THR Web Features   /   August 11, 2021

Looking Under the Hood of AI’s Dubious Models

The lucrative partnership of obfuscation and hype.

Ethan Edwards

( Oak Ridge National Laboratory super computer. Via Wikimedia Commons.)

In 2018, McKinsey Global Institute released “Notes from the AI Frontier,” a report that seeks to predict the economic impact of artificial intelligence. Looming over the report is how the changing nature of work might transform society and pose challenges for policymakers. The good news is that the experts at McKinsey think that automation will create more jobs than it eliminates, but obviously it’s not a simple question. And the answer they give rests on sophisticated econometric models that include a variety of qualifications and estimates. Such models are necessarily simplified, and even reductionistic, but are they useful? And for whom?

Without a doubt, when it comes to predictive modeling, the center of the action in our society—and the industry through which intense entrepreneurial energy and venture capital flows—is artificial intelligence itself. AI, of course, is nothing new. A subdiscipline dedicated to mimicking human capacities in sensing, language, and thought, it’s nearly as old as computer science itself. But for the last ten years or so the promise and the hype of AI have only accelerated. The most impressive results have come from something called “neural nets,” which has used linear algebra to mimic some of the biological structures of our brain cells and has been combined with far better hardware developed for video games. In only a few years, neural nets have revolutionized image processing, language processing, audio analysis, and media recommendation. The hype is that they can do far more.

If we are—as many promoters assert—close to AIs that can do everything a human knowledge worker can and more, that is obviously a disruptive, even revolutionary, prospect. It’s also a claim that has turned on the spigot of investment capital. And that’s one reason it’s difficult to know the true potential of the industry. Talking about AI is a winning formula for startups, researchers, and anyone who wants funding, enough that the term AI gets used for more than just neural nets and is now a label for computer-based automation in general. Older methods that have nothing to do with the new boom have been rebranded under AI. Think tanks and universities are hosting seminars on the impact of AI on fields on which it has so far had no impact. Some startups who have built their company’s future profitability on the promise of their AI systems have actually had to hire low-wage humans to act like the hoped-for intelligences for customers and investors while they wait for the technology to catch up. Such hype produces a funhouse mirror effect that distorts the potential and therefore the value of firms and all but guarantees that some startups will squander valuable resources with broken (or empty) promises. But as long as some companies do keep their promises it’s gamble that many investors are still willing to take.

It remains uncertain for now how much such hype and hope will deliver, but if some of what is promised comes to pass, the effects may be more than just be cost-offsetting linear advancement. If this technology advances in just the wrong way, labor may not have enough to do, and there are fears that excessive automation might lead to a major employment crisis and a worsening of class inequality. Andrew Yang justified his signature universal basic income program on these terms. Setting aside the merits of his particular policy proposals, Yang had this much right: If AI displaces jobs, we should be making moves now to mitigate that effect. On the other hand, if it will create jobs, we should be driving it further, or, if it does neither, maybe we should stop talking about it so much.

These topics are not new. Neural nets in their contemporary incarnation have been around for almost a decade, and we can look at the areas they have changed and the ones they likely will. AIs are very good at getting results comparable to human experts as long as they have a sufficiently large and well structured dataset to train on. But AIs, because the methods are opaque to human reasoning—absent further research—are very bad at explaining or adjusting their results, not to mention communicating and cooperating with people, which only augurs the fear that AI will replace rather than complement current workers.

The earliest successful application of neural nets was for image recognition—the task of identifying with words what an image contains—and it has continued to be one of the dominant areas for research. This has revolutionized software products like Google Image Search. Facial recognition is probably the most talked about application of images and is being used by businesses and governments for rather worrying projects, but it’s not clear that these are contributing to unemployment. The new Amazon Fresh stores, which likely use neural nets to detect products taken from the store, would entirely displace register operators, but with only about a dozen stores opened, the subsidiary doesn’t exactly represent an existential threat to the US retail industry. One notable prediction of where this technology would be successful is medical imaging, enough that AI pioneer Geoffrey Hinton predicted in 2015 that we no longer needed to train radiologists because by 2020 they would be rendered obsolete. The technology works fairly well, but there does not appear to have been major efforts by technologists or doctors to use the technology in clinical contexts. There were more radiologists trained in 2020 than 2015 and there’s demand for even more. Even if the technology works there are major barriers to actual adoption.

Predictions that AI will transform the American economy look rather premature, if not like wishful thinking. Language processing can successfully generate realistic conversations and has primarily been used in customer service applications. The labyrinthian phone menus known to everyone who has had to complain to a cable company or ask about health insurance have been upgraded to slightly less frustrating AI chat agents. While it may seem like only a small upgrade, these systems are able to answer more questions than their predecessors and have put actual human customer service agents out of jobs, even if they have not made the customer experience much better. Self-driving cars have the potential to be the biggest employment disruptor—about 1 percent of the US population works as a truck driver—but the technology to replace them may still be a long way off and would have to go through countless regulatory hurdles before it would warrant worry. Previous predictions that AI will advance at a steady pace and change each of these industries have so far failed to pan out, so the best one can say for them is that research progress has been slower than predicted.

