Ethicists were hired to save tech’s soul. Will anyone let them?
Firms are adding ethical thinking to their processes, but ethical outcomes are optional.
Fifty-two floors below the top of Salesforce Tower, I meet Paula Goldman in a glass-paneled conference room where the words EQUALITY OFFICE are spelled out on a patchwork bunting banner, the kind of decoration you might buy for a child's birthday party.
Goldman has a master's degree from Princeton and a Ph.D. from Harvard, where she studied how controversial ideas become mainstream. She arrived at Salesforce just over a year ago to become its first-ever Chief Ethical and Humane Use Officer, taking on an unprecedented and decidedly ambiguous title that was created specifically for her unprecedented, ambiguous, yet highly specific job: see to it that Salesforce makes the world better, not worse.
Get what matters in tech, in your inbox every morning. Sign up for Source Code.
"I think we're at a moment in the industry where we're at this inflection point," Goldman tells me. "I think the tech industry was here before, with security in the '80s. All of a sudden there were viruses and worms, and there needed to be a whole new way of thinking about it and dealing with it. And you saw a security industry grow up after that. And now it's just standard protocol. You wouldn't ship a major product without red-teaming it or making sure the right security safeguards are in it."
"I think we're at a similar moment with ethics," she says. "It requires not only having a set of tools by which to do the work, but also a set of norms, that it's important. So how do you scale those norms?"
I ask her how those norms are decided in the first place.
"In some sense, it's the billion-dollar question," she says. "All of these issues are extremely complicated, and there's very few of them where the answer is just absolutely clear. Right? A lot of it does come down to, which values are you holding up highest in your calculus?"
In the wake of the Cambridge Analytica scandal, employee walkouts, and other political and privacy incidents, tech companies faced a wave of calls to hire what researchers at the Data & Society Research Institute call "ethics owners," people responsible for operationalizing "the ancient, domain-jumping, and irresolvable debates about human values that underlie ethical inquiry" in practical and demonstrable ways.
Salesforce hired Goldman away from the Omidyar Network as the culmination of a seven-month crisis-management process that came after Salesforce employees protested the company's involvement in the Trump administration's immigration work. Other companies, responding to their own respective crises and concerns, have hired a small cadre of similar professionals — philosophers, policy experts, linguists and artists — all to make sure that when they promise not to be evil, they actually have a coherent idea of what that entails.
So then what happened?
While some tech firms have taken concrete steps to insert ethical thinking into their processes, Catherine Miller, interim CEO of the ethical consultancy Doteveryone, says there's also been a lot of "flapping round" the subject.
Critics dismiss it as "ethics-washing," the practice of merely kowtowing in the direction of moral values in order to stave off government regulation and media criticism. The term belongs to the growing lexicon around technology ethics, or "tethics," an abbreviation that began as satire on the TV show "Silicon Valley," but has since crossed over into occasionally earnest usage.
"If you don't apply this stuff in actual practices and in your incentive structures, if you don't have review processes, well, then, it becomes like moral vaporware," says Shannon Vallor, a philosopher of technology at the Markkula Center for Applied Ethics at Santa Clara University. "It's something that you've promised and you meant to deliver, but it never actually arrived."
Google, infamously, created an AI Council and then, in April of last year, disbanded it after employees protested the inclusion of an anti-LGBTQ advocate. Today, Google's approach to ethics includes the use of "Model Cards" that aim to explain its AI.
"That's not anything that has any teeth," says Michael Brent, a data ethicist at Enigma and a philosophy professor at the University of Denver. "That's just like, 'Here's a really beautiful card.'"
The company has made more-substantial efforts: Vallor just completed a tour of duty at Google, where she taught ethics seminars to engineers and helped the company implement governance structures for product development. "When I talk about ethics in organizational settings, the way I often present it is that it's the body of moral knowledge and moral skill that helps people and organizations meet their responsibilities to others," Vallor tells me.
More than 100 Google employees have attended ethics trainings developed at the Markkula center. The company also developed a fairness module as part of its Machine Learning Crash Course, and updates its list of "responsible AI practices" quarterly. "The vast majority of the people who make up these companies want to build products that are good for people," Vallor says. "They really don't want to break democracy, and they really don't want to create threats to human welfare, and they really don't want to decrease literacy and awareness of reality in society. They want to make things they're proud of. So am I going to do what I can to help them achieve that? Yes."
The Markkula center, where Vallor works, is named after Mike Markkula Jr., the "unknown" Apple co-founder who, in 1986, gave the center a starting seed grant in the same manner that he gave the young Steve Jobs an initial loan. He never wanted his name to be on the building — that was a surprise, a token of gratitude, from the university.
