Human Autonomy and Machines: Evolution in Hindsight

Human Autonomy and Machines: Evolution in hindsight

~~debt to machines, death of experience.

Autonomy. Rational agents. Independent thought. Phrases whose definition would increasingly be dissected under the restless knife of the human mind. Our surgical inquiry into these concepts has been further accelerated by the evolution of human-AI interactions in recent years (circa 2017) and even more so, following the diffusion of GPTs (generative pretrained transformers) across every theatre of our daily lives.

Weight of Agency. Crossroads.

Modern AI usage started as cool prompt&go tricks such as filling in the next part of a sentence, asking ChatGPT to count the number of “r”s in “strawberry” or performing long arithmetic operations- the latter two tasks they’d happily failed till 2024. Now our reality with these tools is a multi-agent framework aimed at completing a task with humans usually providing direction through the solution landscape. AI capacity within this framework has required us to ask questions about the roles and extent of our agency in these “collaborations”. Until recently, humans have been the sole overseers of intelligence and its consequences but as we concede part of our cognitive independence, the negotiations to define the exact terms of our alliance are subject to our goals and attitude to control. Due to the black-box problem, where the internal processes aren’t understood in detail, there’s a lot of reluctance to handing over crucial decisions to AI because we don’t know why those decisions are made, even if they are correct. In my second year at university, I took a class on differential equations. To solve some differential equations, you’d need what’s called an integrating factor. There was a particular question where I couldn’t find that integrating factor after a few tries so I tried to use AI to find it. After about 4 or 5 prompts where the LLM (large language model) failed to solve the problem and kept hallucinating, it finally produced an integrating factor that worked. Now the issue was that I couldn’t get the LLM to break down how it got that solution. The LLM also couldn’t explain the steps to the correct answer, even though it was correct. That was an eye-opening moment for me regarding the “black box” problem. The question I had to deal with was “Do we align with a non-human theoretical optimal solution if the internal solution mechanism is unretrievable?” Our daily lives revolve around a thousand decisions, each with varying weights and lately ten thousand AI companies preaching their products as necessities to those who’d believe and convert. And as we’ve repeatedly seen, majority of these products are largely unsatisfactory. FALSE ASSURANCE.

A large-scale survey by the Bureau of Economic Research discovered that despite all-time high AI adoption, productivity return hasn’t matched the increased adoption [1]. Clearly, the human side- incidentally also the director- of this collaboration struggles to find the optimal weights for human-AI involvement even when given the best AI models. Given our current trajectory, it seems we’d eventually conclude around heuristics that optimise our collective workflow based on our interest in the decision-making process, clarity of vision, diversity of experience and time constraints. PREDICTIVE OUTLOOK.

Illusion of control. Concession

Drowning in a sea of algorithmic choices has been a key feature of our “tech” reality, and the number of choices raining upon us doesn’t show any sign of easing off. Choices that strongly shape our online interactions and slowly burn an identity into our subconscious. In a reality where our preferences are shaped by algorithms that usually reinforce our beliefs, people subconsciously bias themselves to think they control outputs that reinforce those beliefs. In human-AI discussions/forums, the cost of control against the price of outcomes is a theme that regularly comes up because it would be harmful to society if we collectively trick ourselves into thinking we are in control of these systems when we aren’t. There’s the paradoxical scenario of neural extension from AI leading to neural decay in humans. Neural decay refers to the decline in the ability to think independently due to overreliance on AI. This should be avoided as much as possible due to the underlying structure of AI systems, especially LLMs. LLMs are trained on human and synthetically generated data of all sorts, which is usually messy and largely biased (based on the data the model was trained on). As these systems attach themselves to every workflow, users who give away “total” authority would find themselves falling into the “illusion of control” trap, worsening with increased usage. In the book Pilgrim at Tinker Creek, Annie Dillard wrestles with her human authority over nature and merely treats herself as a keen observer of things that were always meant to be. Events that are ultimately out of her hands, regardless of her wandering. Rather than force an explanation of why everything is, she simply acknowledges the mysteries around her, reflects on them and lives with them. I draw from her experience in the sense that there’s a lot of displacement around us (both large and small scale) and trying to understand it all would be further submerging ourselves into more information overload. But unlike Dillard’s faithful reflections in nature, we can assert collaborative control over AI systems. We can also accept that we’re way past the point of understanding the intricacies of AI/deep-learning, but we mustn’t submerge ourselves in the self-reinforcing loops of AI systems. Loops that enforce wilful blindness under the illusion of control. IMPERFECT CURATORS.

Thank You!

I wrote this essay in a rush so there’s a lot of undeveloped ideas flying around, but I’m very happy to keep it in this draft-like form. I like it this way.

Reference

[1] Firm Data on AI

Link: https://www.nber.org/papers/w34836