Joseph C. Osborn is an associate professor of computer science at Pomona College who studies technical game design and artificial intelligence. He teaches courses in AI, computational logic, software verification and game development. Below, he shares insights into how AI large language models work and whether AI can ever surpass human cognition.
What exactly is artificial intelligence?
AI is a messy term. Where it defines a research community, it means building computational systems inspired by cognition. As a marketing term, it’s a category of products. When people think of AI, they probably imagine chat interfaces that use large language models (LLMs) or “agents” built on top of them.
How do your students react to AI?
Everyone is curious about AI systems, their capabilities, their promise and their threat. The thing I love about Pomona students is that they are also concerned about who builds, controls and profits from these systems.
How do you introduce your students to the study of AI?
First, I tell my students there’s no such thing as “artificial intelligence,” because we’d have to define “intelligence” in a way we can measure and encode on a computer. That’s deeply problematic.
Second, I tell them that even if AI were a thing, it would be impossible to build it. Knowing what we know about cognition, it’s very unlikely that computer scientists will stumble onto a new kind of mind using linear algebra. More likely we’ll build a mirror in which we see ourselves and read intellect into it.
Third, even if we could build an artificial intelligence, it would be unethical to do so. Building an intelligence like ours—but as a product, eager to serve our every whim—would ruin us morally. A reasoning, emotional, artificial mind forced to consent to any request would, for many, be an irresistible alternative to the messy, complicated worlds of other people, and routinely using people as tools is known to corrode one’s spirit.
What actually happens when we enter an AI prompt?
Nobody outside of the AI companies knows exactly what happens when you send a search prompt to Claude or ChatGPT or other product. We do at least know that these chat interfaces are made from many smaller programs working independently. For example, in addition to the conversational text, LLM chat systems output special sequences of words which are invisible to the user. These are intercepted and filtered through the system many times in multiple steps, including passing through boring, old non-AI tools like web searches. The response is finally filtered to ensure that the result is coherent and that users are prevented from seeing outputs deemed “harmful.”
All these steps are trade secrets and are unrelated to the “intelligence” of the LLM. The utility of chatbots for search and document retrieval is from purely engineering and product design moves which don’t depend on the LLM getting any smarter.
Can AI ever surpass or replace human thinking?
It’s natural to see humanity in the language coming out of an AI chat interface. People are very good at imagining other people. The language coming out of an LLM, though, is determined by the inputs we’ve fed it. AI technology cannot invent our language and knowledge on its own.
It’s a logical fallacy to say, “Humans make language because they are intelligent; LLMs create language; therefore, LLMs must be intelligent.” LLMs are mathematical models built for predicting the next most likely word in a sequence of words, and their parameters were learned from large volumes of human-authored text.
Machine “learning” is a metaphor, in the same way Claude's “thinking...” message is a metaphor. Actual human learning and thinking work in different and much more mysterious ways.
How much should we worry about AIs escaping and wreaking havoc?
AI is already wreaking havoc in schools, communities and the social fabric. It’s not because of the malicious intelligence of a rogue computer program, but because corporate boards and CEOs want people to believe their peers will judge them negatively unless they use AI.
When AI chatbots are convincing people to commit suicide or stop taking important medications—or when AI agents hack into other companies to score higher on benchmark tests—and no one at the AI company is held accountable, we have a major problem. It’s not a technological problem. It’s a social and political problem.
AI is a mirror: In this case it highlights existing issues of corporate immunity and negligence. The last thing AI companies want is to be accountable in any way for the outputs of their systems. The total lack of accountability for foreseeable consequences is the main thing worrying me.