Designing Agentive Technology: AI That Works for People

eBay Followed Searches: Interests in Stuff

Most folks know of eBay as a great place to go and find something to purchase at a good price, but it also has an agentive feature called Followed Searches. Launched back in 1999 as the Personal Shopper, this feature lets users take any search and keep it going. Even if I don’t find one right now in a style, size, and price I want, I can ask the site to keep an eye on all new items that go on sale there for me, and let me know when any match.

eBay has search tools for that moment I hope to find something for now, and agentive tools to help me keep track of things I’m interested in.

What’s interesting for the kind of search embodied in these three agentive examples is that creating too specific of a search term can work against you. Following a search for green Ted Baker shirts selling for $53 in San Francisco might find you exactly the shirt you’re looking for right now, but it would not provide a fruitful persistent search in the future. Users need to be able to set up more abstract searches, and agents may need to help them do it.

These kinds of persistent searches move you from having to go out and find stuff of interest yourself to letting stuff of interest find you—for example, to subscribe to your favorite artists, or authors, or to notify you when cool things are mentioned online. The dark side of this might be opt-out advertising, but in the best cases, these agents turn the tables so that your interests find you. How will marketing and advertising change when this becomes the norm?

They’ll Do the Work You’re Not Good At

Autopilots are handy because getting from point A to far-away point B can be monotonous in the middle, and people aren’t reliable at those kinds of attention-endurance tasks. Fortunately, for the past century, people have been developing systems to help with that part of the journey in boats, planes, and increasingly, cars.

Autohelm Steers the Boat

Boat captains navigating easy seas simply need to keep the boat pointed in the same direction. Early mechanical systems worked by using a wind vane to keep the boat at the same angle against the wind. Today these systems are known by a couple of names, with marine autopilots being the most common and autohelm being a genericized trademark of Raymarine’s product line, which nicely distinguishes it from aviation autopilots. In their simplest mode, the captain presses the AUTO button, demonstrates the angle at which it must be held, and then lets go to attend to other things around the boat, like perhaps a nice hot buttered rum or the next track of yacht rock.

Image source: Norm Bundek, Montgomery Sailboat Owners Group Photos

More expensive configurations pair with GPS, sonar, and course-plotting devices to make the autohelm aware of the heading, obstacles above and below the water, location—and corresponding maritime law—and planned course. Touching the tiller or ship’s wheel can be the exception and not the rule.

Understanding this function of the autohelm, it’s easy to see how a captain on a short pleasure ride might get distracted by other things on the boat and need to recover quickly when she finds herself off course or hears the drone of an alarm from the autohelm, or finds her vessel in all sorts of possible trouble. Since the device is an add-on to the normal mechanics of a ship, the pilot can use normal means to assess the problem, quickly disengage the device to take manual control if necessary, or troubleshoot the electronics if the problem is the technology itself.

Autoflight Pilots the Plane

Airplane pilots have to manage more complex variables with less margin for error than their maritime counterparts, but long trips can still be fatiguing. The earliest mechanical autopilots worked like the simple autohelm, but with rudder control being augmented by an attitude sensor adjusting the plane’s horizontal tail flaps, or elevators. Nowadays, autopilots are required for most long-range passenger planes over a certain size, and they consist of many subsystem controls for altitude, speed, throttle, heading, and course. While airplane pilots are always busy and can’t just zone out, they do rely on the auto flight systems to manage some of the tedious aspects of flying and to warn them when there is a problem.

Image source: Wikimedia commons

Don Norman has studied the interaction of pilots with autopilot. Norman estimates that a pilot flying at 25,000 feet up has about five minutes to figure out that there’s something wrong, then decide what’s going wrong, and finally to recover in order to save the plane and the people aboard.

Autodrive Drives the Car

Even though driverless cars aren’t yet common on roads, we’re already dealing with autodrive. (Aside: Dear future, forgive us. We’re still in that transitional phase where we have to call them driverless, to distinguish them from the manually driven variety, but you’ll know them simply as “cars.”) Manufacturers like Tesla, Volvo, and Mercedes-Benz car models have the driver facing forward, ready to take over. Google’s driverless cars are eventually meant to be wholly agentive, so there won’t be any need for passengers to suddenly take control. But in the near term, while the technology is being introduced to roads, riders and even legislators are most comfortable with a driver sitting at the helm of the driverless car, ready to take over should the agent fail.

