I work in advertising, in a mysterious region called “media planning” where we have an even smaller island in the shallows labeled “programmatic.” If you’re not familiar with how advertising works, just imagine that our media team focuses on the mathematical predictions of what could work in communications campaigns designed to influence people to buy things … and we are starting to use automated software systems, called programmatic, to assist us in targeting ads.
So I’m really interested, as you may be, in whether robots are coming for our jobs.
Earlier this month on a stage in Fort Lauderdale, Florida, Jon Iwata walked over to a small white robot. Iwata is senior vice president of marketing and communications for IBM, and he’s spent the past few months conducting a road tour promoting IBM’s new global brand slogan, “The Era of Cognitive Business.” The cute plasticky robot was tied into Watson, IBM’s artificial intelligence experiment – powering everything from Under Armour workout wearables to weather predictions. Watson, famous for winning a Jeopardy match in 2011, is at core a computer that can answer questions by processing vast amounts of data. But Watson has two other levels for IBM: Outside, it’s a humanized brand face for a vast technological monolith, and inside it is really an ecosystem of machine learning. Iwata and the robot explained how Watson could help with almost everything humans do, from research to healthcare to, well, advertising. Watson, it turns out, is not a singular robot that is learning, but a vast series of knowledge pools each of which could be siphoned off to perform a specific mission.
“It will be inevitable that artificial intelligence, or digital intelligence, will be embedded and integrated into all things digital,” Iwata said. “Why? Data is exploding today, and most of the data is unstructured.” The volume of data in healthcare, government, utility and media doubles annually. The Weather Channel, for instance, gathers more than 3 billion data points from weather stations to build forecasts. Daily, our human species now produces 2.5 billion gigabytes of data, enough each week to build a stack of thumb drives holding 1 gigabyte each from the Earth’s surface to the moon.
Machines are learning to manage this complexity, finding patterns that lead to insights that in turn push controls. Fly in a modern airplane, and the pilot assists in the takeoff and landing, with most other actions completely automated by algorithms.
So it’s only natural that artificial intelligence would encroach on the ad industry. The art of influencing consumers or business partners to take action is moving rapidly into science. Someone at the end of Iwata’s talk posed the question, via tweet as people do at conferences nowadays, had IBM ever deployed Watson to run a digital ad campaign? “Sure,” Iwata said cheerfully. “Our team tested Watson running programmatic digital, and the results went up 2x over anything we’ve seen before.”
Marketers can predict what you’ll do next (say, catch the flu)
Advertising has always been based on data — marketers at core want to place the message about their supposed value against a human brain that can only be found by some form of data targeting — but the vague concentric circles of targeting have tightened from demographics to individuals to psychological prediction. Old qualitative systems (focus groups, radio ratings panels) and quantitative systems (Portable People Meters that accurately monitor radio tune-in signals among volunteers) are migrating to huge sensor-based systems that pick up individual motion, for extrapolation to what you’ll do next.
Sensors like, say, the ones in your phone.
For instance, algorithms can now suggest where you should go, or not go, to avoid catching the flu, based on mapping smartphones around you. Google for years has mined search data for flu-outbreak patterns faster than reporting from the U.S. Centers for Disease Control, but Alex Pentland, creator of the MIT Media Lab, has gone even further in finding clusters of people who are about to come down with fevers. Pentland uses so-called “reality mining” to evaluate signals from the smartphones people now carry in their pockets. He’s built algorithms that not only can tell who has the flu, by iPhones breaking their daily commutes back-and-forth to work, but which people are just getting ill by sensing the common variances in travel and communications we all make when we start to feel ill. A few days before you get really sick, you make uncharacteristic changes in behavior; people make more calls in the evening to friends or family, seeking an unconscious consolation, as they fall under the weather. By picking up and modeling locations of phones with these early flu signals, Pentland can build maps showing which movie theaters on a weekend night should best be avoided — because more people there are about to get ill.
Data predictions are moving far beyond traffic alerts to forecasting nuances of human desire, health, and behavior. A few years ago, Target sent coupons promoting maternity wear to a Minneapolis household when it picked up signals, from shopping behavior, that a woman who lived there was pregnant. The woman was a high school girl, and her father didn’t know she was pregnant yet.
Robots writing ad creative
While many in the ad industry have boxed this robotic targeting-and-prediction trend into “programmatic,” thinking it just applies to digital banner ads or online video, the reality-mining bleeds into creative, too. Algorithms aren’t just for digital breakfast any more.
Consider the company Automated Insights, which turns datasets into nearly perfect prose. The AP uses it to write more than 20,000 news and sports stories every year, and companies from Allstate to Samsung deploy it for automated business writing. Here’s a real example:
You started this month with $1,800,000 in total pipeline. You have $900,000 in closed/won revenue against your 2015 annual quota of $1,000,000, and this is 150% of what you closed by this same time last year. Damn well done!…”
Play this auto-content out, and the noblest of advertising human innovation, creative for television ads, could soon be automated. In 1978, Donald Gunn, creative director for Leo Burnett, took a year sabbatical to study patterns in television advertising. He formulated that all TV ad creative falls into 12 master formats. There is the product demo ad (HeadOn), the contrast-with-competition ad (Audi vs. BMW), associated user imagery (Justin Bieber relays his cool persona to Calvin Klein), and only nine others. Humans in advertising hold The Big Idea sacred, but computers that can automatically write AP stories surely are not far away from algebraically thinking up a funny Super Bowl spot based on core formulas.