Markkula has retreated to living a quiet life, working from his sprawling gated estate in Woodside. These days, he doesn't have much contact with the company he started — "only when I have something go wrong with my computer," he tells me. But when he arrived at the Santa Clara campus for an orientation with his daughter in the mid-'80s, he was Apple's chairman, and he was worried about the way things were going in the Valley. "It was clear to us both, Linda [his wife] and I, that there were quite a few people who were in decision-making positions who just didn't have ethics on their radar screen," he says. "It's not that they were unethical, they just didn't have any tools to work with."
At Apple, he spent a year drafting the company's "Apple Values" and composed its famous marketing philosophy ("Empathy, Focus, Impute."). He says that there were many moments, starting out, when he had to make hard ethical choices, but "fortunately, I was the guy running the company, so I could do whatever I wanted."
"I'd have a heck of a time running Apple today, running Google today," he says. "I would do a lot of things differently, and some of them would have to do with philosophy, ethics, and some of them would have to do with what our vision of the world looks like 20 years out."
Advanced issue found
The Markkula Center for Applied Ethics is one of the most prominent voices in tech's ethical awakening. On its website, it offers a compendium of materials on technology ethics, including a toolkit ("Tool 6: Think About the Terrible People"), a list of "best ethical practices" ("No. 2: Highlight the Human Lives and Interests behind the Technology"), and an app ("Ethics: There's an App for That!" reads a flier posted at the entrance).
Every one of these tools is an attempt to operationalize the basic tenets of moral philosophy in a way that engineers can quickly understand and implement. But Don Heider, the Markkula center's executive director, is quick to acknowledge that it's an uphill fight. "I'd say the rank-and-file is more open to it than the C-suite," he says.
Even at Salesforce, practitioners like Yoav Schlesinger, the company's principal of ethical AI Practice, worry about imposing an "ethics tax" on their teams — an ethical requirement that might call for "heavy lifting" and would slow down their process.
Under Goldman's direction, the company has rolled out a set of tools and processes to help Salesforce employees and its clients "stretch their moral imagination, effectively," as Schlesinger puts it. The company offers an educational module that trains developers in how to build "trusted AI" and holds employee focus groups on ethical questions. "Our essential task is not teaching ethics like teaching deontological versus Kantian or utilitarian approaches to ethics — that's probably not what our engineers need," he says. "What people need is training in ethical risk spotting: How do you identify a risk, and what do you do about it when you see it from a process perspective, not from a moral perspective."
Goldman agrees: "It's not not from a moral perspective," she says. "It's just more that we're focused on the practical, 'what do you do about it,' than we are about the theory."
The company has also created explainability features, confidential hotlines, and protected fields that warn Salesforce clients that things like ZIP code data is highly correlated with race. They have refined their acceptable use policy to prevent their e-commerce platform from being used to sell a wide variety of firearms and to prevent their AI from being used to make the final call in legal decision-making. The Ethical and Humane Use team holds office hours where employees can drop by to ask questions. They have also begun to make their teams participate in an exercise called "consequence scanning," developed by researchers at Doteveryone.
Teams are asked to answer three questions: "What are the intended and unintended consequences of this product or feature?" "What are the positive consequences we want to focus on?" "What are the consequences we want to mitigate?" The whole process is designed to fit into Agile software development, to be as minimally intrusive as possible. Like most ethical interventions currently in use, it's not really supposed to slow things down, or change how business operates. Beware the "ethics tax."
"Nobody touches running code," says Subbu Vincent, a former software engineer and now the director of media ethics at the Markkula center. Engineers, he says, "always want to layer their new effort on top of this system of software that's handling billions of users. If they don't, it could end their career."
And therein lies the problem. These approaches, while well-intentioned and potentially impactful, tend to suggest that ethics is something that can be quantified, that living a more ethical life is merely a matter of sitting through the right number of trainings and exercises.
"What's notable is that the solutions that are coming out are using the language of, 'hey, we'll fit within the things you're already familiar with,'" says Jacob Metcalf, a researcher at Data & Society. "They're not saying, 'hey, maybe don't be so voracious about user data, maybe you don't need to grow to scale using these exploitative methods.' They're not forcing a change in the diversity of who is in the room."
With colleagues danah boyd and Emmanuel Moss, Metcalf recently surveyed a group of 17 "ethics owners" at different companies. One engineer told them that people in tech "are not yet moved by ethics." An executive told them that market pressures got in the way: "If we play by these rules that kind of don't even exist, then we're at a disadvantage," the executive said. The "ethics owners" they spoke to were all experimenting with different approaches to solving problems, but often tried to push for simple, practical solutions adopted from other fields, like checklists and educational modules.
"By framing ethics as a difficult but tractable technological problem amenable to familiar approaches, ethics owners are able to enroll the technical and managerial experts they feel they need as full participants in the project of 'doing ethics,'" the researchers write. "However, building a solution in the same mold that was used to build the problem is itself a form of failure."
