The language we use to describe AI matters

Here we go again. Another AI researcher emerged this week warning us of humanity's impending doom due to the power of AI. This one was a former Anthropic employee who quit his job because he is convinced these models are going to kill us all. I thought we were past the doom-speak, but apparently not.

Alongside these types of pronouncements, which we have been hearing on a regular basis since ChatGPT emerged in 2022, there is a charged language developing around describing this generation's AI capabilities. It seems deliberately designed to be scary and threatening to make the AI software sound more capable than it really is, and attribute actual agency, intelligence and even menace to the AI.

Consider that when describing recent agentic AI-led hacks of Hugging Face and a German wiki, the nomenclature used by OpenAI was "swarms of agents," wording which was picked up and used frequently in the press. Meanwhile, some went so far as to describe agents as actually creating "civilizations of agents." Words like this raise images of the zombie apocalypse movies and dystopian science fiction instead of what podcaster and Georgetown professor Cal Newport described as merely a way to deal with complex prompts. 

"So what's going on with an agent swarm? It's not about some sort of exotic type of intelligence. It's prompt management,” Newport said in a recent episode of his Deep Questions podcast. "So this is a swarm in the same sense that having multiple programs open on your Mac is a swarm of programs," he added. (I encourage you to listen to the whole episode for his complete explanation.) 

Even relatively benign language like hallucinations, thinking and reasoning assigns human-like capabilities to software that in my experience, while able to do some pretty amazing things, isn't close to the kind of human-like intelligence that we keep hearing attributed to current AI. 

That is not to say that recent agent hacking incidents, however you choose to describe them, didn't happen. There's plenty of evidence they did, even if most of it comes from the labs themselves. The issue is the language we use to describe them, and the motivation behind the words.

Real world versus alarmists and true believers

There is definitely a gap between what I call the reality of everyday use and AI true believers, who are ready to accept every researcher's alarmist call that this software is so good, it's capable of destroying civilization. In fact, the more time I spend working with these tools, the greater the disconnect I find between the language the industry is using to describe them and what they actually do.

I spend time with AI every day. I use Claude, ChatGPT, Perplexity and Writer (which gave me a free subscription). These tools have become a regular part of my research, editing and review workflow, and even business document creation.

I've been able to generate a number of documents for the business side of FastForward including proposals, invoices, contracts and marketing documents; all from a description, which would have taken hours before. But it doesn't just magically happen on the first try. It takes a lot of work, sometimes incredibly frustrating work, to get it done. These models backslide. They forget. They make dumb formatting mistakes and struggle to fix them. They are not all-knowing or anywhere close to human-like intelligence.

More recently, I've tried to teach a Claude agent to become my news gatherer for this newsletter. Each day I check the Techmeme headlines and choose some of them for news of the week, commentary I'm considering, what I'm reading and what I'm watching. Each morning I review the Techmeme email, make my choices and quiz the agent on what it chose. While it's improving incrementally, after a couple of weeks of back and forth and refinement upon refinement, it's still not really close. This isn't a simple task, but the agent lacks human flexibility. It looks for fixed, rigid rules where none exist and human judgment comes into play. Wouldn't a human-like intelligence pick this up much faster without so much input from an actual human?

I keep coming back to the idea that AI has some astonishing capabilities today. Why do we need to make it sound more powerful and actually panic people about it? The labs have their agendas, of course, to make the models sound bigger, badder and scarier. Industry insiders appear to be stuck in a feedback loop, reinforcing one another’s perception about the danger of these models. But for the people outside of this bubble, words matter because they can suggest levels of intelligence and agency that everyday use simply doesn't support.

~Ron