There’s one main thing to understand upfront about artificial intelligence (AI). Take 99% of everything you’ve heard the media say about AI. Now find an actual expert on the subject and show them the latest AI story headline from the likes of Wired, The Guardian, or even The New York Times. That expert will literally turn into Patrick Stewart on the spot:
That image sums up the public understanding of AI. It is a hugely misreported field because AI unites two deeply technical branches of science: computer logic and human brain neuroscience. You need a master’s degree and a lifetime of study to become an expert in just one of those fields. So finding an expert in AI means finding somebody with a deep grasp of both of those fields.
First We Must Understand AI…
Before we dive into how AI works in marketing, here are a few buzzwords we should demystify upfront:
- Automation: This just means “having a machine do the work,” not necessarily with AI involved.
- Machine assistance: Another term for automation, where we still have a human in control of decisions, but computers do the data processing.
- Machine learning: A very new, pioneering branch of AI we will discuss more below. The equivalent of turning a toddler loose in a playroom and letting them learn on their own.
- “Black Box” AI: A term recently in vogue, which loosely means “quick and dirty result-oriented machine learning without caring how it’s done.”
- Simulation: “Faking it.” You can program a system to solve a problem through brute force, which makes it look like it “understands” the problem. Chess-playing software works this way.
- Algorithm: A set rule with concrete conditions producing a predictable result. “X + Y” always returns the exact sum of X and Y.
- Heuristic: A generalized method of producing a close estimate to an answer. “355 / 113” returns a fast, close approximation of the number Pi good enough for small construction projects.
- Cognitive computing: As opposed to a simulation, this is the attempt to make a computer use human-like reasoning to accomplish tasks.
In a nutshell, artificial intelligence is the attempt to have computers simulate cognitive tasks normally associated with human intelligence. This is a challenge and an open problem, with sub-problems where we don’t even know yet whether or not they can be solved. It turns out that humans are really good at some things that computers can’t do yet. So far, we settle for “close enough” with a mixed approach using any and all of the above methods.
Let’s look at some FAQ on the limits of AI:
- Why can’t we just build a computer brain? We have attempted it with a neural network, but brains have more to them than just electronic impulses between synapses, such as the whole chemical neurotransmitter system.
- Will we ever have a complete computer simulation of a brain? It may be doubtful, considering that neuroscience is an ongoing field in itself with breakthrough discoveries happening every day.
- How come a machine can solve big math problems but can’t find my car keys? Computers are great at logic. Humans have other skills, like great spatial ability, super visual pattern recognition, attention control, creativity, social cognition, and so on, which computers have difficulty replicating.
- If computers can’t create, how come I’ve seen them do random things? Computers actually cannot make choices on their own. The random behavior you see in a screensaver or video game is synthesized from external entropy, such as the timing of events on the system clock measured in fractions of a millisecond, or timing of user feedback on the keys and mouse, combined with some fancy math to produce pseudorandomness.
- Will AIs ever “take over”? The chief hurdle to this scenario is that computers do not possess agency. Lacking their own desires, they will only ever do what we tell them to do.
We will of course not be able to plunge into the depths of AI science in a humble blog post. As much as our geeky research staff would love it! Instead, we will propose to dive into…
How Artificial Intelligence is Used in Digital Marketing
Simple AI begins with textual pattern matching, which even the simplest computer can do using basic algorithms. This is how Google indexes web pages to deliver in search results, how spell-checkers on your office suite work, and how software can grade a multiple-choice test. However, this is more a matter of automation rather than true AI.
Many industries use machine-assistance algorithms to facilitate small tasks. Some examples are:
- Netflix recommending movies based on what you have watched
- Steam recommending games based on what you’ve played
- Amazon recommending products based on what you’ve bought
- Twitter recommending users to follow based on your follow list
- Facebook displaying ads based on your posts and what you view
The list goes on! All of these actions rely on the concept of tagging. Every video you view, game you play, product you buy, and profile you visit either has tags in plain view, or an algorithm behind the scenes indexing the text to form a tag database. On Steam, if you played a first-person shooter game set in a zombie apocalypse, the tags “3D,” “shooter,” and “zombies” are part of your preferred game queue.
A great deal of targeted marketing, including PPC ads, works based on the same concept. Facebook serves ads based on various metrics including location, demographics, interests, topics you follow, lifetime events, and even actions you’ve taken on third-party sites if those other sites use your data and also market on Facebook. Programmatic display advertising also relies on AI to display ads to the desired target group as they move through the internet.
Raw tag matching only gets you so far. What do we do for the user who is tight-lipped about what they prefer? Or how do we implement more targeted marketing based on the limited facts we know about a user? We have a concept of a “similar user” or “lookalike audience,” and now we’re getting closer to true artificial intelligence here. This is the hot, trending field of…
We humans cringe at stereotyping. It introduces all kinds of biases and bigotry. But marketing algorithms love it! We don’t even tell the computer about stereotypes, we just give it a pool of users and say “find similarities between any of them and form your own stereotypes.” Suddenly, user A sees an ad for leather boots, only because he bought a Garth Brooks album like user B, who also bought leather boots.
Predictive analytics uses statistical methods to classify people into groups, and exploit patterns derived from those groups. This is done through machine learning, where we can give the computer a full set of data for cases A and B and tell it to compare every data point between the two cases. We can do this now because computers are so much faster than they used to be, thanks to Moore’s Law.