If and when ethics does "arrive" at a company, it often does so quietly, and ideally invisibly. "Success is bad stuff not happening, and that's a very hard thing to measure," says Miller, the acting CEO of Doteveryone. In a recent survey of UK tech workers, Miller and her team found that 28% had seen decisions made about a technology that they believed would have a negative effect upon people or society. Among them, one in five went on to leave their companies as a result.
At Enigma, a small business data and intelligence startup in New York City, all new hires must gather for a series of talks with Brent, the philosophy professor working as the company's first data ethics officer.
At these gatherings, Brent opens his slides and says, "All right, everyone. Now we're going to do an hourlong introduction to the European-based, massively influential moral theories that have been suggested in the past 2,400 years. We have an hour to do it."
The idea is that starting at the beginning is the only way to figure out the way forward, to come up with new answers. "The theories that we're looking at don't obviously have any direct application — yet — to these new issues. So it's up to us, right? We're the ones who have to figure it out," he says.
The engineers he works with — "25-year-olds, fresh out of grad school, they're young, they're fucking brilliant" — inevitably ask him whether all this talk about morals and ethics isn't just subjective, in the eye of the beholder. They come to his office and ask him to explain. "By the end," he says, "they realize that it's not mere subjectivism, but there are also no objective answers, and to be comfortable with that gray area."
Brent met Enigma's founders, Marc DaCosta and Hicham Oudghiri, in a philosophy class at Columbia when he was studying for his doctorate. They became fast friends, and the founders later invited him to apply to join their company. Soon after he came on board, a data scientist at Enigma called him over to look at his screen. It was a list of names of individuals and their personal data. "I was like, whoa, wow, OK. So there we go. What are you going to do with this data? Where did it come from? How are we keeping it secure?" The engineer hadn't realized that he would be able to access identifying information. "I'm like, OK, let's talk about how you can use it properly."
The fact that Brent, and many others like him, are even in the room to ask those questions is a meaningful shift. Trained philosophers are consulting with companies and co-authoring reports on what it means to act ethically while building unpredictable technology in a world full of unknowns.
Another humanities Ph.D. who works at a big tech company tells me that the interventions he ends up making on his team often involve having his colleagues simply do less of what they're doing, and articulating ideas in a sharper, more precise manner. "It's hard, because to actually make these products and do our jobs, all the machine learning is built around data. You can't really avoid that for now," he tells me. "There are a lot of stops in place … It's basically really hard to do our job right now."
It's the Wild West, and anyone who wants to say they are ethicists can just say it. It's nonsense. — Reid Blackman
But for every professional entering the field, there are just as many — and probably more — players whom Reid Blackman, a philosophy professor turned ethics consultant, calls "enthusiastic amateurs." "You have engineers who care, and who somehow confuse their caring with an expertise. So then they bill themselves as, for instance, AI ethicists, and they are most certainly not ethicists. I see the things that they write, and I hear the things that they say, and they are the kinds of things that students in my introduction to ethics class would say, and I would have to correct them on," he tells me. "They're reinventing the wheel, talking about principles or whatever. It's the Wild West, and anyone who wants to say they are ethicists can just say it. It's nonsense."
The result is that to wade into this field is to encounter a veritable tower of Babel. A recent study of 84 AI ethics guidelines from around the world found that "no single ethical principle appeared to be common to the entire corpus of documents, although there is an emerging convergence around the following principles: transparency, justice and fairness, non-maleficence, responsibility and privacy."
This is also, in part, a geopolitical problem: Every nation wants its code of AI ethics to be the one that wins. (See, for instance, the White House's recent explication of "AI with American Values.") "All of these AI ethics principles are brought out to support a particular worldview, and a particular idea of what 'good' is," Miller says.
"It is early days," says Goldman, so it's not necessarily surprising that people would be using different vocabularies to talk about ethics. "That is true of how fields get created. I'm sure it's true of how security got created."
I asked her and Schlesinger what would happen if a Salesforce client decided to ignore all of the ethical warnings they had worked into the system and use data that might lead to biased results. Goldman paused. The thing is, ethics at this point is still something you can opt out of. Schlesinger explains that Salesforce's system is right now designed to give the customer "the opportunity to decide whether or not they want to use the code," he says. "We really believe that customers should be empowered with all the information to make their decisions, but that their use cases are going to be specific to them and their goals."
Likewise, at Enigma, the company's co-founders and leadership team can choose not to listen to Brent's suggestions. "I'm going to say, OK, here's what I think are the ethical risks of developing this kind of product," he says. "You guys are the moneymakers, so you can decide what level of risk you're comfortable with, as a business proposition."