A fully autonomous self-driving car doesn’t really need a steering wheel, or a rearview mirror, or even windows to get where it’s going. But the first models are still likely to have them. (And not just because such features could be legally required.)

In the coming years and decades, as the public decides how to feel about autonomous cars, the way that self-driving vehicles appear will be arguably as important as how they function. And people, Americans in particular, have clearly defined expectations about what cars ought to look like.

“When we’re looking at new devices, you could make them anything, right? Any shape, any form,” said Robert Brunner, the industrial designer who worked for many years at Apple and now runs his own design studio. “But we’re also trying to get people to relate to and understand the technology.” Self-driving vehicles, he says, should feel inviting and friendly, and should inspire confidence. The way to do this might be to follow Google’s lead, and make driverless cars cute. At the very least, Brunner told me, the ideal self-driving car probably shouldn’t be a “black menacing thing with lots of red lights.”

Engineers and designers will also have to take into account some of the new challenges that accompany driverlessness. For instance: How will self-driving vehicles communicate with human drivers, pedestrians, and bicyclists? The use of blinkers, brake lights, and hazard lights can be automated, surely, but there are many human gestures and cues that are a crucial part of how people navigate the roads—eye contact, the waving of a hand at an intersection—which a machine can’t precisely emulate.

“There are ways of drastically reducing the level of complexity of these systems and making them logical and understandable and reliable,” said Sam Arbesman, the author of Overcomplicated: Technology at the Limits of Comprehension. “The problem is—because of the fact we build the new on top of the old—we aim for these really, really pristine constructions that are built with all these design practices and principles, but things cannot be perfect.”

A more slippery existential problem is that new ways for driverless cars to communicate with pedestrians will only work if people respond to them. But getting people to respond to a new kind of design signal—just getting them to understand it in the first place—is iffy at best.

“People hate ambiguity and unpredictability,” said Chris Rockwell, the CEO and founder of Lextant, a design consulting firm. “I don’t care if it’s your toaster or your car; if you’re confused, you’re not having a great experience. And if it acts in strange or unpredictable ways, it’s not acceptable.”

The trouble is, people are unpredictable. So designing new ways for machines to communicate with them isn’t exactly straightforward. Many ideas for new communications systems have been proposed—driverless cars might feature audible chimes, voice instructions, or text displays to communicate their next moves—but few if any such systems have been tested. “The ideas aren’t the problem; it’s raining ideas,” Rockwell said. “The challenge is really understanding what problem we’re solving. These are human systems, ultimately.”

“From our standpoint, autonomous vehicles and self-driving systems will happen,” he added. “It’s kind of an inevitability. But the challenge won’t be around the technology as much as it will be around the psychology. It’s going to be critical to gain trust—and that trust can be designed into these systems. Trust not only with the passengers, but also the pedestrians outside.”

In an attempt to better understand how pedestrians might respond to self-driving vehicles, roboticists at Duke recently carried an experiment that involved comparing the effectiveness of several different prototypes for vehicle-to-pedestrian communications. (They detailed their findings in a paper that’s now under review for presentation at the Transportation Research Board’s annual meeting.)

The researchers used a van meant to look like a driverless vehicle, and outfitted it with a large display that could feature “walk” and “don’t walk” signals, as well as a numeric display of the speed at which the vehicle was traveling. “The idea was that the participants would use the speedometer to determine whether it was safe to cross,” said Michael Clamann, a roboticist at Duke and one of the lead authors of the paper. “Reading ‘0’ would be the safest, but the objective was to provide a display that would indicate the vehicle was decelerating.”

As it turned out, most pedestrians ignored the new-fangled display, whichever iteration was used. Pedestrians were more likely to rely on “legacy behaviors”—like eyeballing an approaching car’s speed and inferring how quickly to dart across the street—rather than external displays.

“As we start the transition to driverless vehicles, designers need to be aware that people will rely on old habits when interacting with the new technologies,” Clamann said. Part of the problem is that for a display to be useful, the text has to be huge. And even when it’s big enough, a lot of people seemed to ignore it anyway. To be visible from a distance of 100 feet, a single letter would need to be six inches tall and nearly four inches wide. “So a screen designed to display a simple message like ‘safe to cross’ without scrolling horizontally would require a screen at least 47 inches wide,” the researchers wrote. From 200 feet away, the same message would have to be over 100 inches wide; wider than most cars.

“Right now, a pedestrian communicates with a driver. In the future, this communication will be between a human and a machine, which is an area that requires exploration and careful design decisions,” Clamann told me. “We learn from birth how to communicate with other people, but communicating with machines is a very different skill. We need to make sure the displays and signals work as intended before we release them.”

The team at Duke found no significant differences between any of the three different displays they tested, meaning each was “as effective as the current status quo of having no display at all.”

Nearly 5,000 pedestrians were killed by cars in 2014 in the United States alone, according to data from the National Highway Traffic Safety Administration. Driverless cars—with their famously sterling safety records—may be able to reduce those statistics significantly. Still, about half of the pedestrian deaths in the 10-year period ending with 2014 occurred because the pedestrian ran into the road, failed to yield to a vehicle with the right of way, or otherwise crossed the street improperly, the Duke researchers said. Even the best-programmed autonomous cars will be unable to prevent every pedestrian death unless those vehicles can find a way to prompt safer pedestrian behaviors. In other words, with self-driving cars facing a critical test period for the public’s trust, the status quo isn’t going to be good enough.