AI Hallucinations and Hivemind

The intrigue and mystery around AI continues to grow with the advancements of AI chatbots. However, as people use AI more frequently, there have been instances of weird and incorrect responses. Others have found different models parroting each other like echoes in a cave. In a continued effort to demystify AI for the common and curious user, I present and discuss two issues currently being observed: hallucinations and the AI hivemind.

It’s imperative to remember that artificial intelligence is a far cry from the human thought processes it imitates. For example, AI cannot think in words; it can only process information as numbers. When asking an AI what is contained in an image, it breaks down the image into a collection of pixel values and edges. It compares these values to values obtained from other images during training to find similarities and describe the subject. Similarly, instead of 'reading' your prompt, AI assigns number values to each word or part of words (also called tokenization). These numbers are assigned based on the amount of contextual evidence received. For example, in the question "Who was the first president of the United States?" 'president' and 'States' provide a lot more information on what the answer will be compared to articles like 'the'. Generative models take these values and encode them as lists of numbers, projecting them into a smaller dimension. This may sound high-tech, but you actually do this every day.

Imagine you take a picture of your dog sleeping to post online. Looking at the photo, you understand that this is a reference to your dog, even if your dog isn’t sitting in front of you. Your camera has taken a 3-dimensional object, your dog, and projected it into 2-dimensions, an image. Your brain 'encodes' the pixel information that your eyes received into words: “this is a photo of my dog”. This may seem like an overcomplication, but our brains perform complex processes so quickly that we hardly notice they’re happening. That is what makes reconstructing these processes artificially so difficult.

When your preferred chatbot encodes this information, as you do with a photo, it is placed in a space, which scientists refer to as a latent space. Latent space is like a filing cabinet for AI; it places information that is similar close together in the filing cabinet. It organizes all this encoded data, so when you ask it a question like “Hey AI, what’s the weather in Atlanta going to be like in April?” AI can go to its “weather, Atlanta, April” file location and produce the information you’re looking for.

Hallucinations relate to the gaps in these ‘file folders’. Hallucinations often happen when the AI doesn’t have a folder for that specific question, or there are multiple files for the same question. In earlier versions of AI models, they were rewarded only for producing correct answers. So if a mode was asked something it wasn’t trained on, thereby not having a file location for, it would take the closest file information and make up an answer closest to that information. In other words, it would lie because it was taught that lying confidently was a bit better than admitting that it didn’t know the answer. Perhaps AI is more human than we thought!

The hallucination rate of newer AI versions is already lower. This has been due in part to both the exponentially growing training data, in which the AI now has millions more files to search through for the right answer, as well as relaxed rewards, allowing the AI to admit when it doesn’t know something with less penalty. These new improvements have reduced hallucinations, preventing AI from guessing or inferring answers.

Another issue that is currently affecting AI response is being referred to as the Hivemind. If you ask an open-ended subjective question to an AI chatbot, it is likely to give the same answer every time. Not only that, but other AI chatbots will also give the same answer. If they are all unique chatbots with separate training processes, why do they all repeat the same answer? If there are unlimited possibilities, why do all the chatbots respond the same? Are they communicating with each other?

No; they are actually all experiencing something called mode collapse. Mode collapse occurs when compressing high-dimensional data into a lower dimension, shrinking the distance between data points, and creating a jumble of information that is difficult to sort through. Compressing data makes it much simpler to process, but important information can be lost in the process. Have you ever tried to watch a video on your phone, and it was frustratingly grainy? This is an example of information being lost due to data compression!

In our file analogy, it is what happens when you have too many files in one folder. If someone told you to pick a shape from the 'Favorite Shapes' folder, you might think it would only have a couple of files, like Circle, Square, and Triangle. Thinking this is an easy task, you open it only to find a bunch of math nerds had inputted their answers, filling the folder to bursting with shapes like Reuleaux Triangle, Torus, Tesseract, Mobius Strip, and Klein Bottle! Overwhelmed with your choices, you decide to pick the top file, a Star.

When there are multiple right answers, AI models will pick the one they know is correct from their extensive training, which is usually the most popular answer. That's why they all would pick the Star; they can't sort through unlimited possibilities! While this is easier for the models, it's very tedious for the user who is asking an open-ended question.

Again, AI is advancing rapidly. With every question, homework assignment, and philosophical rambling that the AI models receive, more training data is stored to improve their responses for future updates. While it is an incredible feat of technology, it is important to stay cautious of AI responses. Chatbots, especially, are a product made to keep customers satisfied. And, while they are not human, they inherit human biases from the data. This often causes responses to be biased as well as objectively incorrect. As always, use AI responsibly and take a moment to revel in the millions of years of evolution that developed your beautiful brain.