Image source: Google

But if a, uhhh, driver is just sitting there not driving, can they really stay, just sitting there, keeping their attention on their non-task constantly? Ten seconds is all it takes for a user’s attention to drift while using slow software, and I’m pretty sure a car trip won’t be worth taking if it’s under ten seconds long. Perhaps the car will have to introduce some means of keeping the driver actively engaged in the driving, such as a game that drivers play by trying to match the software’s driving. But if not, then the sitrep-and-takeover will present major problems to the driver who’s just about to win a difficult, timed game on their phone and has to drop that to wonder what that alarm is all about.

They’ll Do the Things We’re Unwilling to Do

Shotspotter is a civic agent that constantly listens to a large number of microphones that are sprinkled across a neighborhood. When it hears gunshots, it compares the timestamp on each microphone to triangulate the location of the shots to within a meter’s accuracy. Within seconds of the gunfire, officers can be on their way to investigate.

When I spoke with a representative from Shotspotter, she explained that the service is helpful for more reasons than just decreased police response time. It also helps ensure that the shots-fired signal reaches the police at all. One factor is the bystander effect, in which people presume that surely someone else has reported it already. This might affect people living anywhere. But citizens living in high-crime areas, she explained, can often fear being labeled a snitch and suffering consequences for reporting crimes. Shotspotter takes this responsibility that no one wants unto itself.

Image source: Shotspotter, Inc.

They’ll Do the Embarrassing Things

Going on first dates can be harrowing. Who knows if that person is genuinely charming, or a well-practiced sociopath? You should have a backup, someone who will check up on you. But then again, you don’t want to burden friends with remote chaperone duties for every single date you go on. Enter the safety agent that is kitestring.io. You tell kitestring when to check up on you, information about the date you’re going on, and an emergency contact. When time is up, kitestring sends you a text to make sure you’re OK. If you don’t answer, or reply with your fake safe word, it forwards the information about the date to the emergency contact, presumably so they can take immediate action to find you and ensure your safety. If you answer with the real safe word, kitestring erases the information about the date and stops checking up on you. It’s not a replacement for being careful, but an additional tool in your arsenal.

They Will Allow Play…

In some domains, users are happy to let the agent run, and only think about it when there’s a problem. Managing a long-term investment portfolio is one where you specifically don’t want to look at it every day. But in other domains, you’ll need to recognize that users will still want to play.

Here I’m thinking of iOS Autocorrect. (It’s closer to assistive tech, but is still an instructive example for our purposes here.) In this low-level typing function, if users mistype a word—or correctly type an unusual word that is not in its dictionary—the operating system will offer to replace the word with its best guess of the intended word. Spell checks have been around almost since the beginning of computers, but Autocorrect has two differences. First, the interaction design is such that most of the time, people don’t realize their words have been autocorrected until after they’ve sent text messages or made status updates. Second, being on a mobile OS with an on-screen keyboard means there are plenty of mistakes to be corrected.

Most of the time, Autocorrect works pretty seamlessly, changing people’s mitsakes to mistakes before they realize they’ve happened. Some of the time the corrections are nonsensical. And a few times they can be genuinely funny.

But if you are a playful user of language, Autocorrect is much more of a damned nuisance than a help. There’s a whole philosophical tangent I’ll avoid about why being a playful user of language is important at any age, but even if you’re personally prudish with your words, note that it’s a fact of teenage life and part of the currency of subcultures. It’s a way to create and celebrate a shared identity. Consider the recent popularity of turnt, bae, yas, and nudnik. OK, that last one is from the 1920s, but a flapper with an iPhone, can you imagine? Since both declining the Autocorrections and adding to its dictionary slow your text entry down significantly, it’s much easier to just turn the feature completely off. Better would be an agent that gauges your degree of playfulness, or updates itself with language trends of your peers, and backs off accordingly.

(Note that we are on the second edition of Damn You Auto Correct compilations from the blog of the same name, so it’s kind of a thing.)

…and They Will Encourage Discovery Through Drift

IBM has been working on its deep learning engine Watson since 2005. Although its first public implementation was to compete against humans on the game show Jeopardy!, since winning that show in 2011, the project has evolved in some other directions. Chef Watson is one spinoff that came online in 2015. At first, it seems like any other recipe database: a search form lets you input ingredients to find recipes.

