By Ellie Su ’28
Artificial Intelligence (AI) often is so advanced that it may appear sentient. Robots speak using human speech patterns; driverless vehicles motor through busy roads; programs create paintings, essays, or compositions in a fraction of a second. For some, these are magical feats, indicators that machines are developing minds of their own.
However, this notion is incorrect. AI isn’t alive or sentient. It doesn’t think, feel, or know things as we do. Instead, AI is a product of human intelligence and consists of mathematical systems used to recognize complex patterns at extremely high speeds. Understanding this fact is essential to determining how society deploys and regulates AI.
Artificial Intelligence at Its Foundations: Algorithms, Data, and Probability
At the core of AI are algorithms: a series of instructions that tell machines what to do with data. Machine learning, by far the most prevailing subset of AI today, teaches algorithms to identify patterns from enormous collections of data. These patterns enable the system to “predict” the results of other inquiries without having to hard-code each rule (Goodfellow et al., 2016). For instance, a machine learning model that learns to identify cats in images doesn’t really know what a cat is. It doesn’t represent fur, whiskers, or paws in its ‘mind’. It merely picks up pixel patterns that it determines are statistically connected to the term “cat” from scouring millions of confirmed cat photos. Thus, AI is controlled by probability, not comprehension. Neural networks, or the multiple layers of processing that characterize the “deep learning” (a subset of machine learning) used by various AI models, adjust millions of internal variables (weights) to reduce error in making a forecast. The outputs of these neural networks can be surprisingly precise, but they’re still mechanical and statistics-based (Marcus, 2022).
Why AI Feels Alive
Artificial Intelligence feels “alive” owing to three primary reasons: speed and scale, the complexity of its outputs, and its ability to surprise us.
AI has the capacity to do millions of calculations per second. It can identify correlations and patterns from vast data sets, patterns that are way too subtle for humans to identify. Such capabilities are a strong advantage in functions such as image recognition, writing out natural language, or a game of Go (Silver et al., 2016).
In addition, GPT-4-style systems have the ability to generate essays, poems, and emails that are human-sounding. Such fluency has nothing to do with creativity or self-awareness, but is a matter of anticipating using its training data, word by word, what phrase is most statistically likely to come next. This explains why we can use AI to detect AI-generated writing, alongside other works.
Finally, because AI’s training teaches it to notice patterns we aren’t even consciously aware of, its work occasionally yields results we find unexpected or “creative.” A suggestion of a song you didn’t know you would adore, or an AI-written piece of music that is unsettling yet poignant, feels like magic. But behind the scenes, it’s simply probability unleashed at enormous scale.
The Limits of AI
Despite its spectacular capabilities, AIs are extremely limited by the humans who created them. There’s a saying, “garbage in, garbage out”. If an AI’s training data is faulty, biased, incomplete, or inaccurate, then its output will be faulty too, as AI cannot check facts or make moral decisions independently (Gebru et al., 2021). In addition, language models like ChatGPT are prone to “hallucination,” or generating false or misleading information when asked questions outside their training knowledge. They create plausible-sounding answers not to intentionally lie, but because they lack any real knowledge or truth-checking ability (Maynez et al., 2020). Finally, each AI system is created, constructed, trained, and fine-tuned by human researchers. Programmers select the architectures, data sets, loss functions (algorithms that evaluate the amount of error of a prediction), safeguards, and usage policies of each AI (Marcus & Davis, 2020), and in a similar vein, AI doesn’t independently evolve or design itself, at least so far.
Failure to understand the nature of AI poses serious risks. The notion that AI can be thought of as “neutral” or “objective” merely because it is technology is simply false; like all technology, AI’s design reflects its makers’ values and assumptions, both positively and negatively. Already, AI has been applied in a way that amplifies inequality, misinformation, and bias. It has been featured in mechanisms employed for hiring, policing, and loan sanctioning, helping perpetuate racial and socioeconomic biases, since AI training data mirrors societal injustices. We might over-rely on AI in areas such as medicine, law, or government, if we approach AI systems like sentient beings; we must remember that they do not have moral agency, and not treat its work like the work of a human, when it in fact needs stringent, skeptical scrutiny (O’Neil, 2016).
Conclusion: Embracing Tools As They Truly Are
AI isn’t a mind; instead, it merely reflects and amplifies human creativity, prejudice, ambition, and fallibility. Use AI intelligently: create ethical systems, have realistic expectations, and avoid harm. The future of artificial intelligence isn’t about sentient machines; instead, it’s about human beings responsibly crafting tools that help them navigate the world, while understanding that these simulations are not reality.
