AI is a bad word

It’s actually an initialism, and two words, and it’s not a swear word, at least not yet. But multivalent puns aside, I’m feeling 2012 vibes again.

For those who may have missed it, in 2012 Siri was the darling new technology from Apple and everyone was chasing this emergent intelligent service, AKA the personal voice assistant. Many called Siri an “AI,” and in many ways it was, and still is. But in the AI research community, it’s a specific type of Natural Language Processing (NLP) that uses a technique called “slot filling” — a predefined ontology of applied functions, and far from unstructured idea comprehension.

At that time I was working at Microsoft on what would become our answer to Siri: Cortana. Our team, under Blaise Agüera y Arcas, bridged the design, engineering, and research teams. Our focus was “applied intelligence” — the natural progression of ability we assumed was ahead. Working with Microsoft Research (MSR) was eye-opening, in a delightful and humbling way. Some of the smartest and most thoughtful people I’ve worked with.

Back then, Artificial Intelligence, AKA AI, was a bad word within Microsoft’s walls. When it was used at all, it was used carefully, or derogatorily. This was because of two factors. The first was that Microsoft was still nursing a hangover from its own software agent era — Bob in ‘95, Clippy in ‘97, the Office Assistant switched off by default in Office XP. The second was that all the serious computer scientists at the time were focused on machine learning, and ML was about math and statistics, not philosophy. Anyone working close to products like Siri and Cortana could clearly see the gap between their work and the science fiction portrayal of AI.

For my entire tenure at Microsoft, we used “machine intelligence,” “intelligent services,” and “intelligent/autonomous agents” to describe the potential of our aspirations. We called the connective tissue of that potential the “intelligence substrate” — the shared layer of models, signals, and services that every product would eventually draw from. The one mention of AI I can find in all my work is a note from my boss Blaise questioning some of Google’s claims in 2013, alongside a quote from Ray Kurzweil. Years of notes, one hit. And yet our expectation for where intelligent services were headed, and the role of intelligent autonomous agents in just about everything, was no different from what we see now — we just didn’t call it AI.

So now, with data centers under attack, speakers getting booed off commencement stages for praising AI, a generation (or two, or three) feeling AI is robbing them of potential, I’m beginning to wonder if it might be time to return to the language we used at Microsoft. To be clear, I’m not trying to sweep the problems under the rug — I’m deeply invested in addressing them in any way I can. But the thing we were describing back then did arrive. It just arrived under a different name, and with worse manners. And I still see part of this emergent technology as inevitable. It’s what I’ve been sharing for a while: the creation of a new abstraction layer that sits on top of the old.

The part of AI that is so destructive and generating so much ire at the moment is rooted in the disregard of human rights and dignity. It likely started in the bowels of industry — the desire to capitalize on innovation. But the first critical mistake was allowing OpenAI and Google to extract the world’s knowledge without any contract or mandate to fairly compensate the authors. The lack of consequence created a contagion that has poisoned the systems of human governance globally — a selfish one. Harsh words, but what’s emerged is a general narrative of AI as human replacement, and this largely stems from the choice to break ethical and moral societal contracts in the name of technological and individual dominance.

So where do we go from here? Part of the solution may be a natural defense mechanism of humans fighting back. The focus on framing the value as human replacement will come back to bite the companies — because companies need consumers, and people who are not working can’t consume. Then there’s market equilibrium — when AI models become a commodity, the value becomes a direct reflection of the cost (both resource and societal), which will make many of the current AI applications, especially the unethical ones, unsustainable.

Those are forces that act on the industry from the outside, on their own schedule. The lever I actually control is what I call the thing. So going forward I’ll revert to my Microsoft lexicon and drop the “artificial.”

The word comes from the Latin artificialis, by way of artificium, craft, and artifex, craftsman. It meant “belonging to art” — made by skill rather than by nature. From the Renaissance through the 1700s, artisans got very good at mimicking it: silk flowers, glass paste cut and set to pass for emeralds, scagliola plaster polished until it read as marble, and eventually porcelain teeth. The artificial label was a nod to the artistry and craftsmanship. But as the copies became harder to tell from the real thing, the meaning drifted toward deception and insincerity. As we drift into a world where AI is indistinguishable from reality, does the word take on new meaning? A swear word, maybe?

Don’t mind me, I’m over here building intelligent services on the intelligence substrate.