How to tell when someone is talking BS about AI

It may seem like a drop in the bucket, but both the number of deals and the amount raised is on the rise. What’s more, Google launched its own AI-specific fund earlier this year, and, more recently, IBM and MIT inked a ten-year, $240m deal to create the MIT-IBM Watson AI Lab. And the topic has even inspired public spats between some of the biggest names in tech.

In short: AI is unquestionably the biggest buzzword in digital marketing, joining the ranks of 'big data', 'programmatic' and 'mobile-first,' which occupied a smilar role in recent years.

And with buzz comes BS.

Mike Grehan, chief marketing officer at digital marketing firm Acronym Media, says the hype around AI is reminiscent of the dotcom era, adding that many PR-hungry startups spout BS about AI and machine learning.

“And yet, when they explain what the core business of the new company is, it’s not difficult to detect that they are making stuff up for the sake of securing funding, shall we say,” he adds.

This is not just limited to smaller startups eager for their first round of funding, according to Grehan, adding that larger, more established tech companies are guilty of it too.

It's a thought not lost on Danny Hopwood, executive vice president of programmatic in EMEA and strategic advisor at Publicis Media, who says his BS alarm goes off when he considers the number of companies that were started in the last year with AI in their business models.

“We saw this a few years ago – companies that had nothing to do with programmatic – it was not part of their DNA – and then when programmatic became a popular thing and the way to do business, companies picked up that name as well, riding on the coattails,” he says.

For his part, Paul Martino, partner at early stage venture fund Bullpen Capital, called the space massively over-invested. In fact, he recounts how Bullpen has an office joke in which any startup that could make it to the end of a presentation without saying, “AI” or “machine learning,” would receive a check the spot.

“Guess how many checks we wrote?” he quips.

First: What exactly is AI?

Part of the problem is that “AI” is often used as an umbrella term for related topics like machine learning, deep learning, sentience, singularity and even killer robots, but these terms are not interchangeable. Although that’s not to say AI is easy to define, which is perhaps illustrated best by this Venn diagram:

Want to understand the AI, machine learning?

This graphic will NOT help you

h/t @stevenkuyan pic.twitter.com/Xsgo8eZGug

— CB Insights (@CBinsights) September 7, 2017

As Grehan notes, another hot topic – machine learning – is actually a subset of AI.“It’s all about machines learning from experience, from real world examples: the more data, the more it learns. Most AI applications and processes are still made of fixed rules and do not learn from data,” he points out.

“It’s rules-based with lines of ‘if this, then that’ code written by humans. That means many so-called AI technologies being touted are, in reality, not ‘strong AI’, but most likely machine learning systems. It’s a subtle difference but it matters when it comes to perception.”

Today, there is weak AI in the form of Siri, autonomous vehicles and face recognition, but Grehan adds a further development – deep learning – is not likely to be mastered by many companies independently until most have sufficient/relevant digital data to train a machine learning/AI process reliably.

It’s also worth noting that Google doesn’t answer questions – it returns factoids, but cannot answer questions that require understanding, prior knowledge and real thought, which would be the case if it had artificial intelligence, according to Grehan.

He adds: “So when people start talking about machines ‘thinking and making decisions like humans do – bullshit!

“Machines can learn to perform systems and processes when programmed by humans. And machines can be taught to learn by other machines. But whether by human, or by machines, they can’t be taught to think for themselves yet.”

In other words, both Siri and driverless cars are great at what they do, but they are weak AI. Case in point: Siri can order you a pizza, but can’t drive you home; a driverless car can get you home, but it can’t get you a pizza.

Further, Martino points to John McCarthy, the professor emeritus of computer science at Stanford who coined the term artificial intelligence in 1955. According to McCarthy the problem with AI is that as soon as an AI masters something, it’s no longer AI, so the target keeps moving. For example, in ad targeting, figuring out the demographics an ad best performs against would have been called AI 20 years ago, but is now it’s standard.

“Every year the definition gets larger and larger,” Martino says, adding that most vendor claims of AI in marketing are probably slightly incorrect.

“We’ve been using predictive systems, and self-optimizing click-through rates for 20 years now,” he adds. “Putting on a new coat of paint and calling it AI is disingenuous.”

Real-world thinking on AI and marketing

AI offers the most value when data scientists operationalize models and automate millions of decisions on a scale that would be unfeasible to expect of a human being; like ranking paid search results, or up-selling in a commerce platform.

