How smart does AI have to be to help marketers?
How smart does AI have to be to help marketers?
It would seem to some that AI marketing is a natural marriage for a marketer looking to deliver value by way of insights. However is that still a dream? Way back in 1965, Herbert Simon at MIT predicted that machines would be able to do anything a human can do in twenty years’ time. Fifty years later we’re still not there, but today’s AI hype cycle would suggest otherwise.
That’s why I thought I’d clean up some of the misunderstandings in this blog post. Here’s my best effort to describe what is currently possible, what is currently not possible, and why you as a decision maker in marketing technology should care.
It’s easy to see there are massively inflated expectations toward some of the highly publicized AI systems, which are fuelled by remarkable demonstrations of machines outperforming humans at games like jeopardy and go.
These demonstrations pair state of the art machine-learned models with handcrafted programs to manage a huge volume of domain insights, such as the complete rule book for Go and every possible move you can make.
The results are impressive, to say the least, showcasing breathtaking performance in an admittedly complex problem domain. But the reality is these demonstrators only simulate one generic form of AI. They may well beat the best trained human brains in the ancient game of Go, but that machine wouldn’t stand a chance answering questions a 3-year old could answer, like “is ice warm?” Why is that?
Today’s AI systems can be categorized as either pattern recognition or algorithmic intelligence. To match the intellectual capabilities of a 3-year-old child, the system needs to combine both.
Pattern recognition intelligence
World-renowned thinkers in the domain of economic principles long held the view that computers are only good at following specific rules. For example, they presumed that in the domain of credit risk rating, a human programmer could define a set of IF, THEN, ELSE rules to describe whether someone should be granted financial credit, so a computer could then ingest inputs to evaluate rule compliance at amazing speeds, without error.
So far so good. But the same experts also assumed the computer would never be able to design the rules and patterns that define the problem-solving decision process.
Today, we all know this was a wrong assumption. Computers are mastering the art of pattern recognition because human data scientists have found mathematical ways to recognize patterns, paired with the inherent advantages computer systems have in general (such as signal transmission at the speed of light or the capability to tirelessly repeat a process ad infinitum without theoretically ever making an error).
For example, computers are already at human quality levels when it comes to solving geometric problems like finding a pattern that describes a low-risk credit; or the meaning of the word ‘benefit’ in the hair care business of an FMCG company
No surprises here, computers are way better than humans at executing algorithms. AI systems show superior performance solving abstract problems that combine pattern recognition intelligence with, for instance, graph or search-based methods.
Back to parlour games, where a great example would be the part of the AlphaGo system that applies pattern-recognition based evaluation of the positions of stones on the board to the decision about which move to make next. No human is capable of executing the algorithm required to make this decision with the speed and continuous quality a machine can achieve here.
The big difference between the “intelligent” systems we see today, which only appear to display superhuman cognitive capabilities, and actual, super-human intelligence capabilities (compared to our three-year-old), is not only the ability to combine pattern recognition and algorithmic intelligence, but to integrate far more complex architectures that solve the two problem domains at much more detailed levels.
What does that mean?
To solve a problem like Go, programmers have to create the modules that recognize the pattern of a promising board position, and also write the algorithm that defines what to do with these evaluations.
A human knows how to do both. Given a problem, we develop strategies to solve it, while simultaneously training our mental patterns to recognize and auto-evaluate the board position we see, and also cognitively dealing with all the latent problems of existing in a physical world, like walking, talking, eating and breathing.
The current state of the art in AI is far from being so universal
I am not saying that these combinations are completely unexplored. Yes, technical exploration of these combinations is underway. Methods also exist to automatically search for solutions to algorithmic problems, even to learn how to construct pattern recognition architectures. But the combination of the two is far removed from what is needed to approximate human universality.
So there’s nothing wrong with being impressed by the fascinating new capabilities of deep learning architectures, but it’s dangerous to assume that general artificial intelligence has been achieved yet, or to call more or less every AI service “cognitive.”
Over-promising what to expect from AI systems, poses a serious change management problem as, and I assume that is really where everyone agrees, nothing is more expensive than winning back a user who has been disappointed by the real capabilities of an AI.
That’s why it’s important for AI marketing, marketers should look to invest in AI technologies AS WELL AS verticalization in the domain of their core business model, which combines pattern recognition intelligence with algorithmic intelligence.
At Market Logic, we help clients to solve these problems by integrating supervised pattern recognition capabilities (for the specific vertical) with supervised algorithmic capabilities, which is the minimum a buyer should invest into with a marketing insights platform that will leverage the power of AI marketing.
Both are then combined with an AI marketing insight for the problem domain, from the market logic knowledge graph. Of course, our system is not yet able to solve that domain all by itself, no system can do so currently – but we can make sure it solves already known problems with supervised machine learning.
This is where our data science team crafts models and architectures that generate ROI from day one while working on methods of higher abstracted problem-solving capabilities in the medium term.