The Janus Face of AI

The world was initially taken by storm with the arrival of ChatGPT, an artificial intelligence (AI) that could answer any of your questions with very good accuracy. Though its answers were sometimes made up and often at least partially incorrect, its replies could, to a certain extent, even be called human. Many feared they would lose their jobs, and what is certain is that most people’s jobs changed at least to a certain degree. Being a programmer, I felt a significant influence on my career as a result of this state-of-the-art technology. However, I incorporated it into my workflow, and it helps me accomplish certain tasks more efficiently and quickly.

Starting with some positive aspects of this new type of AI, ChatGPT has become a part of my daily work routine. It provides suggestions on how I could improve my code, offers code examples so that I don’t have to look them up, and sometimes even presents useful solutions to my problems. However, it is not perfect—far from it, to be fair. It often produces incomplete and sometimes even incorrect code. That is why we still need programmers, and it’s why AI will not replace us just yet. Additionally, I’ve recently been using ChatGPT to assist with writing my blog posts. While I could have written entire posts using it, I see no point in doing so—it is my post, after all. What would be the point of writing in the first place if it was all generated by AI? I write because I enjoy it, not because I am paid to do so. Nevertheless, I have been using it as an assistant to offer opinions, correct my grammar here and there, and provide additional ideas for what to write about. It is honestly great for this purpose. I’ve also tried using ChatGPT 4 recently, which feels like quite an upgrade when compared to GPT 3.5. I can think of many other useful scenarios where ChatGPT might be really helpful, such as writing email responses (you obviously still have to check if they make sense), writing reports, or any other template-like written content that is mostly seen as a chore. Conversely, I am strongly against its use for writing entire posts, articles, etc.

Thus, many view this language model as a stepping stone toward AGI (Artificial General Intelligence)—the holy grail of AI and the ultimate goal for many AI researchers. AGI has the potential to exponentially improve the rate of technological advancement. However, achieving the so-called technological singularity could spell the end for humanity as we know it. The technological singularity is a hypothetical future point at which technological growth becomes unstoppable and uncontrollable. With the advent of modern AI and the internet, I sometimes wonder if we have already crossed the point of no return?

And now to address the elephant in the room, which prompted me to write this post in the first place: Sora. For those not familiar with it yet, Sora is another generative AI model. Unlike its predecessors, such as the language model ChatGPT or the generative image models like DALL-E and Midjourney, Sora specializes in generating videos. And it’s exceptionally good at that! It has truly surpassed my expectations for AI. In less than a year this technology has evolved from funny Will Smith eating spaghetti videos to hyper-realistic, often indistinguishable videos of anything that you desire. While I haven’t been able to test the model yet since it is still not publicly available, the provided video samples generated by Sora are simply astounding. It can conjure up almost anything: fictional characters, drone shots or a video of a commuter on a train.

This is probably one of the best video examples generated by Sora. If I didn’t know this video was generated by AI, I probably would never have noticed. I looked for any giveaways but I just couldn’t really find any obvious clues. However, after carefully replaying the video a couple of times, I did notice something weird, something that shouldn’t be physically possible. Around the 10-second mark, the left dog’s paw enters the frame and moves towards the middle dog’s head. Suddenly, the left dog’s paw morphs into the middle dog’s paw, and the left dog seemingly acquires a new leg out of nowhere. This highlights one of the problems that Sora faces: not actually understanding the world. It is a flaw shared by other similar generative models such as ChatGPT and DALL-E. They too do not really understand what they are drawing or discussing – they just predict what should follow based on the given input prompt. This limitation is reminiscent of the issues I’ve encountered with ChatGPT, where the code it produces often fails to function correctly, likely stemming from a similar lack of understanding.

