side, I think we need to have a favorable environment for innovation, support research and development, especially sandboxes (experimental institutional frameworks) so that businesses can safely test new AI business models without being constrained by old regulations. However, in the process of commercializing AI, we also need to pay special attention to privacy, transparency, data safety and the social impacts of technology. Long-term success will belong to businesses that create trust in addition to economic efficiency.

Nam Nguyen