3 AI Myths That 2025 Will Debunk

The AI revolution has generated many false starts, not necessarily in the technology, but in how people react to it and their assumptions. Here are 3 assumptions widely spread in the last year or two, which I expect to be thoroughly debunked in 2025.

Prompt Engineering Is A Career

When ChatGPT and other Large Language Models (LLMs) first arrived, Prompt Engineering (the art of crafting an effective query to these applications) also arrived. Prompt Engineering was hailed as an emerging career (and certifications also emerged). Just as Google searching is not a career (even though Google’s search engine is a complex software which will return better results if you make more precise queries), prompt engineering is unlikely to be a career. Also worth noting that it is very much in the interests of the providers of these tools to make them easy to use, i.e. provide useful answers that address your intent regardless of how well you frame the question.

AI Detectors Work

Along with LLMs, AI Detectors emerged that claimed to be able to tell whether content was AI generated. This was structurally unlikely given how AIs work internally. AIs that generate realistic content often have internal means to assess the quality of content and continue iterating until it “looks real”. For example, a common AI technique, Generative Adversarial Networks (GANs) operate as two competing AIs, one generating and one discriminating. The iteration continues inside it till the generator comes up with something the discriminator cannot detect. If AIs use these methods to improve their content, how can we believe another AI can detect it? Methods like AI watermarking, and legislation to enforce such methods, are emerging and may become a more reliable way to detect and track AI content. Mistakes by AI detectors are starting to make the news. I expect 2025 is the year when we widely desist from reliance on such tools.

Using AI And Learning AI Are The Same Thing

The previous two are relatively safe bets, opinions to the same effect were already emerging in 2024. This one is harder. 2024 saw a massive growth in interest in AI Literacy, but for many, AI Literacy means knowing how to use AI tools. Why is this not enough? From my own experience, I can make the argument that I am an effective car user, but I could not explain how the engine works. Similarly, I know how a computer works, but I don’t rely on that knowledge day to day as I use my computer. What is different here? The difference is that my car and my computer are not gaining skills almost daily, and are not increasingly becoming able to do more and more of my job. If that were to happen, I would become very interested in how they work. I will need to know how they work to figure out how to add value in my job above and beyond them. I expect 2025 is the year when we start to realize that, for a large fraction of teh working population, AI Literacy also means understanding how AIs work.