Are We Nearing an AI Winter? Evaluating the AI Landscape Through Promises
An AI Winter, to quote Wikipedia, “is a period of reduced funding and interest in artificial intelligence research.” It is the opposite of an AI Spring, which is, as you might guess, a period of increased funding and interest.
Needless to say, we’ve been in an AI Spring since the early 2010s. A few landmark moments include:
- The launch of Amazon Alexa in 2014, which brought voice-activated assistants into the mainstream. (Let’s be honest, in 2014 Siri was more useful as a drinking game than an actual assistant.)
- Google’s AlphaGo defeating the world’s #1 ranked Go player, Ke Jie, in 2017.
- AI’s near-human cancer-detection success rate when analyzing radiographic images. There are a number of companies/academics in this space, IBM Watson reported successes in 2013.
- Every time you’re persuaded to click an ad online. Or vote for someone you wouldn’t usually vote for. *cough*
Doubtless an AI Spring. (Though, as with any technology, dubious applications may raise the question: “Spring for who else?”)
With that context, the notion of an AI winter might seem ridiculous. Everywhere we look are headlines extolling the advances of the AI Revolution of the 2010s.
Here are a few commonly-cited examples:
On the B2C side of things, we have Alexa’s natural language processing, Google Maps’ predictive route planning, and cancer detection, which is provided by a number of vendors.
On the B2B side of things, we have JPMC’s’ automated legal reviews, approximately 10,000 chatbot vendors spread across a number of use cases, and AutomatedInsights’ Wordsmith, which automates sports and financial reporting.
An AI Winter, really?
Let’s start here: an AI Winter does not entail the unravelling of all previous AI-related innovation. It means, to quote Luciano Floridi in his Logic of Information, reaching the top of the tree, rather than taking the first step towards the moon. For example, after the collapse of AI hype in the 1970s, DARPA still leveraged previous AI investment — DART — to save billions in logistical costs during the first Gulf War.
So what’s the reasoning?
Evaluating AI Through the Lens of Promises
There are a number of lenses you might look through to evaluate where we sit in relation to previous cycles of AI hype and AI disillusionment.
You might look at the underlying technology, which has not changed much in form, only in application and quantity of data, over the past several years. [Update: the same day this was published, Wired published an article featuring Facebook’s Head of AI, who claims the field will soon “hit a wall.”]
You might look at the struggles of contemporary data scientists to innovate beyond the frontiers of their teachers.
You might look at increasingly negative press coverage, emerging ethical dilemmas, emerging legislation, or Gartner’s Hype Cycle. There are others who are better equipped to look through these lenses, and solid starting points for each are linked here.
The lens I’d like to look through is that of promises. At the beginning of any new innovation, whether it be a long-term trend or a flash-in-the-pan fad, promises are made. The difference between long-term trends and fads is made clear over time by the technology’s ability to keep those promises.
I think that we can all agree that, for a large part, the promises made by AI proponents at the beginning of the Spring have largely been fulfilled (with a few notable exceptions). The level of prediction we’re getting is groundbreaking: from the kitchen, to the doctor’s office, to the legal department. In many cases, it’s even “transformational.” It’s clear that artificial intelligence is not a flash-in-the-pan fad.
But in any healthy innovation, promises are not only made at the beginning of the upswing. They continue to be made. As promises are fulfilled, they are replaced with new promises: new frontiers for technological exploration. In this regard, we seem to have hit a roadblock.
The examples of groundbreaking AI that journalists and vendors trot out haven’t really changed since 2015. With the exception of AlphaGo’s win over Ke Jie, the examples I listed above are all from 2015 or earlier. Where are the radical AI innovations of 2019? Where are the new frontiers? The most dramatic applications of AI seem to involve taking tech from 2015 and porting it to an overlooked or headline-provoking use case.
There are unkempt promises too.
Anything but the most basic robot seems severely limited. Chatbots are still pretty darn dumb. Amazon’s Alexa hasn’t seemed to improve much since it was launched, despite immense quantities of newly-generated data to improve with. And self-driving cars seem to have more in common with Waiting for Godot than any available technology.
And remember the hype when IBM Watson could identify cancerous tumors with near-expert-human accuracy in 2013? We might imagine that five years would make even the most skilled radiologist obsolete. Instead, all we’ve learned is that the technology is at best overhyped, and at worst actually harmful to patients.
Looking into 2020, what new promises do you see on the horizon? We’ve got talking assistants, talking refrigerators, and better game-playing AI than ever. But these innovations are all nearly five years old. This isn’t a matter of innovation; it’s a matter of consumer adoption, and transposing existing technology onto new use cases.
The lens of promises is far from the only way to look at this question. But combined with other lenses, I think we can create a more comprehensive, if less optimistic, picture. Will the shift happen immediately? Of course not. It takes time for the frustrations of AI academics to coincide with the frustrations of AI consumers. Perhaps it would be more accurate to describe the current landscape as an AI Autumn, with all the pros and cons that entails.
Maybe an AI Winter Isn’t So Bad?
Again, this is not to say that today’s existing innovations will fade away. “AI” carries with it a sense of mystique, a sense of something new. As the AI innovations of today become “just the way things are,” we might even argue that AI needs an AI Winter to rebrand, and reemerge 5, 10, or 15 years later with all the hype and gusto of 2016.
If we look at the history of innovation (and even modern theories of evolution), progress is not a linear. It is a series of leaps.