But there are several fantastic things about Chef Watson that make it different. For one, it automatically searches for other ingredients based on synergy with your starting ingredient, for example, the presence of similarly flavored chemical compounds. Next, it finds existing recipes for the set of ingredients, and it drifts them. By drifting—my term—I mean taking a known recipe and finding replacements for ingredients. For instance, with a little drift, a lemon tart might become a lime tart, and with a lot of drift it might become a mangosteen tart. Or perhaps a Yuzu Kouign-amann. Seriously, this software will surprise you. In the pictured example, Watson has taken the recipe for Tea and Lemon Gravy from Bon Appétit and drifted it so that the Cara Cara orange replaces the lemon.

Additionally, it checks its database to ensure that the starting point it gives you is unique. As far as it knows, that recipe has never been tried before in recorded history. That’s a pretty amazing thought. Sure, you’re going to run into some culinary dead-ends, but think about what deliciousness you’re going to discover.

But hang on, it even goes one better, because Watson knows, Hey, I’m just a computer semi-randomly modifying recipes, not a human with taste buds—and limited by access to ingredients—think you’ll find a yuzu at your local bodega?. So it lets users drift recipes manually, too. In the above picture, I’ve clicked on one cara cara orange to see that I can pick any of the fruits in the list to suit my tastes or drift it further.

In true agentive form, Watson even lets me store my food preferences—ovo-lacto vegetarian, distaste for mushroom and sea flavors—and will send me recipes occasionally based on the general calendar. In the future, I fully expect to give it permission to look at my personal calendar and receive suggestions when I am booked for a potluck.

Between play and drift, your agents won’t ask you to adhere to rigid rules, but rather to channel and encourage creativity and discovery, all while making it easy to make smart choices.

They Help Achieve Goals with Minimal Effort

Tasks by definition are small things, bound by time or the simplicity of their undertaking, but performed by people in the service of some larger goal. For instance, a person can take photos (task) to preserve the happy vacation memories for a group of friends (goal). Another person can take photos (the same task) to capture the elusive beauty of natural forms (goal). Most of design focuses on designing for the task, and that can produce perfectly functional things, but design that focuses on users’ goals are much more loved because they help us do so much more. They fit into our lives and identities. You can even say that an agent is the ultimate expression of goal-focused design thinking, because it gets users to their goals with the least effort possible.

That simple design principle explains why agentive tech that is focused on the goal will win out over agents that focus on the task. Take, for example, cameras. Point-and-shoot cameras have existed for a long time now, but these tools eliminate the work of managing all the complex and interrelated settings involved in taking a picture—but you’re still taking a picture. Contrast that with the life blog cameras like the Narrative Clip. It’s a small square device with a clip on the back and a small lens on the front. As long as the lens sees light, it takes a picture every 30 seconds. Clip it on your shirt, and by the end of the day, you’ll have around 2,000 pictures. That would be an overwhelming number to sort and sift through, except the Narrative uploads all the images it took to a server, where smart algorithms first divide them into segments of the day, pick the best ones from each segment, and share those good few with you via an app. The app lets you second-guess its selections and annotate or share them on social media. So as you can see, it is kind-of a camera. But it’s also not. It’s an agentive camera that focuses on your living life and having great photos, rather than taking photos.

Image source: Narrative

Similarly, the Roomba is a vacuum unlike the ones that came before it. The goal of the design of prior vacuum cleaners was to make it light, powerful, and ergonomic—that is, to make it easy for a user to clean their floors. But the Roomba rethinks the problem. It’s an agentive vacuum that sits in its charging cradle until scheduled cleaning times. Then it roams around vacuuming until it gets close to running out of battery, when it returns to the cradle. In typical use, users only have to empty its dustbin occasionally.

Image source: iRobot

Both the Narrative Clip and the Roomba turn the tables on their predecessors by focusing on meeting the goal with an agentive technology, rather than being a good tool for users to complete tasks.

The Scenario Is—a Lifetime

Most technologies focus on the moment of use. Even the tools you use to embody users—that is, personas, are typically tied to an age, a small moment in time. But since agents handle things for their users over the long haul, agentive technology encourages you to think about the scenarios that make up a whole life.

Betterment is a roboinvestor that helps users keep investment portfolios balanced to specific risk targets. It lets investors specify long-term goals—from drawing a particular income in retirement to making a down payment on a home. Since these are long term goals—playing out across decades—they have to take into account the change in persona goals from when they are young, bright-eyed, and bushy tailed, to when they’re about to hit retirement and be cautious about losing what they’ve gained, all the way through retirement, drawing down on their savings and enjoying their golden years.