Omer Efrat, director of product management at mobile ad exchange Inneractive, says marketers still struggle to use data to reach the right consumer/device/moment. As a result, AI-powered automation will eventually help make ads work better – and will eventually help companies smaller than Google and Facebook utilize this technology. In other words, it’s not just showing an ad for cat food to cat lovers, but matching exact experiences to exact users, such as those who want coupons or are concerned about ingredients.

Similarly, angel investor Eric Franchi points to the ability of AI to constantly learn and to be responsive to users’ needs, as well as to use exponential learning and improvement to transform content.

So if applied correctly, AI-produced creative copy can transform from being nonsensical, to pretty good, to being able to match what a human is able to do in a relatively short period of time. "But then it is able to start to learn how to make it better and apply to 1000 different audience segments and apply multi-variates and literally use AI to develop creative copy at a pace/scale that would be very challenging for one human not only many humans,” he adds.

So with an ever-changing definition and the promise to change marketing forever, how can you tell which vendors touting AI products are legit and which are, well, full of it?

5 questions you need to call out BS in AI:

1. Does this company have enough data?

Hopwood says companies that can utilize AI technology are few and far between.

“More often than not, companies that end up doing machine learning are ones that have massive amounts of data with multiple dimensions to it [including identities and geographies] – all sorts of different data points and they use machine learning to harness data in one place,” he adds.

That means it’s limited to companies like Google, Amazon, Facebook, IBM and Adobe, along with newer entrants like Netflix, Uber and Dollar Shave Club, which have massive infrastructures.

“Within that infrastructure, there’s massive data and they’re… making use of the broader sense of AI… to assess and look at that data to create personalization,” Hopwood says. “Uber’s surge pricing is based on a form of machine learning, I suspect… most programmatic buying is predicated on an algorithm…[and the] algorithm is probably using a form of machine learning to figure out how to optimize campaigns in real time.”

Grehan agrees the most exciting examples of machine learning and AI come from universities and tech giants. “So you need to be aware of experts who promise to almost instantly revolutionize your company with the latest AI technology when the likelihood can be that they are just rebranding their old rules-based solutions as AI,” he adds.

2. What was this company doing before?

Franchi, too, says he likes to look at what any such company making these claims was doing last year. “If last year, this was a DSP that was touting its mobile and data capabilities and all of a sudden today, ‘we’re an AI company that…’ – insert what your business does there – I would be/am suspect of how you became an AI company over the span of months to a year, and where is this expertise is coming from,” he adds.

Hopwood says a pretty good dipstick when talking to companies using AI is to find out what they can offer you – and if it isn’t much more than optimizing campaigns, it should raise flags.

“If someone is coming to you with the greatest thing in the industry now, but the offering is just as simplistic as running ad campaigns for you, I would question are we making the best use [of the technology]?” he adds.

3. Was the technology developed in-house?

Franchi also suggests asking whether the company is tapping into technology from, say, IBM Watson or if it has developed its own? “From a customer perspective, [using Watson is] okay, but, from an investor perspective… that’s important,” Franchi adds.

“If IP is going to matter over time in the case of the tech business, building on top of Watson or something like that could be good to start, but at the end of the day, you want to own your own IP.”

4. Who works for the company?

Franchi also advises talking to the people involved in the business in order to really understand what their qualifications are – including degrees and university programs – and whether they are data scientists or have advanced degrees in machine learning and computer vision.

“To me, that’s been the most helpful filter,” Franchi adds. “Some of it is so cutting edge, you need to be deeply engaged in the study of it to have the most current understanding.”

5. Can I try before I buy?

“In a hype-cycle like this, sales people [use] buzzwords, CEOs spin a tale, [but], as a customer, you have to go use them,” Martino says.

Further, he adds, it is impossible to tell from vendors’ glossies and case studies which system will work best. Martino says when he was the chief of an ad tech company, he would let customers such as Omnicom and WPP use the product for a set time period before they had to buy it, and when they saw results, they became customers.

And, of course, you can also avail yourself of online resources to beef-up your own understanding of machine learning, Hopwood adds.

If all else fails, you can always ask a data scientist for help, Grehan advises.

AI will be discussed at length during this year's DMEXCO conference where The Drum will be reporting from the floor. Click here for more coverage