Why Every "AI" Conversation Feels Like Nonsense
Another “AI” Post?
I really didn’t want to write about artificial intelligence (AI) again so soon, and I promise this will be the last one for a little while, but I came away from my last two posts with a very unsettled feeling. It wasn’t the subject matter itself that bothered me, but the way I ended up writing about it.
In writing those posts, I made an intentional decision to be consistent with the vernacular use of the term “AI” when discussing a specific set of popular tools for generating text. Surely, you’ve heard of tools like ChatGPT, Claude, and Gemini. I did this because using more precise language is (1) awkward, (2) inconsistent with the terminology used in the podcast that prompted my posts, and (3) so unfamiliar to a layperson that accurate language comes off as pedantic and unapproachable. In the interest of making the posts simpler and more focused on the issues raised by the way these tools are being used, I ended up just accepting the common language used to talk about the tools.
But the truth is that I know better than to do that, and it bothered me. ChatGPT ≠ AI, even if it’s very common to talk and write about it that way. As time passed, I felt worse and worse about the possibility that by accepting that language, I gave the impression that I accept the conceptual framing that these tools are AI. I do not. In this post, I intend to make an addendum/correction to clarify my position and add some context.
What’s Wrong with the Term “Artificial Intelligence”?
People have been confused about what “artificial intelligence” means for as long as they’ve used the term. AI, like all of computer science, is very young. Most sources seem to point to the 1955 proposal for the Dartmouth Workshop as its origin. Right from the start, it seems to have been chosen in part to encompass a broad, uncoordinated field of academic inquiry.
By the time I was a computer science undergraduate (and, later, graduate student responsible for TAing undergrads) in the late 90s, little had changed. The textbook I used, Artificial Intelligence: A Modern Approach (AIMA) by Russell and Norvig, didn’t even attempt to offer its own definition of the term in its introduction. Instead, it surveyed other introductory textbooks and classified their definitions into four groups:
- Systems that think like humans
- Systems that act like humans
- Systems that think rationally
- Systems that act rationally
The authors describe the advantages and drawbacks of each definition and point out later chapters that relate to them, but they never return to the definitional question itself. I don’t think this is an accident. If you take AIMA as an accurate but incomplete outline, AI consists of a vast range of technologies including search algorithms, logic and proof systems, planners, knowledge representation schemes, adaptive systems, language processors, image processors, robotic control systems, and more.
Machine Learning Takes Over
So, what happened to all this variety? The very short answer is that one approach really started to bear fruit in a way nothing else had before. Since the publication of the first edition of AIMA, researchers produced a host of important results in the field of machine learning (ML) that led to its successful application across many domains. These successes attracted more attention and interest, and over the course of a generation, ML became the center of gravity of the entire field of AI.
If you think back to the technologies that were attracting investment and interest about 10 years ago, many (perhaps most) were being driven by advances in ML. One example is image processing—also called computer vision (CV)—which rapidly progressed to make object and facial recognition systems widely available. As someone who worked on CV back in the bad old days of the 90s, I can tell you these used to be Hard Problems. Another familiar example is the ominous “algorithm” that drives social media feeds, which largely refers to recommendation systems based on learning models. Netflix’s movie suggestions, Amazon’s product recommendations, and even your smartphone’s autocorrect all rely on ML techniques that predate ChatGPT.
Herein lies the first problem with the way people talk about AI today. Though much of the diversity in the field has withered away, AI originally encompassed an expansive collection of disciplines and techniques outside of machine learning. Today, I imagine most technologists don’t even consider things like informed search or propositional logic models to be AI any more.
The Emergence of Large Language Models
In the midst of this ML boom, something unexpected happened. One branch of the AI family tree advanced unexpectedly rapidly when ML techniques were applied: natural language processing (NLP). Going back to my copy of AIMA, the chapter on NLP describes an incredibly tedious formalization of the structure and interpretation of human language. But bringing ML tools to bear on this domain obviated the need for formal analysis of language almost entirely. In fact, one of the breakthrough approaches is literally called the “bag-of-words model”.
What’s more, ML-based language systems demonstrated emergent behavior, meaning they do things that aren’t clearly explained by the behavior of the components from which they are built. Even though early large learning networks trained on language data contained no explicit reasoning functionality, they seemed to exhibit that behavior. This is the dawn of the large language model (LLM), the basis of all of the major AI products in the chatbot space. In fact, this technology is the core of all of the most talked-about products under the colloquial AI umbrella today.