Our jobs, including those of artists and other professions influenced by this technology, are, in my view, not immediately at risk of becoming obsolete. However, I think we are at a point where we should tread carefully, or this new technology might just come back at us at a certain point in the future. This truly is a double-edged sword. On one hand, tools like ChatGPT allow us to automate some of our work, such as writing emails, speeches, or code. These tools save me a couple of hours every so often. But on the other hand, there are a handful of problems that have emerged and are yet to be solved. Things like deep fakes, image (and now video) copyright problems, and the devaluing of artists’ works—the devaluing of human creativity.

First, there are deepfakes and misinformation. What an awful use of AI this is. Image and video generation models can be used to generate propaganda and misinformation videos on demand. Want to discredit the opposing political party? Why not create a fake video of them saying something controversial? And this is just scratching the surface! Deepfakes face pretty much the same issue. Most often they are used for shady, nefarious purposes, including pornography. Images of celebrities are used to generate inappropriate content involving these individuals without their consent or knowledge. The scary thing is, as long as the perpetrator has photos of a person, they can create a deepfake of that person in a matter of hours if not minutes. A recent example is the huge Taylor Swift controversy, which unfolded exactly as I’ve described. The infamous video was viewed by millions before its eventual removal. Such incidents are frequent and are not limited to deepfakes but also extend to generative image models like Diffusion, which are sometimes used for similar purposes. There are some solutions to these problems such as tools for AI image/video/deepfake detection. However, this often turns into a relentless cat-and-mouse game. Moreover, misinformation spreads rapidly, with significant damage often done before effective measures can be implemented. Stricter moderation and censorship present potential solutions but intriduce data privacy and other concerns. There simply isn’t a straightforward solution. Moreover, some individuals monetize such content, raising significant copyright law issues. This leads me to my next point.

Secondly, and significantly, is the problem of copyright. My primary concern lies with the somewhat unfair usage of artistic works in datasets used for teaching the models to draw or as of recently, make videos. This issue arises because art pieces are often used without the artists’ permission in the creation of these models. Consequently, artists are not compensated for it and often aren’t even given credit. Furthermore, people who make these models sometimes even profit from this instead of the original artists whose works this model is based on! The AI can now draw in their style and spit out hundreds of images every day. While these images may lack the quality of the original, their sheer volume might overshadow the genuine artistry. As long as the artwork looks appealing, most people sadly won’t pay attention to the details that usually give away the AI nature of the image. As far as I am concerned, this is borderline copyright infringement.

Addressing this issue, I propose a few potential solutions, though I am not an expert in copyright law, so these are just my thoughts. First of all, stricter rules regarding the usage of artworks in datasets and compensation for the works used in datasets. This however is not an easy problem to solve – the internet is free and censorship isn’t the answer to everything. Once the arts are posted on the internet everyone has access to them. Secondly the separation of AI arts from the normal arts. I have seen most image boards and websites already take this approach which is a step in the right direction. But I do think that more could be done. In my opinion, all posts containing AI-generated content should be clearly labeled. Especially on social media networks. This would benefit both the artists as well as play a part in solving the problem of misinformation, which I have partially touched upon in the previous paragraph. And lastly as mentioned before, tools for detection of AI-generated content. Though, once again, this is yet another cat-and-mouse situation. The users of AI technology will always find a workaround for these detection tools.

These examples highlight both the pros and cons of AI. Thus, as we journey through the evolving landscape of AI, from ChatGPT’s assistance to Sora’s cinematic marvels, it’s evident that these technologies stretch far beyond mere tools; they are mirrors reflecting our complex relationship with innovation. While AI brings efficiency and creativity to my programming and writing, it also brings challenges that test our ethical boundaries. The balance we must find is delicate, navigating between harnessing AI’s potential and protecting our core human values. The path forward should be shaped with optimism and a healthy amount of caution and scepticism. The mentioned challenges and problems need to be dealt with or these tools might just become weapons of the future. Like nuclear energy that can be used both for powering millions of homes as well as killing thousands of people, so too can AI be used for both good and bad and it is in our hands to shape our future the way we want.