AI is cognitive automation, not cognitive autonomy
Like the rest of computer science, AI is about making computers do more, not replacing humans.
The way we think about AI is shaped by works of science-fiction. In the big picture, fiction provides the conceptual building blocks we use to make sense of the long-term significance of “thinking machines” for our civilization and even our species. Zooming in, fiction provides the familiar narrative frame leveraged by the media coverage of new AI-powered product releases.
As a result, the dominant view in the popular imagination today is that AI is about creating artificial minds, agents with a will of their own. These agents, since they possess a similar kind of autonomy as their human creators, may decide to pursue their own goals, and eventually turn against humans. Anxiety about being replaced by machines is a key part of the media coverage about AI. A closely related theme, AI-caused mass unemployment, is often brought up. Back in 2015 or so many saw this as one of our most pressing threats. (The US unemployment rate is now lower than it was in 2015, while AI saw an unprecedented boom.)
It’s the classic Frankenstein monster archetype. But these are really just stories, cultural archetypes with no connection with the actual technological and social reality of AI.
Here’s an alternative frame for you to make sense of AI.
Three kinds of AI
There are three kinds of AI that we could build. Three A’s: Cognitive Automation, Cognitive Assistance, and Cognitive Autonomy.
Cognitive Automation
The first kind is cognitive automation: encoding human abstractions in a piece of software, then using that software to automate tasks normally performed by humans. Nearly all of current AI fall into this category.
Cognitive automation can happen via explicitly hard-coding human-generated rules (so-called symbolic AI or GOFAI), or via collecting a dense sampling of labeled inputs and fitting a curve to it (such as a deep learning model). This curve then functions as a sort of interpolative database — while it doesn’t store the exact data points used to fit it, you can query it to retrieve interpolated points, much like you can query a model like StableDiffusion to retrieve arbitrary images generated by combining existing images.
This second form of automation is especially powerful, since encoding implicit abstractions only via labeled training examples is far more practical and versatile than explicitly programming abstractions by hand, for all kinds of historically difficult problems.
Cognitive Assistance
The second kind of AI is cognitive assistance: using AI to help us make sense of the world and make better decisions. AI to help us perceive, think, understand, and do more. AI that you could use like an extension of your own mind. Today, some applications of machine learning fall into this category, but they’re few and far between. Yet, I believe this is where the true potential of AI lies.
Do note that cognitive assistance is not a different kind of technology, per se, separate from deep learning or GOFAI. It’s a different kind of application of the same technologies. For instance, if you take a model like StableDiffusion and integrate it into a visual design product to support and expand human workflows, you’re turning cognitive automation into cognitive assistance.
Cognitive Autonomy
The last kind is cognitive autonomy: creating artificial minds that could thrive independently of us, that would exist for their own sake. The old dream of the field of AI. Autonomous agents, that could set their own goals in an open-ended way. That could adapt to new situations and circumstances — even ones unforeseen by their creators. That might even feel emotions or experience consciousness.
Today and for the foreseeable future, this is stuff of science-fiction.
Modern AI is cognitive automation
So what is modern AI really? As it stands today, our field isn’t quite “artificial intelligence” — the “intelligence” label is a category error. It’s “cognitive automation”, which is to say, the encoding and operationalization of human skills and concepts.
AI is about solving problems where you’re able to define what needs to be done very narrowly or you’re able to provide lots of precise examples of what needs to be done. It’s not about creating artificial minds. It’s about making computers do more. It is along the usual, old spectrum of programming, etc. It isn't something fundamentally new.
In particular, it isn’t a magic wand that you can wave to become able to solve problems far beyond what you engineered or to produce infinite returns. Take self-driving, for instance. We’ve invested about $100B in the field over the past 10 years — roughly half of the inflation-adjusted cost of the Apollo program. And we’re now just starting to see fully driverless cars able to handle a controlled subset of all possible driving situations. You can ride in one in SF from Cruise (in private-access beta) or in SF or Phoenix from Waymo (in public access). Other actors tend to be far behind on the generality curve. Crucially, these results were not achieved via some kind of “just add more data and scale up the deep learning model” near-free lunch. It’s the result of years of engineering that went into crafting systems that encompass millions of lines of human-written code. And we’re still very, very far from generalizing to all locales.
The difference between cognitive automation and cognitive autonomy
Cognitive automation is incredibly useful. But intelligence is a different creature altogether.
Intelligence is to automation as a new lifeform is to an animated cartoon character. AI is a cognitive cartoon. An automaton. Much like you can create cartoons via drawing every frame by hand, or via CG and motion capture, you can create cognitive cartoons either by coding up every rule by hand, or via deep learning-driven abstraction capture from data.
“If the cartoon is drawn with sufficient realism and covers sufficiently many scenes, what’s the difference?”, you may ask. If a large language model can output a sufficiently human-sounding answer when asked a question, does it matter it is possesses cognitive autonomy?
The difference is adaptability to the unknown. A lifeform will autonomously adapt to a changing future. A cognitive cartoon will perform the scenes you planned for.
Intelligence is adaption to unknown unknowns across an unknown range of tasks and domains. Automation is, at best, robustly handling known unknowns over known tasks. Which is already incredibly difficult and resource-intensive in the real world — whether we’re talking about engineering resources or training data. (The resource-intensiveness, naturally, comes from the lack of adaptability: you need to plan for every possible unknown, whether explicitly or via a dense sampling of possible situations, assuming a fixed distribution.)
Today AI has no autonomy. It has no ability to adapt to situations it wasn’t designed for. To situations its creators didn’t themselves plan for — and didn’t even understand.
Lessons
So what should you expect of modern AI? What can we see if we extrapolate current trends far into the future?
You should expect broader applications and greater business value. You should expect AI to make its way into every industry, every product, every process. But do keep in mind that AI is not a free lunch — it’s not going to be a source of infinite wealth and power, as some people have been claiming. It can yield transformational change (like driverless cars) and dramatically disrupt countess domains (search, design, retail, biotech, etc.) but such change is the result of hard work, with outcomes proportionate to the underlying investment.
And you should not expect current AI technology to suddenly become autonomous, develop a will of its own, and take over the world. This is not where the current technological path is leading — if you extrapolate existing cognitive automation systems far into the future, they still look like cognitive automation. Better cognitive cartoons, but still not alive. Much like dramatically improving clock technology does not lead to a time travel device. Intelligence and automation are simply two different categories.
Like the rest of computer science, AI is about making computers do more, not replacing humans.