I had the pleasure of doing a podcast interview with Steve Hogan of Drone Law Today last week, and we had a fascinating, wide-ranging discussion on the future of artificial intelligence. The podcast episode is now available on Drone Law Today‘s website. If you haven’t heard the podcast before, check it out and subscribe–there are not a lot of regularly updated resources out there for people interested in the intersection of law and emerging technologies, and Steve’s podcast is a great one.
As background, the new California regulations specify in greater detail than before what type of technology counts as “autonomous” in the context of cars. Specifically, it says that an “autonomous vehicle” must qualify as Level 3, Level 4, or Level 5 under the Society of Automotive Engineers (SAE) framework. That framework classifies vehicles along a 0 to 5 scale, with 5 being fully autonomous and 0 being a standard, fully human-controlled car.
The kicker came in the very last section of the draft regulations:
A short round up of recent news of interest to Law and AI.
In the Financial Times, John Thornhill writes on “the darker side of AI if left unmanaged: the impact on jobs, inequality ethics, privacy and democratic expression.” Thornhill takes several proverbial pages from the Stanford 100-year study on AI, but does not ultimately offer his view of what effective AI “management” might look like.
Patrick Tucker writes in Defense One that a survey funded by the Future of Life Institute found “that the U.S. military more commonly uses AI not to help but to replace human operators, and, increasingly, human decision making.” In the process, he gives voice to the fears held by many people (well, at least by me) of how an autonomous weapons arms race might play out:
Today, the United States continues to affirm that it isn’t interested in removing the human decision-maker from “the loop” in offensive operations like drone strikes (at least not completely). That moral stand might begin to look like a strategic disadvantage against an adversary that can fire much faster, conduct more operations, hit more targets in a smaller amount of time by removing the human from loop.
Microsoft CEO Satya Nadella sat down for an interview with Dave Gershgorn of Quartz. Among other things, Nadella discusses the lessons Microsoft learned from Tay the Racist Chatbot–namely the need to build “resiliency” into learning AI systems to protect them from threats that might cause them to “learn” bad things. In the case of Tay, Microsoft failed to make the chatbot resilient to trolls, with results that were at once amusing and troubling.
Well, by far the biggest AI news story to hit the papers this week was the announcement that a collection of tech industry heavyweights–Microsoft, IBM, Amazon, Facebook, and Google–are joining forces to form a “Partnership on AI”:
The group’s goal is to create the first industry-led consortium that would also include academic and nonprofit researchers, leading the effort to essentially ensure AI’s trustworthiness: driving research toward technologies that are ethical, secure and reliable — that help rather than hurt — while also helping to diffuse fears and misperceptions about it.
“We plan to discuss, we plan to publish, we plan to also potentially sponsor some research projects that dive into specific issues,” Banavar says, “but foremost, this is a platform for open discussion across industry.”
There’s no question this is welcome news. Each of the five companies who formed this group had been part of the “AI arms race” that has played out over the past few years, when major tech companies have invested massive amounts of money in expanding their AI research, both by acquiring other companies and by recruiting talent. To a mostly-outside observer such as myself, it seemed for a time like the arms race was becoming an end unto itself–companies were making huge investments in AI without thinking about the long-term implications of AI development. The Partnership is a good sign that the titans of tech are, indeed, seeing the bigger picture.
The most interesting story that came up during Law and AI’s little hiatus came from decidedly outside the usual topics covered here–the world of beauty pageants. Well, sort of:
An online beauty contest called Beauty.ai, run by Youth Laboratories . . . ., solicited 600,000 entries by saying they would be graded by artificial intelligence. The algorithm would look at wrinkles, face symmetry, amount of pimples and blemishes, race, and perceived age.
Sounds harmless enough, aside from the whole “we’re teaching computers to objectify women” aspect. But the results of this contest carry some troubling implications.
Of the 44 winners in the pageant, 36 (or 82%) were white. In other words, white people were disproportionately represented among the pageant’s “winners.” This couldn’t help but remind me of discrimination law in the legal world. The algorithm’s beauty assessments had what lawyers would recognize as a disparate impact–that is, despite the fact that the algorithm seemed objective and non-discriminatory at first glance, it ultimately favored whites at the expense of other racial groups.
The concept of disparate impact is best known in employment law and in college admissions, where a company or college can be liable for discrimination if its policies have a disproportionate negative impact on protected groups, even if the people who came up with the policy had no discriminatory intent. For example, a hypothetical engineering company might select which applicants to interview for a set of open job positions by coming up with a formula that awards 1 point to an applicant with a college degree in engineering, 3 years for a Master’s degree, and 6 points for a doctorate, and additional points for certain prestigious fellowships. Facially, this system appears neutral in terms of race, gender, and socioeconomic status. But in its outcomes, it may (and probably would) end up having a disparate impact if the components of the test score are things that wealthy white men are disproportionately more likely to have due to their social and economic advantages.
The easiest way to get around this problem might be to use a quota–i.e., set aside a certain proportion of the positions for applicants from underserved minority groups and then apply the ‘objective’ test to rate applicants within each group. But such overt quotas are also illegal (according to the Supreme Court) because they constitute disparate treatment. What about awarding “bonus points” under the objective test to people from disadvantaged groups? Well, that would also be disparate treatment. Certainly, nothing prevents an employer from using race as, to borrow a phrase from Equal Protection law, a subjective “plus factor” to help ensure diversity. But you can’t assign a specific number related to the race or gender of applicants. The bottom line is that the law likes to keep assessments very subjective when they involve sensitive personal characteristics such as race and gender.
Which brings us back to AI. You can have an algorithm that approximates or simulates a subjective assessment, but you still have to find a way to program that assessment into the AI–which means reducing the subjective assessment to an objective and concrete form. It would be difficult-to-impossible to program a truly subjective set of criteria into an AI system because a subjective algorithm is almost a contradiction in terms.
