What I learned about AI

AI gains control of a spaceship overpowers the human guards and steers it to a new planet. In 1960, Arthur C. Clarke wrote the novel 2001: A Space Odyssey, in which he imagined what AI would be like in the 21st century. Though he turned out to be mostly correct, he also overestimated AI. Over the summer, I worked with Felidae (felidaefund.org), a non-profit organization dedicated to protecting big cats worldwide. To protect animals and their habitat, the researchers must first know where the cats hunt & live. To do this, they created Wildepod, a project to catalog trail camera data. The Wildepod team had been collecting images for over five years and was looking to analyze and improve the efficiency of their image processing pipeline. Given my love for animals and computer science, I was thrilled to be part of this project. I learned a plethora of new information about AI by working with Felidae. I learned that AI is sometimes surprisingly correct, but also sometimes obviously wrong, but still only as good as its training data.

Sometimes, AI can find things a human eye would never notice. I once saw an image that the AI algorithm had labeled as a fox. At first, I thought it was a tree. However, I zoomed in and realized that what I thought were branches were the limbs and the tail of a fox! This experience made me realize the incredible potential of AI in image recognition. While patterns and shapes can trick our human eyes, AI algorithms like You Look Only Once (YOLO) can detect and classify objects with remarkable accuracy. 

However, AI can make critical mistakes as well. While working with the data, I saw an image YOLO labeled as an American Badger. I was surprised because I didn’t think there were any American badgers in the area where the camera trap data came from. I opened the image, and lo and behold, there was, drum roll please, a sparrow. This incident highlighted a critical aspect of AI: the amount of errors it makes. Despite the advanced algorithms and extensive training data, AI systems can still misinterpret information. This experience underscored the importance of human oversight in AI applications. While AI can process vast amounts of data quickly and identify patterns that might elude human observers, it is not perfect. Human oversight is essential to verify and correct AI outputs, ensuring accuracy and reliability.

AI is often said to be the epitome of human science, an infallible technology that will one day surpass human intellect. However, AI is extremely expensive and time-consuming due to the heavy computing requirements.  Another key takeaway is that AI is only as good as its training data. If you train it on data that says 1+1 = 3, the AI will believe that 1+1 = 3. This means that AI can only make inferences about the past because it is trained on data from the past.

In conclusion, I learned a lot about AI from working with Felidae. I gained a deeper understanding of the capabilities and limitations of AI in real-world applications. While AI has the potential to revolutionize fields like wildlife conservation,  it is not without its flaws. The instances of AI’s remarkable accuracy, as well as its stupendous errors, emphasized the need for continuous human oversight and intervention. Moreover, the high costs and time-consuming nature of AI processing highlighted practical challenges that must be addressed.  While Chat GPT might be good about studying for your history test, it won't be able to tell you who will win the next presidential election. This journey not only strengthened my passion for computer science and animal conservation but also instilled in me a greater appreciation for the balance between technology and human expertise. As AI continues to improve, we must remember that AI is a great tool for human aid, not a replacement.