SAN FRANCISCO — On a rainy December afternoon, Anthony Levandowski sat on a stool near the front of a coffee shop, engrossed in his phone.

He scrolled through pictures on a website that dares users to discern the difference between images of humans generated by artificial intelligence and actual photos of humans. For the uninitiated, it’s a crapshoot. The more time spent staring at the images, the more convincing the fakes appear. Doubt seeps in. The more unnerving the game becomes.

For Levandowski, more skilled in separating what’s real from Silicon Valley creations, spotting the telltale signs of images designed to deceive is an easier task. A small discoloration. A stray hair. He has aced six in a row, and his enthusiasm spills over as he gains momentum.

“Look, look at the freckle on the side,” he exclaims, pointing to an otherwise innocuous blemish.

While he examines images with dermatological care, he’s oblivious to the gawkers and passersby captivated by the presence of the self-driving pioneer whose stints at Google and Uber remain daily news fodder. With his 6-foot-7 frame, thinning hairline and boyish smile, there’s no doubt the real deal sits a few feet from the counter where they await their lattes.

What they make of Levandowski remains a murkier question. In a region known for turning its technology wunderkinds into outsized caricatures of themselves, nobody cuts a longer shadow than the one prosecutors say stole trade secrets from one Big Tech behemoth and brought them to another.

He’s a visionary. As a college student, Levandowski entered a self-driving motorcycle nicknamed GhostRider into a Department of Defense competition designed to stimulate interest in autonomous vehicles. A technical mistake at the start spoiled the undertaking, but the chutzpah necessary to enter a motorcycle in the first place impressed friends and future employers.

He’s an opportunist. In 2011, when he led hardware development for Google’s self-driving car project, he engineered the sale of two companies, 510 Systems and Anthony’s Robots, to Google. Many members of the project were unaware Levandowski owned them.

He’s the personification of Silicon Valley’s more unseemly impulses. He’s flouted regulations in multiple states, allegedly endangered a colleague and a member of the public during self-driving testing, rankled others with an abrasive management style and hopscotched from his job at Google to Uber with at least one confidential document in hand.

Levandowski, now 40, has heard it all.

“There’s a whole narrative about me,” he tells Automotive News. “I got branded as an asshole if I said, ‘Here, how do we build something that succeeds?’ Everything comes down to this: Ideas should be challenged. Conversations about our ideas are healthy. And ultimately, we’re here to make a product.”

Over the past two years, those conversations often occurred inside a courtroom. Waymo, Google’s self-driving car unit, accused Uber of absconding with proprietary knowledge of the company’s essential lidar technology, via Levandowski, and of using those documents to accelerate Uber’s self-driving progress. The dispute was settled in February 2018, with Waymo receiving a stake in Uber worth approximately $245 million at the time.

For Levandowski, legal reverberations continued. The judge presiding over the civil suit recommended him for criminal prosecution. Last summer, Levandowski was indicted on 33 felony counts of theft and attempted theft of trade secrets. One national magazine called the case “the most prominent federal trade-secret criminal prosecution in Silicon Valley history,” one that pitted a brilliant engineer against a multibillion-dollar company and a government emboldened to use new legal tools to deter corporate espionage.

A case some observers characterized as Big Tech’s turn against innovation itself, the United States of America v. Anthony Scott Levandowski reached its crescendo this month. For all the potential implications and precedents, the outcome underscored one point: In Silicon Valley, things may not always be as they appear.

Early tensions

Following the foundational events sponsored by the Defense Advanced Research Project Agency, known as DARPA, that brought Levandowski and other roboticists together in the Nevada and California deserts during the mid-aughts, Google in 2009 launched Project Chauffeur. That was the first iteration of its foray into self-driving developments.

Ambitions were grandiose. Through the mass deployment of self-driving vehicles, Google promised nothing less than a wholesale transportation revolution, one with potential to eradicate more than a million worldwide traffic deaths every year and end the monotony of driving for everyday commuters.

DARPA alums Sebastian Thrun and Chris Urmson served as early leaders of the project, and Levandowski, who had worked on Google’s Street View project since the DARPA challenges, soon joined them as director of the team’s hardware division.

During his stint at Street View, Levandowski started a side company, 510 Systems, that engineered tools that helped turn camera-based images into data. By 2011, the same underlying technology was needed to boost Google’s self-driving car program. Whether through entrepreneurial cunning or genuine interest in helping, Levandowski engineered the sale of 510 Systems to Google for $20 million.

