The Mix 2025: Go under the hood on GenAI
By Patrick Gohl (Pompano Beach High School), Deolu Akingbade (Davidson Day School), and Teeba Reab Agga (Cary High School)
“I’ll ask ChatGPT.”
Since generative artificial intelligence) entered the mainstream in the early 2020s, this phrase has become ubiquitous, uttered by everyone from students seeking quick research for a paper to people looking for a recipe that uses whatever they have in their fridge.
In recent years, generative AI has found its way into nearly every facet of life. Most people, however, have little idea how the answers generate.
“Most people don’t realize what technological breakthrough led to generative AI,” said Scott Geier, UNC professor of journalism and media.
What is generative AI?
The term “AI” encompasses a variety of technological systems, all of which are designed to mimic human intelligence in some way. This imitation can be in pattern recognition, organization, content generation or other forms. This is achieved by “training” the AI on a set of existing data, which could include text or images.
A key feature of AI systems is that they are able to refine their accuracy based on new data, without having their programming updated by a human. This process is known as machine learning.
Machine learning, a key feature of AI systems, allows AI to refine their accuracy based on new data without being updated by a human.
Generative AI is the type of AI most people use on a daily basis: AI designed to generate new content. This can include chatbots, such as ChatGPT and Google Gemini, as well as image and video generation software, such as Midjourney and Adobe Firefly.
How does generative AI work?
When you provide a generative AI with a prompt — What large dog breeds are good with kids? for example — it first breaks the prompt down into tokens, or units of meaning which the AI can compare. These tokens can be words or parts of words for text and groups of pixels for images.
The large language model (LLM) can then compare the tokens in the prompt to those in its dataset. It will focus on what it detects as being the most significant tokens in a sentence and compare these tokens with ones that have similar assigned “positions” in its data set. In the aforementioned example, these would likely include words like “large,” “dog” and “kids.”
Geier said the transformer — the ‘T’ in GPT — was the breakthrough that developed generative AI. Before this, AI could only make predictions in a linear fashion.
“That was the algorithm behind predictive text in email and smartphones, e.g., when you typed ‘I like ice ___,’ AI could only predict the most likely word after ‘ice,’ e.g., ‘cream,’” Geier said.
Transformers, however, are a technology that enables AI models to analyze billions of tokens at once, generating answers based on the entire prompt. This is what elevated AI from making basic predictions to generating new content, enabling the emergence of complex chatbots and image generators.
However, since AI systems evolve without having their code directly altered by human coders, experts often cannot fully understand how AI systems, such as LLMs, arrive at specific answers. They will often demonstrate new “emergent behaviors” as they are trained. This model of AI, which nearly all widely-used generative AI tools use, is known as a “black box.”
How are people using AI?
As AI continues to evolve and adapt, it has found a place for itself in day-to-day life. According to a study by the Harvard Business Review, people commonly use AI for personal help, like therapy, healthier living, and finding purpose, largely because of its accessibility.
AI is also used by professionals in specialized industries. Programmers, public relations specialists, and hiring agencies all use AI to streamline tedious tasks and augment human creativity. Navigation, weather, and streaming services are all examples of AI being embedded in many services without users realizing it.
What’s the good, bad and ugly of AI?
Thanks to its logical reasoning and pattern recognition, AI can do tedious work — like taking meeting notes, sorting documents, or handling basic customer support — that humans generally detest.
Many firms have adopted specialized language models, freeing up employees who can focus on urgent and important tasks. Generative AI is good for providing structure and getting rid of repetitive work, but not for original thoughts that require high accuracy and intuition.
Despite its ability to synthesize large amounts of data quickly into written works, AI is not ideal for generating works where accuracy is paramount.
When presented with a prompt, an AI chatbot will try to generate an answer for the user from its data, regardless of whether it can provide an accurate response. This can result in hallucinations, a phenomenon where AI presents inaccurate information or fabricated sources to try to provide an answer.
As AI becomes indistinguishable from human products, it has a murky ethical position due to its lack of government oversight and struggles with transparency. Creatives have criticized AI-generated images and copy, with many fearing that AI could take over entry-level jobs.
What environmental concerns are associated with AI?
The significant environmental impact of AI is also a concern. According to researchers at the Massachusetts Institute of Technology, AI data centers use more electricity than the entire nation of Saudi Arabia and consume 39 million gallons of freshwater daily.
“The data centers that power AI are huge energy sucks that might eventually overwhelm our outdated power grid, and that power has to come from somewhere,” Geier said. “Until renewable energy becomes ubiquitous and cost-efficient, that power will be generated by fossil fuels.”
AI’s heavy usage of fossil fuel, and the resulting impact on the environment has added to its controversy.
Experts are working on solutions to make AI more environmentally sustainable.
“One of our computer science professors at UNC, Tianlong Chen, is working on creating a modular framework for AI models that divides the processing power into a series of smaller, less energy-intensive systems,” Geier said.
However, solutions like these, as well as those utilizing renewable energy and a lowered emphasis on water cooling, could take decades to complete. Generative AI has become the latest battlefield in political debates surrounding climate change and greenhouse gas emissions.
“There is a Dickensian quality to the use of AI when it comes to our environment: It can make our planet better, and it can make our planet worse,” Sen. Ed Markey, D-Mass., said in a statement. “The development of the next generation of AI tools cannot come at the expense of the health of our planet.”
Despite these political controversies, Geier still said AI will continue playing a crucial role in the future.
“AI will be ingrained in nearly everything we do within 5-10 years,” he said. “And often, we won’t even realize it.”