Fortunately for Beauty.ai, it can probably solve its particular “disparate impact” problem without having the algorithm discriminate based on race. The reason why Beauty.ai generated a disproportionate number of white winners is that the data sets (i.e. images of people) that were used to build the AI’s ‘objective’ algorithm for assessing beauty consisted primarily of images of white people.
As a result, the algorithm’s accuracy dropped when it runs into the images of people who don’t fit the patterns in the data set that was used to prime the algorithm. To fix that, the humans just need to include a more diverse data set–and since humans are doing that bit, the process of choosing who is included in the original data set could be subjective, even if the algorithm that uses the data set cannot be.
For various reasons, however, it would be difficult to replicate that process in the contexts of employment, college admissions, and other socially and economically vital spheres. I’ll be exploring this topic in greater detail in a forthcoming law practice article that should be appearing this winter. Stay tuned!
A brief item that I could not resist leaving a quick comment on. The Atlantic posted a fascinating story last week on a machine learning program that could help make more accurate psychiatric diagnoses. The system currently in place is a “schizophrenia screener” that analyzes primary care patients’ speech patterns for some of the tell-tale verbal ‘tics’ that can be a predictor of psychosis. For now, as the author points out, there are many weaknesses with widespread deployment of such a system because there are so many cultural, ethnic, and other differences in speech and behavior that could throw the system off. But still, the prospect of an AI system playing a role in determining whether a person has a mental disorder raises some intriguing questions.
The lawyer in me immediately thought “could the Tarasoff rule apply to AI systems?” For those of you who are normal, well-adjusted human beings (i.e., not lawyers), Tarasoff was a case where the California Supreme Court held that a psychiatrist could be held liable if the psychiatrist knows that a patient under his or her care poses a physical danger to someone and fails to take protective measures (e.g., by calling the police or warning the potential victim(s)).
Now granted, predicting violence is probably a much more difficult task than determining whether someone has a specific mental disorder. But it’s certainly not out of the realm of possibility that a psychiatric AI system could be designed that analyzes a patient’s history, the tone and content of a patient’s speech, etc, and comes up with a probability that the patient will commit a violent act in the near future.
Let’s say that such a violence-predicting AI system is designed for use in medical and psychiatric settings. The system is programmed to report to a psychiatrist when it determines that the probability of violence is above a certain threshold–say 40%. The designers set up the system so that once makes its report, its job is done; it’s ultimately up to the psychiatrist to determine whether a real threat of violence exists and, if so, what protective measures to take.
But let’s say that the AI system determines that there is a 95% probability of violence, and that studies have shown that the system does better than even experienced human psychiatrists in predicting violence. Should the system still be designed so it can do nothing except report the probability of violence to a psychiatrist, despite the risk that the psychiatrist may not take appropriate action? Or should AI systems have a freestanding Tarasoff-like duty to warn police?
Given that psychiatry is one of the more subjective fields of medicine, it will be interesting to see how the integration of AI in the mental health sector plays out. If AI systems prove to be, on average, better than humans at making psychiatric diagnoses and assessing risks of violence, would we still want a human psychiatrist to have the final say–even though it might mean worse decisions on balance? I have a feeling we’ll have to confront that question some day.
The last post in this series on “Digital Analogues”–which explores the various areas of law that courts could use as a model for liability when AI systems cause harm–examined animal liability law. Under traditional animal liability law, the owner of a “wild” animal is strictly liable for any injury or damage caused by that animal. For domesticated animals, however, an owner is only liable if that particular animal had shown dangerous tendencies and the owner failed to take adequate precautions.
So what lessons might animal liability law offer for AI? Well, if we believe that AI systems are inherently risky (or if we just want to be extra cautious), we could treat all AI systems like “wild” animals and hold their owners strictly liable for harms that they cause. That would certainly encourage safety precautions, but it might also stifle innovation. Such a blanket rule would seem particularly unfair for AI systems whose functions are so narrow that they do not present much risk to anyone. It would seem somewhat silly to impose a blanket rule that treats AlphaGo as if it is just as dangerous as an autonomous weapon system.
The last two entries in this series focused on the possibility of treating AI systems like “persons” in their own right. As with corporations, these posts suggested, legal systems could develop a doctrine of artificial “personhood” for AI, through which AI systems would be given some of the legal rights and responsibilities that human beings have. Of course, treating AI systems like people in the eyes of the law will be a bridge too far for many people both inside the legal world and in the public at large. (If you doubt that, consider that corporate personhood is a concept that goes back to the Roman Empire’s legal system, and it still is highly controversial)
In the short-to-medium term, it is far more likely that instead of focusing on what rights and responsibilities an AI system should have, legal systems will instead focus on the responsibilities of the humans who have possession or control of such systems. From that perspective, the legal treatment of animals provides an interesting model.
As discussed in a prior post, the White House Office of Science and Technology Policy (OSTP) published a request for information (RFI) on AI back in June. IBM released a response that was the subject of a very positive write-up on TechCrunch. As the TechCrunch piece correctly notes, most of IBM’s responses were very informative and interesting. They nicely summarize many of the key topics and concerns that are brought up regularly in the conferences I’ve attended.
But their coverage of the legal and governance implications of AI was disappointing. Perhaps IBM was just being cautious because they don’t want to say anything that could invite closer government regulation or draw the attention of plaintiff’s lawyers, but their write-up on the subject was quite vague and somewhat off-topic.
I was a guest on today’s edition of AirTalk, a talk program on the Los Angeles area NPR affiliate KPCC. The topic was a new initiative by OpenAI to recruit a team of AI cops to watch out for potentially harmful AI systems. A recording of the interview is available here.