He sold another company, Anthony’s Robotics, to Google as well. At least initially, most Google employees didn’t understand the connections. Levandowski says he was forthright about his motives.

“I didn’t do that to cash out,” he says. “I did it to scale up.”

The acquisitions nonetheless left uneasy feelings on both sides. Once Google acquired 510 Systems, several employees of the startup were dismissed while others slotted into subordinate roles at Google. Levandowski says he regrets not advocating for the 510 Systems staff in a more aggressive manner. But for the time being, Levandowski retained the backing of Google’s senior leadership, and he and his colleagues coexisted.

That changed in the summer of 2013, following a seminal moment for the self-driving car project. Google entrusted some employees with its fledgling autonomous system for road tests, with the caveat that they understood they were responsible for monitoring the system to ensure safety. Quickly, those employees trusted the system so completely that one fell asleep behind the wheel.

Alarmed, Google scrapped its entire approach to introducing autonomous technology. No longer would the company pursue a path that included interactions between human and machine. Humans, the company concluded, must have no role in the driving process. From Levandowski’s perspective, the company learned the wrong lesson.

Levandowski reached the opposite conclusion. From his perspective, the company learned the wrong lesson from the slumbering driver.

“To me, it was perfect because it showed that we had achieved trust,” he says. “You don’t go from a great product to a hard pivot. You say, ‘Let’s fix that. Let’s detect that.’ But they’re going to take this science project that we’re doing to cure one disease, and saying, ‘No, no, let’s pull that until we cure all diseases.’?”

Tensions mounted. Soon after, he says, another project met a discouraging end.

Spawned from a need to find reliable transportation around its hometown, Mountain View, Calif., Google started its own ride-hailing network on a small scale, Levandowski says. It was called Freebird, and it resembled in spirit a San Francisco startup just beginning to upend transportation.

A new path forward

Had the project been allowed to flourish, Levandowski laments, perhaps Freebird would today be as ubiquitous as Uber. Instead, he says, Google leaders viewed the network as a distraction and shelved the project.

For the first time, Levandowski had doubts about his future at the company. He had disagreed with the self-driving-or-bust pivot, been passed over in favor of Urmson when Thrun departed and now seethed as a novel project was shuttered. Frustrations boiled. The feelings were mutual.

“With his manipulations and lack of enthusiasm and commitment to the [self-driving] project, it became clearer and clearer that this was a lost cause,” Urmson told lawyers during an August 2017 deposition at the outset of Waymo’s civil suit against Uber.

The final straw came in the summer of 2015, when Levandowski approached colleagues about joining him at a new startup dedicated to developing self-driving trucks. Word leaked. Urmson was livid. In an email to his bosses sent Aug. 4, 2015, he was blunt: “We need to fire Anthony Levandowski.”

That didn’t happen. But his departure came soon enough. By that December, Levandowski held discussions with then-Uber CEO Travis Kalanick, who would later tell Bloomberg that self-driving technology was “basically existential” for the ride-hailing network’s long-term survival.

On Dec. 11, 2015, Levandowski downloaded approximately 14,000 documents from a password-protected Google server, according to court documents. A month later, he announced his departure and the creation of autonomous trucking startup Ottomotto, known as Otto. Discussions with Kalanick continued, and in August 2016, Uber acquired Otto for $680 million.

Google was furious.

Meanwhile, Levandowski had charged into his first chance to run his own self-driving program, with little regard for obstacles. In May 2016, Nevada’s top transportation official warned Otto officials they did not have permission to conduct a driverless test on the state’s roads. Levandowski greenlighted a test run anyway.

Over the summer, company documents indicate, Otto tested a “self-driving system” in “self-driving mode” on California public roads. That was problematic because state regulations prohibit the testing of autonomous vehicles with a gross weight of 10,001 pounds or more. In December 2016, a public dispute erupted with California Department of Motor Vehicles officials after Uber, which had placed its entire self-driving program in Levandowski’s hands, began testing AVs on public roads without obtaining the necessary permits. The state ultimately revoked the registrations of Uber’s test vehicles.

“He has a willingness to take risks to advance his cause in near total indifference to what others may think of him and his endeavors,” said Randy Miller, a friend of Levandowski since their days at the University of California, Berkeley and member of the GhostRider team. “He’s a really, really smart guy who’s always been different than everyone else. He’s the hardest worker I know, and he’s completely fearless. That hasn’t served him well all of the time.”

Two months after the spat with California regulators, Waymo filed suit against Uber, alleging patent infringement and violations of the Defend Trade Secrets Act, a law passed in 2016 that allowed corporations to file civil suits against competitors. The long fall of Anthony Levandowski had begun.

