By Abhivarya Kumar
Have you ever thought about how ChatGPT is able to give you the latest fashion trends, do your assignments and help mom with a shopping list? The answer lies in how Artificial Intelligence systems are trained.
The funny thing about artificial intelligence is it can write essays, paint pictures, and beat world champions at chess, but ask it to identify a blurry turtle, and it might call it a rifle. This captures the essence of AI’s biggest challenge, it doesn’t comprehend or know what it is learning. To understand why that happens, we first need to see how it actually “learns.”
AI doesn’t learn like a human does, we don’t sit it down and explain it something, neither do we write something on a board and teach it that way. Instead, we feed it data, millions of examples of something we want it to recognize. Imagine teaching a child what a cat looks like. You’d show lots of pictures: fluffy cats, skinny cats, cats hiding in boxes. Over time, the child starts noticing patterns, pointy ears, whiskers, tails. AI does the same thing, but with math. It learns patterns in the data that helps it make predictions.
At the glowing mathematical centre of this process is a complex web of digital “neurons” called a neural network. These neurons are inspired by the human brain; each neuron makes tiny calculations and passes them on to others. Together, they build a sort of decision-making system. When you show the AI an image, it doesn’t see a cat the way you do, it just detects patterns in pixel colors and shapes and decides that these patterns “look like what I’ve seen in other cat pictures.”
Now you may ask, why does this complex decision-making system made up of neurons inspired by the human brain still mess up?
Well, the answer is simple.
The AI doesn’t really understand what it’s looking at, it only matches patterns. If an image has similar shapes or colors to what it saw before, it can get confused. Show it a picture of a dog wearing sunglasses and it might confidently declare it’s a panda. This isn’t the AI being stupid, it’s just following its training data very literally.
Let me show you just how dangerous this can be when the stakes are high. Back in the times of World War X, researchers trained a simple neural network called a perceptron to detect camouflaged tanks hidden among trees from aerial pictures. It performed amazingly well during testing. But later, they discovered the “AI” wasn’t identifying tanks at all, it was just distinguishing between photos taken on sunny days and those taken on cloudy ones. The training data had a hidden bias: all the “tank” photos were taken in bright light, while the “no tank” photos were taken in shade. The machine hadn’t learned what the researchers thought it had.
If AI can confuse a turtle for a rifle or misclassify tanks based on shade then imagine what could happen when these same systems are used for military intelligence or surveillance. In these contexts, misidentifying an image could be the difference between life or death or identifying a threat correctly or mistaking a civilian object for a weapon. With AI being used increasingly in drone technology, especially on the battlefield in Ukraine, one can’t help but wonder the disastrous consequences of having machines behind the trigger finger.
The implications of AI “guessing wrong” go far beyond harmless image recognition errors. A biased or poorly trained AI might unfairly flag someone as suspicious, misread satellite imagery, or amplify human prejudices encoded in its data. That’s why researchers emphasize that AI is only as good as the data it’s trained on, and this data is never perfect.
So, how can we make AI better at recognizing things, and stop it from repeating mistakes like the perceptron?
The answer lies in better data, transparency, and feedback.
First, AI models need diverse and balanced datasets. If all your “cat” photos come from one environment, like a sunny room then the AI will struggle with cats in darker settings or outdoors. By showing the model a wider variety of examples, it can learn more general patterns that actually reflect reality, not just quirks in the data. While this may be great in theory, however in practice a recent analysis published by MIT, AI models’ data requirements may be outpacing the supply of suitable, usable data available today. The median training dataset contained about 3,300 datapoints in 2020. This figure grew dramatically to over 750,000 datapoints in just the three years that followed. Even though the total data generated is expected to hit 180 ZB by the end of this year, it might not be enough to feed the growing demand.
Second, scientists are working on something called Explainable AI (XAI), systems that can describe why AI made a particular decision. Instead of a black box that just says, “this is a tank,” an explainable AI can tell you which features in the image made it think that. This helps humans catch and correct mistakes faster.
Third, human feedback is becoming a key part of training. You’ve probably seen those little pop-ups asking, “Was this answer helpful?” or “Click all the images with traffic lights.” Every time you respond, you’re helping an AI model learn from its errors.
But the question arises, where does all this data for training come from? Companies source data in two ways, internally or externally. Examples of internal data might include Spotify using your music history to generate a list of trending songs in your area, external data might be made up of datasets provided by the government or research agencies. There have already been several reports of AI companies using data unethically, so the search for better, more accurate answers reminds us to sometimes take a breather and think of the consequences behind the things we do.
So, the next time you ask ChatGPT to give you a good dosa recipe and it asks you to add sugar instead of salt, think what could have gone wrong in its thinking instead of just calling it stupid.
(The writer is a 4th Year B Tech Plaksha University, Mohali)
