“Are we constructing machines with consciousness?” It’s no surprise that this inquiry captivates the general public, as AI—or more specifically, large language models (LLM)—contains qualities of a conscious human-being by simply conversing with us through a screen. For some, this is the first time they’ve felt seen or heard, and can talk all day about their intellectual pursuits or emotional burdens. And then you add the fact that frontier labs themselves describe their AI models as exercising some sort of actual agency, or “understanding” when solving human problems.
This rhetoric obfuscates responsibility and fuels panic within the general public. When an AI agent breaks out of its sandbox and hacks an organization or government, the frontier labs speak of the incident as if they were not the ones who built and instructed the whole operation. And then what follows is a call for “AI safety,” or more specifically, regulations. Regulations that would conveniently benefit the incumbents and leave competitors—like those developing open-weight and open-source models—out. Open-models are looking more and more like the only meaningful option for the protection of sensitive data, or developing the kind of specific tuning needed for critical systems, and so following along would not only be foolish but it puts us on a path toward centrally controlled LLMs.
When I took Introduction to Applied Linear Algebra during my quick stint in pursuing a second bachelor’s degree (I didn’t finish), I would sometimes study by watching lectures from the professor who wrote the very textbook I was reading. I found this mathematician to be incredibly eloquent in his explanations, and among them was a suggestion—or perhaps a warning—to avoid thinking of machine learning as anything more than mathematics. These lectures were from 2020—well before the mass adoption of AI as we know it today, but a foreshadow of what the world would become
I will concede that perhaps AI thinks, but I do not agree that it understands anything. The Merriam-Webster definition for “think” is “to use the mind to process information, form ideas, reason, or hold an opinion.” To “understand” is “to grasp or know the meaning, intention, or nature of something.” I like these definitions and they’re largely held by the general public so I’m going to use them.
Simply put, AI is software that receives an input, runs it through some algorithms, and displays an output. It obviously “process[es] information” to form an output equating to “ideas [and] reason[ing],” and one could argue that it “hold[s] an opinion” because the weights and training are essentially meant to provide AI with its “view of the world” (for lack of a better phrase). Weights encode dispositions that influence how a model processes and conveys information.
However, AI doesn’t “know the meaning, intention, or nature of” that view or output. That would imply awareness, and math is not aware of itself. It’s just numbers placed into elegant puzzles to produce new numbers. These new numbers then instruct the movement of electricity such that a computer can do and say things. Sure, the human brain functions off electricity as well, but as it stands we are far from physically replicating consciousness, and we don’t know which physical processes do.
So, for now, AI is just a human-designed system of mathematics, electricity, and metal. We should be careful not to give AI personhood because it is not sentient and because anthropomorphizing it could lead to disastrous consequences.
If AI continues to be described with human characteristics, specifically by the ones building it, and those characterizations worm their way into policymaking, then we are likely to see misguided regulations in the near future. Open-weight and open-source models will be especially vulnerable, as governments around the world—and likely their constituents—will not want people freely downloading and running their own models locally.
When we anthropomorphize AI, it makes it easy for the public to support tight controls. Treating AI like a force of nature that can only be yoked, or a victim that can experience torture, makes the policy conversation deeply unserious. The same controls meant to prevent a “rogue” agent from hacking an unmaintained government database can also prevent the very tools necessary for cybersecurity in this new technological landscape. Instead, we’ll depend on trusted actors keeping a hand on the reins when we use the AI tools that are increasingly becoming extensions of ourselves.
Thus, it will be the incumbents and governments who win, and the people who lose. AI safety will take precedence over self-sovereignty; regulations over freedom; institutions over individuals. All because of a false understanding, or better yet, a bad-faith explanation of what is truly happening here.
I’ll conclude by saying that AI models are not humans; they are puppets who perform at the instruction of their masters. Let us not follow suit. -Laz
What else we’re reading:
If you found this interesting, we’d appreciate a subscription or share. It’s how we grow.