Trade secrets take center stage

In June 2018, about 14 months before Levandowski’s personal legal battle would begin, prosecutors from the U.S. District Court of Northern California charged six executives from wearables company Fitbit with stealing trade secrets from a former employer. Between the Waymo settlement with Uber and now these cases, engineers throughout Silicon Valley suddenly felt vulnerable to similar accusations.

But when the first of the Fitbit cases reached trial in January 2020, a federal grand jury took less than two hours of deliberation to acquit the defendant, Katherine Mogal, of all charges, ending what one attorney called “a five-year nightmare.” Charges against the remaining defendants were subsequently dropped.

Evidence in the Mogal case centered on documents on her personal computer. In an era when so many people work remotely and outside traditional business hours, jurors could not determine what constituted stolen documents or ones that employees had merely saved on home devices in the normal course of their work.

“Companies think that if they keep information confidential, that it’s a trade secret, and that’s absolutely not true,” said Sharon Sandeen, director of the Intellectual Property Institute and professor at the Mitchell Hamline School of Law in St. Paul, Minn. “Just because you downloaded these documents, that doesn’t mean they’re trade secrets.”

In the civil case between Waymo and Uber, Levandowski’s lawyers argued the approximately 14,000 documents downloaded in December 2015 were part of automatic and regular transfers to his home computer. Waymo engineer Sasha Zbrozek, who designed the server that held the documents, had told lawyers during the course of the investigation that the files were “low value,” according to court filings.

Separately, Zbrozek would say that conducting a “checkout” of documents was common practice and that “it makes me uncomfortable that lawyers are trying to ascribe suspicion to it.”

Sandeen says there are three distinct requirements for information to be considered a trade secret: It must be a secret, not a generally known bit of knowledge; it must have economic value because of its secrecy; companies must engage in efforts to maintain that secrecy.

Further, provisions in trade-secret law allow employees to bring general knowledge with them to new jobs. Silicon Valley overall has thrived for decades with a culture that promotes such movement.

“The problem is, you have a situation like this, with some new technology or initiative, and all the sudden there’s a switch in culture,” Sandeen said. “People start getting more proprietary, and open-innovation norms that existed don’t exist anymore.”

The government’s defeat in the Fitbit trial and subsequent retreat from further cases served as a backdrop to the conclusion of Levandowski’s case. Levandowski reached a plea agreement with the same Northern California prosecutors on March 19. Headlines in publications across the country noted his guilty plea, but they did not capture the substance of the agreement.

He faced 33 felony counts related to theft and attempted theft of trade secrets. Thirty-two of those counts involved accusations related to Google’s internally developed lidar — receiver schematics, motor designs, circuit boards, calibration instructions, etc.

In the end, he pleaded guilty to a charge that had almost nothing to do with the technology Google described as a business differentiator.

Levandowski confessed to downloading 20 documents, including one called “Chauffeur TL weekly updates – Q4 2015.” Stored on Google Drive, the document was an ongoing status report of Google’s self-driving car program, containing quarterly goals and weekly metrics, according to his plea agreement. It contained his team’s OKRs — objectives and key results — and summaries of technical challenges faced by the overall program.

It was not lidar secrets, but it still carries substantial consequences. He admitted to taking the file to use for his own benefit and that he intended to use the weekly update for “the economic benefit of somebody other than the owner” with the knowledge that it would injure Google’s economic position.

“Mr. Levandowski’s guilty plea in a criminal hearing today brings to an end a seminal case for our company and the self-driving industry,” a Waymo spokesperson said. “We are successfully protecting our intellectual property as we build the world’s most experienced driver.”

In a separate hearing two weeks earlier, an arbitration panel had ordered Levandowski to pay Google $179 million in damages, upholding a ruling that he engaged in deceptive and unfair practices when he siphoned Google employees to his trucking startup. He then filed for bankruptcy.

An uncertain future

At the coffee shop in December, Levandowski declined to discuss the specifics of the charges against him. That was partially tactical; the case was ongoing. But beyond the advice of his lawyers, it seemed Levandowski viewed the wringing over his own particular fate as a distraction from a larger concern — a self-driving industry beset with problems and delayed deployment time frames.

Uber shut down its self-driving trucks division in July 2018. Cruise Automation, the General Motors subsidiary, indefinitely delayed its timeline for launching a robotaxi service in 2019. Waymo runs some autonomous vehicles without human safety drivers aboard in a geofenced area of metro Phoenix, but those are a small part of overall operations, of which the company has not disclosed the scope.

