A hesitant way forward: My thoughts on current AI scepticism

A hesitant way forward: My thoughts on current AI scepticism

“When falsehood can look so like the truth, who can assure themselves of certain happiness” — Mary Shelley, “Frankenstein.”

Recently, I held discussions with a few friends concerning their “AI opinions”, and I got a lot of/from them, ranging from total hatred to unrestrained acceptance. My personal opinions (discussed here) are much more nuanced than these extremes, which I attribute to my active involvement with AI systems. Through these discussions, I highlighted a fairly common motif with those who were anti-AI.

“People who hate AI systems do so because they haven’t seen any good come out of it.”

To be fair, the marketing around AI products has been misleading and rather uninspiring. A huge portion of this promotion focuses on large language models, AI image generation, and some sort of human replacement. In AI talks/seminars, it’s a recurring theme to describe the end goal of these systems as “replacing” humans. The word replace in this case might be interpreted as the end of something and its eventual discarding. This is significant in processes that require human emotions and uniquely human processes. The significance stems from the fact that AI systems simply aren’t human. Not because they can’t write or they can’t draw, but because they lack sentience (The quality of being conscious). AI systems try to replicate human feelings and experiences through complex pattern recognition, but sentience isn’t replication; it’s closer to personal experiencing.

In these discussions, one topic we particularly touched on was AI deepfakes and their eventual consequences. Due to the open-sourcing of image generation models, deepfakes have gone viral, typically for malicious purposes. With AI images and videos becoming indistinguishable from real media, fraud losses have been projected to double over the next few years. Slowly but surely, we’re finding ourselves in situations where the already corroding fabric of societal trust is being torn apart. Previously, video evidence was irrefutable proof of events but with AI videos, we’ll have to question the authenticity of every video we come across. It’s even more unfortunate that social media companies are allowing this and might be actively encouraging it.

Although the effects are less distressing, I extend similar distaste towards AI music. I haven’t seen any on my Spotify, but I dread that day {PASS ME BY}. Completely AI-generated music is dead for a couple of reasons. I use the word “completely” intentionally; I haven’t been able to concretely decide what level of AI involvement is unacceptable (I believe using AI to generate lyrics is nonsense, because the struggle in the thought process is part of the art itself), but I’ll focus on completely AI-generated music here. Being Nigerian, I interact a lot with Nigerian music. Creating this sort of music requires interaction with the Nigerian community (local or diaspora) and her ever-changing trends. You can’t simulate this. Understanding the language isn’t understanding the culture. Understanding the musical patterns isn’t understanding the culture. But understanding and partaking in the experience is understanding the essence. Outside cultural involvement, good music requires a decryption of whatever sentiments have been gathered. This is a product of experience and keen observation. To illustrate this, I’ll use music from my top 3 artists on Spotify:

I. Tennis

II. Cocteau Twins; and

III. The Sundays

Tennis is a band comprised of a couple who write music about the experiences within their relationship, growth etc. The intentionality in writing, and the growth in thought from the first album to the last album is something I think is irreplicable. It’s personal. It stretches over time. It’s a career of storytelling. No AI system is capable of this due to a lack of emotion and original experience. The limitations of emotion and original experience would always hold back every non-human system in creative endeavours. The second band on the list is the Cocteau Twins, a group that embodies the zenith of “dreampop” excellence. Majority of their songs are sounds that seem like some other language (thus lyricless), yet you can understand all the feeling Elizabeth Fraser (the lead singer) is trying to get across. This stamping feeling of the Cocteau Twins is also irreplicable because it just represents the genius in human interpretation of what music is. One of my favourite albums of all time is Blind, an album by the Sundays. It’s a beautiful album that questions and touches on philosophical matters. Questions about the fundamental nature of life and the inevitability of time. These questions require sentience, life, and the understanding of an eventual end; therefore, out of the reach of any AI system. To conclude this chapter, AI assistance might be useful, but the idea of completely replacing human-created music/art with AI-generated media is dead. Soulless.

No matter how soulless the current landscape may seem. We mustn’t despair; we mustn’t. In fact, the reason I’m writing this now is to highlight the good of AI. Most people reading this haven’t heard or seen any good use of generative-AI. It isn’t your fault. Such undertakings usually aren’t marketing focal points because they don’t yield immediate profit, they’re likely technical, and they’re not clickbaity (so your favourite news platform doesn’t see the need to promote it). During my co-op (internship), I spent my free time working on human genes, and gen-AI (generative-AI) was of help to me. There’s a protein called tp53 — responsible for preventing gene mutation — and with the help of gen-AI tools, I was able to analyse it, understand the fundamentals of human genes and the sites that are susceptible to genetic mutation. That’s a personal good!!! I wouldn’t make rigid predictions on the future of AI, but I believe science research would accelerate with the “proper” use of AI tools. There’s been loads of computational biology work using gen-AI. Neural networks (the backbone of most AI systems) are probabilistic models that are excellent at optimising search spaces (in simpler words, they use probability to predict complex patterns), so this makes them “perfect” for designing proteins and drugs. In the same sphere, there are language models called protein language models that treat protein sequences like a language and infer syntax from these sequences. The same principles apply to material science where research groups have utilised gen-AI models to create materials that fit a particular criterion.

An additional example worth noting is Google DeepMind’s AlphaEvolve, a coding agent that was able to probe mathematical problems and rediscover state-of-the-art solutions; it also improved existing solutions about 20% of the time. Remember what I said earlier about neural networks being good at optimising search spaces; we can extrapolate this to manufacturing/logistics processes. Given their adaptability to large datasets, gen-ai tools have also been used for resource allocation when provided with real-time data. Real-time analysis powers weather forecasts and physics simulations. One use of gen-AI that I’ve found interesting is its use in decoding lost languages and recreating ancient phonetics. This is inherently a feature of neural networks being able to retrieve information from seemingly obscure data.

Altogether, we are quickly heading into a future where artificial intelligence systems will spark discussions and debate. It is necessary to ask questions about these systems, their decision-making process, and their reliability. Even to me who spends time studying and researching AI systems, they can feel like a blackbox (which is why I perfectly understand all the scepticism). We should note that Artificial Intelligence systems aren’t absolute. They hallucinate and quickly become echo chambers. Even experts aren’t immune to the unnatural ultra-positive feedback from large language models. I think the best way to refute scepticism is to show the good of AI systems, and I also believe that despite these challenges, there’s a lot of good that we can create from these systems. DARING DEFENSE FOR BETTER TIMES.

Notes:

1. I really wanted to touch on the environmental impact of AI but the environmental figures in reports/papers change a lot. Luccioni, Jernite, and Strubell (2024) reported figures of 0.47 Wh, 0.49 Wh, and 2.90 Wh for text generation, text summarisation, and image generation, respectively (Link to paper provided below). In comparison, an 8 second microwave run is about 1.9 Wh but when you consider the amount of prompts these AI systems receive per day, the energy consumption rate becomes very significant. I believe the energy usage per prompt will decrease significantly over the next few years. It’s still very important to create awareness around this issue though. Link to Paper: Power Hungry Processing: Watts Driving the Cost of AI Deployment? (https://arxiv.org/pdf/2311.16863).

2. Thanks to Inemesit Usoro-Udo, Matthew Ohlmann and Kaleb Conquergood for going through the drafts of this essay and providing a lot of constructive feedback.

Thanks a lot.