Technology’s fine – I definitely like texting, and some of the shows on Netflix are tolerable – but the field’s got some serious kinks to work out. Some of these are hardware-related: when, for instance, will quantum computing become practical? Others are of more immediate concern. Is there some way to stop latently homicidal weirdos from getting radicalised online? Can social networks be tweaked in such a way as to not nearly guarantee the outbreak of the second Civil War? As AI advances and proliferates, how can we stop it from perpetuating, or worsening, injustice and discrimination?
For this Giz Asks, we’ve assembled a wide-ranging panel – of futurists, engineers, anthropologists, and experts in privacy and AI – to address these and many other hurdles.
Professor of Electrical Engineering and Computer Science and Director of the Computer Science and Artificial Intelligence Laboratory (CSAIL) at MIT
Here are some broad societal impact challenges for AI. There are so many important and exciting challenges in front of us – I include a few I have been thinking about:
1) virtual 1-1 student-teacher ratios for all children – this will enable personalised education and growth for all children
2) individualised healthcare – this will deliver medical attention to patients that is customised to their own bodies
3) reversing climate change – this will take us beyond mapping climate change into identifying ways to repair the damage; one example is to reverse engineer photosynthesis and incorporate such processes into smart cities to ameliorate pollution
4) interspecies communication – this will enable us to understand and communicate with members of other species, for example to understand what whales are communicating through their song, etc
5) intelligent clothing that will monitor our bodies (1) to ensure we live well and (2) to detect the emergence of a disease before the disease happens
And here are some technical challenges:
1) interpretability and explainability of machine learning systems
2) robustness of machine learning systems
3) learning from small data
4) symbolic decision making with provable guarantees
7) machine learning with provable guarantees
8) unsupervised machine learning
9) new models of machine learning that are closer to nature
“...interpretability and explainability of machine learning systems... robustness of machine learning systems... learning from small data...”
Anthropologist and Research Director at the Centre National de la Recherche Scientifique, Institut Jean Nicod, Paris; Co-Founder of the Centre for the Resolution of Intractable Conflict, University of Oxford, and author of Talking to the Enemy: Faith, Brotherhood and the (Un)Making of Terrorists
How to tell the difference between real vs fake, and between good vs harmful so that we can prevent harmful fake (malign) activity and promote what is real and good?
Malign social media ecologies (hate speech, disinformation, polarising and radicalising campaigns, etc.) have both bottom-up and top-down aspects, each of which is difficult to deal but together stump most counter efforts. These problems are severely compounded by exploitation of cognitive biases (e.g., their tendency to believe in messages that conform to one’s prior believes and to disbelieve messages that don’t), and also by exploitation of cultural belief systems (e.g., gaining trust, as in the West, based on accuracy, objectivity, validation and competence vs. gaining trust, as in most of the rest of the world, based on respect, recognition, honour, and dignity) and preferences (e.g., values associated with family, communitarian, nationalist, traditional mores vs. universal, multicultural, consensual, progressive values).
Malign campaigns exploit psychological biases and political vulnerabilities in the socio-cultural landscape of nations, and among transnational and substate actors, which has already led to new ways of resisting, reinforcing and remaking political authority and alliances. Such campaigns also can be powerful force multipliers for kinetic warfare and affect economies. Although pioneered by state actors, disinformation tools are now readily available to anyone or any group with internet access to deploy at low cost. This “democratisation” of influence operations, coupled with democracies’ vulnerabilities owing to political tolerance and free speech, requires our societies to create new forms of resilience as well as deterrence. This means that a significant portion of malign campaigns involve self-organising “bottom-up” phenomena that self-repair. Policing and banning on any single platform (Twitter, Facebook, Instagram, VKontake, etc.) can be downright counterproductive, with users going to “back doors” even being banned, jumping between countries, continents and languages, and eventually producing global “dark pools,” in which illicit and malign online behaviours will flourish.