Here’s the second problem: people often use the term “AI” when they really this specific set of products and technologies excluding everything else happening in the field of ML. When someone says “AI is revolutionizing healthcare,” they might be referring to diagnostic imaging systems, drug discovery algorithms, or robotic surgery assistance, or they could be talking about a system that processes insurance claim letters or transcribes and catalogs a provider’s notes. The uncertainty makes evaluating these claims almost meaningless.
The Generativity Divide
There’s another important term to consider: “generative AI.” It describes tools that produce content, like LLM chatbots and image generation tools like Midjourney, as opposed to other ML technologies, like image processors, recommendation engines, and robot control systems. Often, replacing the overbroad “AI” in casual use with “generative AI” captures the right distinction.
And that’s an important distinction to draw! One unfortunate result of current “AI” discourse is that the failings of generative tools, such as their tendency to bullshit, become associated with non-generative ML technologies. Analyzing mammograms to diagnose breast cancer earlier is an extraordinarily promising ML application. Helping meteorologists create better forecasts is another. But they get unfairly tainted by association with chatbots when we lump them all together under the “AI” umbrella.
Consider another example: ML-powered traffic optimization that adjusts signal timing based on real-time conditions to reduce congestion. Such systems don’t generate content and don’t lie to their users. But when the public hears “the city is using AI to manage traffic,” they naturally imagine the same unreliable systems that invent bogus sources to cite, despite the vast underlying differences in the technologies involved.
That said, we can’t simply call generative AI risky and other AI safe. “Generative AI” is a classification based on a technology’s use, not its structure. And while most critiques of AI, such as its impact on education, are specific to its use as a content generator, others are not. All learning models, generative and otherwise, require energy and data to train, and there are valid concerns about where that data comes from and whether it contains (and thus perpetuates) undesirable bias.
The Business Case for Vague Language
Why does this all have to be so confusing? The short answer is that the companies developing LLMs and other generative tools are intentionally using imprecise language. It would be easy to blame this on investors, marketing departments, or clueless journalists, but that ignores the ways technical leadership—people who should know better—have introduced and perpetuated this sloppy way of talking about these products.
One possible reason for this relates to another term floating around: artificial general intelligence (AGI). This is also a poorly-defined concept, but researchers who favor building it generally mean systems with some level of consciousness, if not independent agency. For better or worse, many of the people involved in the current AI boom don’t only want to create AGI, they believe they are already doing so. Putting aside questions of both feasibility and desirability, this may explain some of the laxity in the language used. AGI proponents may be intentionally using ambiguous, overgeneralized terminology because they don’t want to get bogged down in the specifics of the way the technology works now. If you keep your audience confused about the difference between what’s currently accurate and what’s speculative, they are more likely to swallow predictions about the future without objection.
But I think that’s only part of what’s happening. Another motivation may be to clear the way for future pivots to newer, more promising approaches. Nobody really understands what allows LLMs to exhibit the emergent behaviors we observe, and nobody knows how much longer we’ll continue to see useful emergent behavior from them. By maintaining a broad, hand-wavey association with the vague concept of “AI” rather than more specific technologies like LLMs, it’s easier for these companies to jump on other, unrelated technologies if the next breakthrough occurs elsewhere.
Speaking Clearly in a Messy World
That makes it all the more important for those of us who do not stand to gain from this confusion to resist it. Writing clearly about these topics is challenging. It’s like walking a tightrope with inaccuracy on one side and verbosity on the other. But acquiescing to simplistic and vague language serves the interests of the AI promoters, not the community of users (much less the larger world!).
From now on, I’m committing to being more intentional about my language choices when discussing these technologies. When I mean large language models, I’ll say LLMs. When I’m writing about generative tools specifically, I’ll use “generative AI.” When I’m talking about machine learning more generally, I’ll be explicit about that, too. It might make my writing a bit more cumbersome, but this is a case where I think precise language and clear thinking go hand in hand. And anyone thinking about this field needs to be very clear about the real capabilities, limitations, and implications of these tools.
The stakes are too high for sloppy language. How we talk about these technologies shapes how we think about them, how we regulate them, and how we integrate them into our lives and work. And those are all things that we have to get right.