Industry lobbyists have blamed a patchwork of regulations that vary from state to state as a contributing factor to the delayed rollout of self-driving technology. But Levandowski defends the regulators he once circumvented.

“Regulators are a scapegoat,” he says. “The real problem is that it’s not ready.”

In 2008, as part of his efforts at 510 Systems, he conducted the first autonomous pizza delivery, using one of the company’s Toyota Priuses, with a police escort, to travel from the San Francisco waterfront, across the Bay Bridge and to Treasure Island. A dozen years later, he laments that not much has improved.

“I’m disappointed,” he says of the industry he helped create. “We’re far out. There’s a fundamental breakthrough that’s needed. This is not that ‘we built an airplane, and we’re figuring out the wing configuration.’ This is, ‘we built an airplane, and we don’t know how to make it fly.’?”

Such a view may not be widely accepted, but it’s no longer an outlier. This month, Toyota Research Institute CEO Gill Pratt told IEEE Spectrum that the gradient of progress has diminished and that it’s getting more difficult to solve remaining challenges. Along similar lines, Stefan Seltz-Axmacher, outgoing CEO of the now-defunct self-driving truck company Starsky Robotics, says supervised machine learning “doesn’t live up to the hype.”

Billions have been spent. Millions of testing miles have been driven, the value of which is increasingly scrutinized. Thousands of employees have been devoted to readying self-driving technology. Fully autonomous vehicles remain a bold and ambitious project, but in the meantime, there’s an intermediate step with lifesaving potential. Instead of waiting for self-driving cars, traffic deaths can be addressed today with advanced driver-assist systems.

In July 2018, Levandowski co- founded Pronto.ai, a company developing driver-assist systems tailored for the trucking industry. He stepped down as CEO when the trade-secret indictment arrived. But over the past year, his faith in driver-assist systems has only grown more resolute.

Levandowski has been accused of disregarding safety over the years. Most notably, a New Yorker article alleges a 2011 incident in which he was behind the wheel of a Google self-driving test vehicle, which boxed in a Toyota Camry. The Camry subsequently spun out of control along a California roadway, and in swerving to avoid it, the article says, Levandowski injured a colleague who was his passenger.

He bristles at the idea he’s cut safety corners, arguing he’s doing more to save lives today than those focused solely on a fully driverless tomorrow.

“Everyone who says they’re building self-driving cars for safety reasons, they’re whitewashing their efforts,” he says. “If you replace a driver, that’s a moneymaking opportunity. If you think that’s happening next week, go work on that. If you think it’s not happening for another 10 years, look at these incremental systems that are saving lives today.”

No limits

Frustrations with the pace of self-driving developments aside, Levandowski remains bullish on the long-term prospects of artificial intelligence.

When Uber fired him in May 2017, he turned his attention toward that future, diving into his self-appointed role as dean of the self-created Way of the Future, a church devoted to the worship of AI. Outsiders thought it was a prank or perhaps a means to shelter his fortune. But he says his intentions are genuine.

“AI has amazing potential that will surpass humans,” he says. “I don’t think there’s a limit to how good it will be, and I want that to be beneficial for society. The whole reason it’s a church is that it strips away financial motivations. It turns a primary motivation of ‘How can we make money?’ to ‘How can we use this to solve problems and potentially take care of the planet?’?”

If exploring those far-reaching impacts of AI are part of Levandowski’s future, he cannot yet escape the present. Sentencing is scheduled for Aug. 4. Federal guidelines suggest he will spend 24 to 30 months in prison.

“Mr. Levandowski is a young man with enormous talents and much to contribute to the fast-moving world of AI and AV,” Miles Ehrlich, his attorney, said in a written statement. “We hope this plea will allow him to move on with his life and focus his energies where they matter most.”

Recently divorced, bankrupt and an admitted thief, Levandowski has seen much of the legacy he created now in tatters. One small part remains intact: GhostRider, the self-driving motorcycle that launched his career, resides in the Smithsonian Institution.

Miller, his friend, laments all that has been squandered. But he’s reassured by the beliefs it’s only served to invigorate Levandowski and that a long road ahead remains.

“Certainly, he made a mistake with the files that has cost him a few years of his life already and will probably cost him a few more,” he said. “In a certain way, that’s humbling. At the same time, he is who he is. He’s not going to get out and take a job at a coding desk. Day one, he’s going to be right back at it. Knowing him, I don’t think this is the final chapter at all.”