(Male) Actor 1 with (product) (stumbles). (Attractive female) Actor 2 (responds) (sexual tension). (Barrier) arises, then (product) solves (barrier) with (unexpected outcome*). (* must match template for human humor.)
Amy Webb, head of the future-forecasting firm Webbmedia Group, has suggested marketers are one of eight jobs that could be replaced by robotic systems in the next 20 years (along with cashiers, finance managers, journalists, and hell yeah, lawyers). In advertising, algorithms could pull in data on consumer habits, desires, and media trends; parse ad creative for what will work best; auto-generate content; select the media; measure results and optimize to best performance. Six levers. Done.
When humans win
However, the history of AI shows the race to replicate human strategy is not a quick one. In digital advertising, many systems in the past years have promised to use automated algorithms to target ads against the right people. AppNexus and The Trade Desk are two examples of systems that learn over time; punch in target data segments, budget, and the campaign goal, and as the advertising runs out over time the bidding system measures what is performing in driving clicks or conversions to a web form, and dials in the variables.
This approach is spreading in other forms of advertising, including online video and television, as the fragmentation of media channels continues and the variables grow more complex. Marketers no longer live in a world where a handful of TV networks can reach most consumers. Instead of targeting “media” to reach a group of people, marketers must target all of the millions of individuals, each aspect of the audience itself. The audience is a kaleidoscope of demographics, psyches, needs, and behaviors. Target groups cluster and break apart to reform, like starlings in murmuration. Automation and software are required to manage the complexity.
Artificial intelligence in marketing today works best when consumers are in a peak state of interest, matched to a vast set of product solutions. The Google search window, Amazon.com product recommendations, and Netflix movies are all examples of a hungry consumer nudged against a near-infinite supply of options. Personalization works best with plethora.
But absent urgent need matched to huge product inventory, automated ad systems often fail. The “funnel” of logic to hone an ad campaign where most consumers are only moderately interested in a specific product can’t match human guidance, because, somewhat counter-intuitively, humans move faster in both ideation and optimization. Our impulsiveness and aggression, ingrained in us by ancestors who had to fight or flee roaring tigers, allow for rapid moves that algorithms, building data over time, are reluctant to make.
People are more likely to realize when a prior assumption is off course. Tell a computer to take a nice walk in the woods, and it will walk. Tell a person who hears a twig snap, and he’ll adjust course anticipating a bear.
A smart human, for instance, would see a pattern of digital ad click-through rates (the percent of people who are served a banner ad who then click on it) averaging in the 0.08% range and pick up a 0.30% outlier looks suspicious, then dig in further to explore fraud. A technology algorithm designed only to optimize to a target of a high response rate would instead push more marketing funds into that 0.30% high performer, rewarding the fraud. Recognizing the power of humans to guide AI, some automated digital companies such as RocketFuel have moved away from their original “black box” algorithm system to a more-open interface, where humans can evaluate and revise the media targeting strategies. Others, such as LinkedIn, have failed in bids to use data on their users to guide automated targeting across the Internet.
When robots fail
Oxford University professor Nick Bostrom predicts in “Superintelligence: Paths, Dangers, Strategies” that artificial intelligent systems will within 50 years outpace human reasoning. But he also worries that algorithms can go astray, creating huge risks, even if they eventually grow smarter than us. The “Flash Crash” stock market collapse of 2010 was one example, where automated algorithms between a mutual fund complex and high-frequency traders began feeding off each other’s erroneous signals, selling off S&P 500 futures contracts in a cascade that wiped out a trillion dollars of value before circuit-breakers kicked in to stop trading.
“These events illustrate several useful lessons…” Bostrom writes. “Smart professionals might give an instruction to a program based on a sensible-seeming and normally sound assumption … and that this can produce catastrophic results when the program continues to act on the instruction with iron-clad logical consistency even in the unanticipated situation where the assumption turns out to be invalid.
“The algorithm just does what it does, and unless it is a very special kind of algorithm, it does not care that we clasp our heads and gasp in dumbstruck horror at the absurd inappropriateness of its actions.”
Imagine an AI designed to run a paperclip factor, he suggests, with instructions to maximize the production of paperclips, that somehow escapes its computer box to control the world, mine the entire planet, and turn our environment into a massive heap of metal parts. Extreme and silly, perhaps, but robots only do what we tell the robots to do.
For this reason, humans are still needed in advertising, from The Big Idea to media planning to digital system management to analytics. Checking my reasoning, I asked a smart colleague at Mediassociates, Nate Carter, if he thought robots would control our future. “No,” Nate said, “because it’s all about aggression. If I’m a person running a campaign I can go in, turn off an advertising source, and make drastic changes over a short period of time, understanding the ramifications of each. If an AI shut things off and turned them back on, you would see bad results, you won’t know why, and you might have to shut down the process.”
Computers can win at chess today, but chessboards have only 64 squares. In a world of millions of marketing variables, humans are still needed to search for patterns or make guesses when we can’t find one. We’re winning for now, perhaps because instead of focusing on one smart goal, we’re not afraid to try many paths that might be stupid.
Yet … Bostrom is right. Solving chess once seemed an impossible challenge for computers, for it would mean matching the perceived height of human intellectual conflict. Today your Mac has a chess program that can easily beat you. Perhaps solving all of advertising someday won’t be too difficult for Watson at all.