Machine learning is a much older concept that has only become possible in recent years due to improved processing speed. The term was first coined in the mid-20th century. Let’s use a really simple example:
In a Tic-Tac-Toe game, we can let the computer teach itself by having it play both sides and record the outcome of each match. After a few hundred games, it has enough data to start collecting patterns. It notices that in openings where “X” plays center and “O” plays an edge, “O” loses a higher percentage of games than when “O” plays a corner. Therefore, it learns that playing the edge on the second move is sub-optimal, disregarding anything else that happens in the game.
That’s easy to do for Tic-Tac-Toe, hard to impossible for bigger games. Tic-Tac-Toe is finished in nine moves with only 255K possible games. The average Chess game is 40 moves, with a total of possible games somewhere near 10^120. Chess has been played using brute-force algorithms looking a few moves ahead, while the Asian board game of Go (possible moves = even astronomical number) has been impossible for computers until modern computer speeds.
This is actually a huge step in computing! For years, games like Checkers and Chess defined the limit of computer game algorithms, while the Asian board game Go was beyond even the best computers. But DeepMind Technologies applied methods exactly like our Tic-Tac-Toe example in training the AlphaGo engine. In 2015, AlphaGo became “the first computer Go program to beat a human professional Go player without handicap on a full-sized 19×19 board.”
That’s the computer programmer equivalent of the Apollo moon landing.
Marketing With Predictive Analytics
Since then, predictive analytics trained by machine learning have been deployed everywhere, chiefly Google (which bought DeepMind). Predictive analytics are deployed in:
- Healthcare: Knowledge systems to match symptoms and conditions, or find interactions between medications
- Sports: Player stats and statistics
- Weather forecasting: The planet’s weather is a huge chaos system of air currents and systems, but deep learning can analyze the data from past statistics
- Insurance and risk assessment: Predicting the trends that flag unsafe drivers and such
- Credit checks: Your credit score right now is based partly on predictive modeling
- Social media analysis: We’ve seen Facebook and Twitter in the headlines lately kicking hate groups off their platform, found through predictive analysis
Now we come down to the nitty-gritty: predictive analytics based on machine learning informs most major branches of marketing today. We use it to optimize market campaigns, predict customer behavior, personalize content, and streamline customer service functions. Social media platforms are a haven for predictive analytics because we volunteer so much information on there that it’s easy to train AI systems from it.
At Google, if you might have noticed, the search results have been getting “smarter” over time, also thanks to predictive modeling. This is how Google is finding content to serve in a featured snippet or recognizing a location-based search to serve up a map response.
Just imagine the billions and billions of searches Google handles per day. How could one human being sit there and assign a movie listings response to this search and a calculator response to that search? Instead, Google served a set of links, and then its algorithm tracked what users clicked on and when they stopped searching. Aha, 90% of users who searched “Custer’s last stand” ended up on the history page about the Battle of Little Bighorn! We’ll start pulling that page up on the results page.
Pattern matching works with image and audio too, even though it’s a bit sloppy. So Alexa and other voice-recognition technology have improved in audio searches, while Facebook algorithms get better at targeting faces. In these cases, painstaking research had to be gathered to train the AI, since, unlike a board game, the computer has no clear way to determine a success.
You may have even helped in this training if you ever checked images in a CAPTCHA.
This kind of trained pattern matching is an example of heuristic machine learning. Predictive analytics hinges on probability; there is always the small chance that somebody searching for “German shepherd” was looking for somebody tending a flock of sheep outside Berlin, but they’re usually looking for the dog breed.
So, with all these recent advances in AI, are we at a new threshold? Could machine learning push computers and marketing into new frontiers? Might we be creating something a little… too sophisticated? Should we break out those dystopian cyberpunk t-shirts from the 1990s again?
Artificial Intelligence Still Won’t Take Your Job
As we pointed out up top, please ignore all of the panic headlines about the robot apocalypse. Thanks to decades of cheesy science fiction movies, public misconceptions about both computers and people, and lack of understanding about true causes of change in society, AI has become the scary bogeyman to some and, well, the crackpot cult of others.
OK, let’s calm down. Artificial Intelligence is not likely to become the devil or god of the future. Instead, it is yet another tool in our computing toolbox. Bear this important fact in mind: Even the latest machine learning does not allow us to do anything new. It only lets us do the same thing faster. That is still a sharp improvement over the past. Facial recognition can recognize faces, yes, as long as they’re posing nicely for a full-frontal portrait. Even then, it still fails…
Whoops, missed somebody!
That’s even weirder.
But AI marketing does also present some frontier territory left to cover. For example, research is looking into using facial recognition to read your expression and guess your mood and adjust marketing based on that. You look hungry, buy one of our delicious sandwiches.
Count on social media to continue developing more ways to use all the data they collect. AI systems are improving at reading social insight, not just how people react to your brand, but how people talk about your brand and general industry and even read intentions based on behavior.
However, we might eventually come up with a question which we never thought we’d have to ask: Is there a point of diminishing returns in targeted, responsive marketing? How much information do you have to collect and analyze on someone to sell them a product or service?
We’re not even asking if customers are going to be creeped out by psychic ad servers that intrude on their most private thoughts. We’re asking if we can get to a point where all our fine-grained analysis and careful action just won’t make much further difference in revenue. Let’s face it, if somebody’s outside in the rain and we recognize that from peeking through their phone camera, they either have an umbrella or they don’t care, you can show them all the umbrella ads you want.