Objectivity. A lost virtue in the age of GenAI
I was just reading an old 2018 paper on computer vision and a thought suddenly struck me.
Objectivity is lost in the age of GenAI.
Or at least, the facade of objectivity is lost. Depending on your personal philosophical leanings.
Take for example, Structure from motion (SfM). It used to be that people had to calculate homography matrices and perform triangulation. It was cumbersome and complicated. But if you had the right information like the camera’s focal length and all, you’d arrive at a physically sensible solution. In other words, you would know the camera’s pose (rotation, position,etc.) and you’d have a measurement of the object’s distance based on hard math. Nowadays with segment anything and depth anything AI models, you’d be able to get a depth map of the objects in the picture with little to no work. The catch is, everything is now relative. The depth map (or the distance each image pixel is from the camera) is what the AI model thinks it is. The change in depth distance between 2 pixels is what the AI model estimate it to be based on the images it has seen before.
The origin of this thought was the Deep Structured Generative Models paper, which led me to think about how I would do the same with GenAI (perhaps more on this in a later post). It then led me to think about how I would use a combination of image generation with background removal to generate assets, use LLMs to generate semantic information and video generation to generate different views of the digital assets, etc..
💡
Actually I started reading the paper because I had problems with pure image generation via ChatGPT when creating thoughtsre’s logo.
Anyway, it then occurred to me that each of these steps feels like I am asking a person to do something for me. I didn’t have to know how to calculate homography matrices anymore. Each step was essentially a black box.
This may be convenient for me. Just like how a company outsources certain functionality so that it can focus on its core business. We can take this analogy further and think about what happens when a company unwittingly outsources its core business functions in pursuit of efficiency. History is replete with bad examples. Another story for another day…
Coming back to the thesis of this article, when one asks another person to do things, the outcome is at least partially determined by the inherent biases and capabilities of the person. This is why we keep comparing between models of different companies and how they perform on different benchmarks. It’s like how we compare the design proposals of various interior designers before we decide to engage one.
Setting success criteria and measurement metrics help. But think about it, if it were a human, there is always some element of subjectivity that is due to the human. Increasingly this will be true for machines as well. The output will be colored by its training data (or experiences). If we blindly accept its outputs, then we are also accepting its subjectivity, so to speak. Just like how if you employ a director to make a movie, you are buying into his artistic direction and vision.
Going back to the analogy of interior designers, it’s like judging the designer’s capabilities by their use of space or adherence to your requirements. Two designers might use space equally efficiently but one is just better at understanding your asks, perhaps because his life experiences and tastes are just more similar to yours.
To try and define success exhaustively would be to make the work of metrics definition an industry unto itself. Then where’s the productivity gain? In any case, it’s a fool’s errand. AI’s capability will far exceed human capability of monitoring. See AI 2027 or recent incidents on AI models employing blackmail to avoid shutdown.
One might argue that this is human society works. We deal with this kind of subjectivity all the time. All humans hold subjective opinions. It can work with machines as well. To wit, I say maybe.
Note that, despite the progress made, human society is still filled with suffering and unfairness. Can we assume that a human society mixed with machines will be better? Moreover, there are many more humans each developed with their unique experiences. The diversity averages things out. And human society has had a lot of time to derive common ground, albeit through painful experiences.
Can we say the same of large AI models in the hands of a few?
Others might also argue that objectivity has never existed. That is not true in select fields like mathematics. But to be honest, mostly true in everyday lives. We have, however, found ways in human society to deal with it. We have laws and cultural norms that are distilled over time.
Importantly, we have, as a species, valued objectivity and at least tried our best to uphold it. This is why we have laws that require proof beyond doubt and chain of evidence to convict crimes. That is why we pursue knowledge in hard science. That is why it’s different when we meet face to face, instead of via Zoom.
With GenAI’s ability to create realistic content, it is hard and mostly too tiring to find out what is true anymore. We will be inundated by generated content delivered to us via the well-hone machinery of social media that constantly sways our thinking and judgement.
At the end, what makes humans want to work with others is trust, which comes from cultural similarity or some other binding force like the law or familial bonds. And I think there are a few ways where we might be able to salvage objectivity.
One, devise a legal framework to treat AI as persons. We urgently need to adapt our current legal framework to treat AI as persons with legal liabilities. Fortunately, we already have what we need. A company is a legal person that is not a human. Hence, just like a company needs to be capitalized and there may be humans (up to a certain point) that are liable for its actions, the same goes for AI.
Two, adapt the use of blockchain as a record of AI actions and human outputs. With the proliferation of GenAI, authenticity becomes a prized commodity. This is where AI actions and human outputs need to be recorded in a globally immutable records like a blockchain. On top of that, with the help of the legal framework, the outputs of AI have to be watermarked, like SynthID.
Three, use small models but lots of them. One of the key dangers in the current GenAI revolution lies in the fact that only a handful of operators own and are responsible for training these models. The reason for that is that these models (even the so-called resource efficient DeepSeek models) are so expensive to run and train. There are some that would advocate stopping or slowing the development of AI. I personally don’t think that will work. The cat’s already out of the bag. In fact, I feel that the research and engineering community needs to urgently work out a way to have the same capabilities but in a much much smaller form factor. In other words, find a way to enable AI models to have high intelligence but lightweight enough to be deployed on mobile devices. Importantly, these AI models should be able to learn quickly, easily and independently, such that they are truly personal assistants to people. There should be at least as many independent AI models as there are people. To do this, we would need to shift the current AI paradigm away from the transformer-based models that are super heavy and achieve quantum leaps in AI hardware. In this way, we achieve a parallel “machine society”, where hopefully diversity of independent AI agents helps to average things out and we can seek some form of objectivity within.
That being said, it is unclear to me whether any of these methods will work in the long run. Maybe it will only serve to delay the inevitable. Take for example the third suggestion of creating a parallel machine society. How we would ensure that the resultant AI agents would really be diverse? Even if they were, would humans be seen by the machines as a cohabitant or a parasite?
I’d love to hear what you think. Leave a comment and let me know your thoughts.