Technology is accelerating. It’s hard to know exactly what the world or technology will look like in five or ten years, so the far-horizon scenarios can be more aspirational and even vague. But agentive tech encourages you to take this long-term view and at least think about how it might be, knowing what you know now.

There May Be an Arms Race of Competing Agents

In this book, I speak mostly about agents acting on behalf of users with benign intentions, but let’s not be Pollyanna about it. Bad actors are out there, causing mayhem and trying to separate you from your hard-earned money. Fortunately, you’re already quite familiar with one agent who does its work almost entirely in the background, saving you from Nigerian Princes and the tedium of unwanted sales messages—that’s the humble spam filter.

You may not think about it as an agent, but that’s what it is. Persistently on, it looks at every incoming message for telltale signs of spamness, sent from any accounts that have been confirmed as spammers, as well as any senders that you’ve personally blocked or personally blessed. And these it silently sweeps into a folder that you’re free to check anytime, but from which it clears out old messages on a regular basis.

Sadly, for every advancement in spam filters, spammers will seek some new way around them. They’re always there. They’re like raptors, testing the fences, trying new strategies. But spam and its related scourge, computer viruses, are enough of a hassle that your immune system filters will keep evolving with them.

It’s Going to Be Big Enough to Affect Our Infrastructure

I spoke about self-driving cars earlier in this chapter. And in addition to the question of what you’re doing while you’re in them—monitoring and ready to take the wheel at a moment’s notice or catching up on the latest holo-toons—what’s even more interesting are the conversations about what these cars will be doing when you’re not in them.

Image source: Mercedes-Benz

They could just park themselves in your garage, driveway, or find a spot on a nearby street. But given agentive cars, why would they do that? They could attend to the small amount of taxi demand for cars at night, earning some money for their owner. They could drive themselves for maintenance or cleaning. They could be rented as a nighttime delivery vehicle, or as a remote-controlled, mobile sensor network for urban planners, scientists, and first responders.

If they’re not being used in any of these ways, they still don’t need to park right near you as long as they can get back to you by the time you need them again. Could they find otherwise unused spaces, allowing homeowners to reclaim driveways for human purposes? Or could they park on a designated lane on a nearby freeway until they’re needed?

Whatever the answer is, we’re going to need to rethink the urban landscape for new efficiencies when agentive things can move themselves out of the way. This will allow us to take some of the massive percentage of urban land use dedicated to parking—in L.A., it’s estimated to be as high as 14% of the total—and reclaim it for people.

Places and Objects Will Need Them

Places and objects have needs. Since the dawn of time, we as their users have needed to routinely keep tabs on them to see if they need maintenance of some sort. Do the kitchen knives need sharpening? Does the office have enough Post-It Notes? Do we need to air out the garage from fumes? But if these places and objects are imbued with agentive technology, we can offload that monitoring onto the things themselves. The knife block can watch the knives for when they need sharpening. The supplies closet can send emails when the paper is low. The garage can watch the air for problems. The responsiveness of objects is captured in conversations around the Internet of Things, but it’s instructive to think of the Thing in IoT as a place that agents inhabit.

A striking example is represented by the services offered by GOBI Library Solutions to academic libraries. Most people don’t think about it, but managing the collections of tens of thousands of books is a major undertaking. Librarians have to keep an eye on the demand of esoteric subjects they themselves may know nothing about, and they must continually adjust the collection to match need. GOBI’s demand-driven acquisitions is an agent that keeps tabs on the borrowing database for the library; both when patrons borrow books from the collection and when the library must borrow from another library to meet demand. It then routinely performs an analysis to determine how the collection needs to adjust, and can send orders for those books with the budget approval of the collections manager. It is an agent that works with the librarians to keep the knowledge base as relevant as possible to the changing needs of its users.

People have similar needs managing places and objects. How about a plant that lets you know when it’s thirsty? Or a civic agent to let you know what museum exhibits touring through town would interest you?

They Will Help Us Overcome Some Human Foibles

For all our awesomeness, humans have some built-in foibles that are really hard for us to overcome. (Check out the full List of Cognitive Biases on Wikipedia if you want a deep dive.) Any time we humans perform a task, we carry those biases with us. They can interfere with our effectiveness. Those foibles may be minor for the individual, but can add up to serious problems when aggregated up to the level of cities, nation-states, or a planet. Traffic is an example most everyone is familiar with. It’s stressful and wasteful to sit in traffic, and moving feels better, so people try to find shortcuts that ultimately cost them more time, use more gasoline, and can cause further traffic snarls for everyone else.