But the problem with AI is not just that the industry hasn’t kept pace with its own hype. That criticism, perhaps trivial, could be lodged against practically any sector that is heavily invested in emergent technologies. No, the far more significant problem lies at the heart of neural net technology: It actively resists accountability and transparency. Beyond a few researchers, the industry demonstrated no interest in producing legible, predictive models. A survey of the literature, just to take one among many questions of concern, shows that no one has enough confidence to give even a ballpark estimate of how many jobs will be lost or how much productivity will increase in the near or long term.

The popular press, of course, thrives on the provocative pronouncements of breathless experts. AI will create $13 trillion in value by 2030. It may also create $150 trillion by 2025. One can even find estimates of the $2.9 trillion value it will create by 2021, which do not appear to have been checked for accuracy now that we’ve passed the date. These estimates, posted excitedly in headlines by the tech and business press ultimately originate from reports by consulting firms like McKinsey and Gartner. Judging by the wide variety of figures and the disinterest in validating the claims, these estimates are not made in order to predict the future accurately.

For a more serious account of what an AI future might look like, one might expect the consultants at McKinsey to do much better. If one looks through the McKinsey report on the AI Frontier, which estimated the creation of $13 trillion in value by 2030, it hardly seems like a firm prediction. The report is very well put together, and explains its methods and judgments clearly, but the methods do not stand up to serious scrutiny. Following the now common usage in the industry, the report includes robotics and automation broadly within the category of AI. The economic model they’ve constructed uses data from a survey of 3,000 firms on how much they use and expect to use AI and how much self-reported value has been generated by it. The report extrapolates these values into a linear estimation of technological progress and economic competition. And although the report clearly states that its figures are not intended to be read as forecasts and only for directional perspective on the impact of AI, those directional perspectives are all pointing the same way.

If one reads the report carefully, it becomes clear how much rests on certain dubious assumptions. How can we assume that advances in computer vision or language modeling will continue at pace? How can we be sure that there will not be another AI winter in which the research will stall and those working on it will be shunned and out of fashion, wiping out the value of AI companies overnight? What about all the messiness of regulation or unexpected roadblocks or industries refusing to adapt for reasons good and bad? The report skates above such inconvenient considerations.

The report maybe be overly optimistic and simplistic, but it is nevertheless honest and tries to make a concrete model for the crucial question of what will happen to the world economy in the next 10 years. We should be discussing it. But the way that the report has been translated into popular discourse has not encouraged a wider discussion of these questions. The figures and numbers were read and digested by a few, and then put into PowerPoints and memos and headlines to show that AI is booming and should be invested in right now. McKinsey, Gartner, and many other firms offer such services, so they are carried upward with this excitement. The tech publications know that hype sells, and these numbers are a gift, so they feed the hype further. Speculation, processed and shined with charts and numbers, becomes knowledge, and that knowledge becomes the basis for a feedback loop of hype and investment.

The numbers produced by the model McKinsey created has the sheen of knowledge produced by top experts. If they were to use AI, it would have an added authority because its numbers would be produced by an objective and intelligent system using the latest, hyped methods. But I was able to read the report, check its methodology, and even though many of the data sources are confidential, I could still walk away with a skeptical opinion of the conclusions.

All of this suggests the disturbing conclusion that one true value of AI is its ability to procure the capital of financial investors and the support of corporate executives. A lot of machine learning and AI modeling is built on methods that are not public or even possibly explainable. We may not even to be able to really tell when this happens if we only trust AI to tell us. Such models and the datasets, shrouded in secrecy, offer little to the skeptical inquirer except one (appropriately computational) binary: believe or doubt.

But perhaps the opacity of AI is a feature more than a bug. And not just for venture capital purposes. There are many elements of society and the state that do not care so much for increases of productivity or labor-saving techniques as they are interested mainly in producing new sources of authority. And let’s be clear: One of the most valuable parts of such a project comes from appealing to a source that—unlike the divine or the traditional—has a currency in today’s society. Governments might use AI, for example, for their bureaucracies—screening potential immigrants and future criminals—not because the techniques are effective, but because they seem objective. Like other technocrats, hiring managers and college admissions committees have much to gain from AI’s opacity and supposed objectivity. If a company’s human employees reject a potential hire because they have a “criminal profile,” it would be outrageous, but systems like AI can make such vague pronouncements palatable with a decision process impervious to audit.

When Facebook’s Mark Zuckerberg was required in 2018 to testify on the company’s content moderation to avoid regulation, AI’s objectivity served him well; he answered any major question about content moderation or disinformation with the assurance that AI tools would somehow solve the problem. Obfuscation, a strange partner to hype, thrives under the same conditions. Even if AI does not advance enough to give us economic growth, it is already more than capable of delivering the goods to those who need them most. If the AI boom continues, the circle will soon be complete: AI models will no doubt predict the impact of AI models.

But the technology is important enough that it deserves far more serious political and cultural scrutiny. If we really care about the accuracy of the results, we will support AI research that advances legible models, which can still be useful even when they are wrong. The short term hype that drives AI may be driving it in exactly in the wrong direction. Models, whether with three variables or three hundred billion, are valuable in the long run only if we are free to take them apart.