But that problem can be addressed when wayfinding is handled by an agentive app. The community-based traffic and navigation app, Waze, has a number of nice agentive features, but for purposes of this point, let’s look at traffic distribution. It constantly re-evaluates an individual’s routing to determine if taking an alternate route—including side streets and shortcuts—would actually buy that driver more time. If it doesn’t, it keeps to the current plan. But if a reroute would mean more time, it alerts the driver that a faster route is available and lets them opt-in. This recommendation is based on actual, real-time data rather than our simple lizard-brain emotions, and ensures that traffic as a whole is being routed across the map in an actually efficient way.

In addition to being freer from our cognitive biases—simply by dint of being algorithms—we can go one step further by adding virtues in. Imagine the agentive car that coordinates with other cars to distribute its rider’s journeys across time as well as routes. Could agents be programmed to avoid other tragedies of the commons?

Using Them, People Will Program the World

Conversations about Pervasive or Ubiquitous Computing have been around since the late 1980s, but we’re now living it. It’s difficult to walk in public out of the view of a camera. We’re carrying and wearing more computing power than it took early astronauts to get to the moon. The technological skin of the planet is evolving and growing all the time, and its disparate parts can be connected by one of the most abstract agentive technologies that is out there.

The If This Then That service (ifttt.com) allows you to create persistent formulas, called recipes, which are exactly what the name implies. If the specified triggers are met (the this), then the service carries out the actions—the that. And that’s it. Each recipe acts like a little agent, watching and measuring the pertinent data stream until it gets to perform its action, and then it settles back down again, eyes firmly on the stream.

Triggers can be made more aware by connecting different data streams: social media, car computers, calendars, email, and phone sensors. Actions can be made more powerful by authorizing the service to add to your social media stream, to text or email you, to control your home automation system, or to create documents that capture important data.

Some of these are pretty basic. One recipe looks for rain in the forecast and sends you a text each morning if you need to take along an umbrella. Another keeps your profile pictures in sync across social media services. If you have a connected car, one can send you an email with a map every time you park in the city.

Authors can make their recipes public, and there are many vast collections of recipes—for work, home, music, health, and even outer space. IFTTT is a huge collection of tiny agents that are all making the world a more connected space. Ultimately, I believe that large brands will offer some of these as features of their products and services, but it’s really nice to have an independent service that gives people a platform to play.

Our Species’ Future May Well Depend on Them

I know, that header sounds overblown. But hear me out. You know how in 600 million years, the increasing brightness of the sun will interrupt the carbonate-silicate geophysical cycle, ultimately interrupting plate tectonics, halting volcanic activity, and stopping C3 photosynthesis? Six hundred million years sounds like a long time until you recall that life is estimated to have begun on this planet around three billion years ago. If we’ve only got one billion left, life is way over half done. Which means that—presuming we make it that far—we’ll have to migrate from this planet to another. Ultimately, space travel is the only way our species will survive, and space exploration is a critical part of that.

Image source: Nasa

Fortunately, that’s already underway, and it’s being done smartly with robots that don’t have our messy, frail, and expensive biological requirements. But space exploration quickly runs into a problem of communication time. Even light speed communications have limits. Sending a message to just Mars involves a 4- to 24-minute delay—depending on our positions in orbit— and even if replies are instantaneous, adds up to an 8- to 48-minute delay between back-and-forth responses. That’s not too bad. NASA’s Mars Rover won’t get into too much trouble in that time.

But the farther out it travels, the longer that delay is. Jupiter is between one and two hours. Pluto is roughly once per day. The farther out we travel, the more time there will be between our communications with exploration robots, and the more we will need the robots to handle things on their own. What should it do if it encounters an immediate problem or opportunity, flying through the blackness of space, or rambling over the cold stones of a distant planet? We need to equip these representatives with the right sensors, actuators, rules, and exceptions to learn what it can learn, send the information back to us, and still keep exploring. Fortunately, this isn’t some gaping hole in space exploration strategy. It’s a known problem, and NASA is already on the case, with its Remote Agent Architecture, or RAA.

Discount for UXmatters Readers—Buy Chris Noessel’s book Designing Agentive Technology: AI That Works for People from Rosenfeld Media, using the discount code uxmatters, and save 20% off the retail price.