Because large clusters that carry hate speech or disinformation arise from small, organic clusters, it follows that “large clusters can hence be reduced by first banning small clusters.” In addition, random banning of a small fraction of the entire user population (say, 10 per cent) would serve “the dual role of lowering the risk of banning many from the same cluster, and inciting a large crowd.” But if, indeed, States and criminal organisations with deep offline presence can create small clusters almost at will, then the problem becomes not one of simply banning small clusters or a small fraction of randomly chosen individuals. Rather, the key involves identifying small clusters that initiate a viral cascade propagating hate or malign influence. Information cascades follow a heavy-tailed distribution, with large-scale information cascades relatively rare (only 2 per cent > 100 re-shares), with 50 per cent of shares in a cascade occurring within an hour; so the problem is to find an appropriate strategy i to identify an incipient malign viral cascade and apply counter measures well within the first hour
There is also a layering strategy evident in State-sponsored and criminally-organised illicit online networks. Layering is a technique where links to disinformation sources are embedded in popular blogs, forums and websites of activists (e.g., environment, guns, healthcare, immigration, etc.) and enthusiasts (e.g., automobiles, music, sports, food and drink, etc.). These layering-networks, masquerading as alternative news and media sources, regularly seek bitcoin donations. Their block chains show contributions made by anonymous donors in orders of tens of thousands of dollars at a time, and hundreds of thousands of dollars over time. We find that these layering-networks often form clusters linking to the same Google Ad accounts, earning advertising dollars for their owners and operators. Social media and advertising companies often have difficulty identifying account owners linked with illicit and malign activity, in part because they often appear to be “organic” and regularly pass messages containing “a kernel of truth.” How, then, to detect layering-networks (Breitbart, One America News Network, etc.), symbols (logos, flags), faces (politicians, leaders), suspicious objects (weapons), hate speech and anti-democracy framing as “suspicious”?
Finally, knowledge of psychology and cultural belief systems are needed to train the data that technology uses to mine, monitor, and manipulate information. Overcoming malign social media campaigns ultimately relies on human appraisal of strategic aspects, such as importance of “core values” and the stakes at play (political, social, economic), and relative strengths of players in those stakes. The critical role of social science goes beyond the expertise of engineers, analysts, and data scientists that platforms like Twitter, Instagram, and Facebook use to moderate propaganda, disinformation, and hateful content.
Yet, an acute problem concerns overwhelming evidence from cognitive and social psychology and anthropology, that truth and evidence – no matter how logically consistent or factually correct – do not sway public opinion or popular allegiance as much as appeals to basic cognitive biases that confirm deep beliefs and core cultural values. Indeed, many so-called “biases” used in argument do not reflect sub-optimal or deficient reasoning but rather suggest their efficient (even optimal) use for persuasion – an evolutionarily privileged form of reasoning to socially recruit others to one’s circle of beliefs for cooperation and mutual defence. Thus, to combat false or faulty reasoning – as in noxious messaging – it’s not enough to target an argument’s empirical and logical deficiencies versus a counterargument’s logical and empirical coherence. Moreover, recent evidence suggests that warning about misinformation has little effect (e.g., despite advanced warning, “yes” voters are more likely than “no” voters to “remember” a fabricated scandal about a vote “no” campaign, and “no” voters are more likely to “remember” a fabricated scandal about a vote “yes” campaign). Evidence is also mounting that value-driven, morally focused information in general, and social media in particular not only drives readiness to believe, but also concerted actions for beliefs.
One counter strategy involves compromising one’s own truth and honesty, and ultimately moral legitimacy, in a disinformation arms race. Another is to remain true to the democratic values upon which our society is based (in principle if not practice), never denying or contradicting them, or threatening to impose them on others.
But how to consistently expose misleading, false, and malicious information while advancing truthful, evidence-based information that never contradicts our core values or threatens the core values of others (to the extent tolerable)? How to encourage people to exit echo chambers of the like-minded to engage in a free and open public deliberation on ideas that challenge preconceived or fed attitudes, a broader awareness of what is on offer and susceptibility to alternatives may be gained however initially strong one’s preconception or fed history?
“How to consistently expose misleading, false, and malicious information while advancing truthful, evidence-based information that never contradicts our core values or threatens the core values of others (to the extent tolerable)?”
Professor, Mechanical Engineering, MIT, whose research focuses on quantum information and control theory
The two greatest technological challenges of our current time are
(a) good mobile phone service, and
(b) a battery with the energy density of extra virgin olive oil
I need say no more about (a). For (b) I could have used diesel fuel instead of olive oil (they have similar energy densities), but I like the thought of giving my computer a squirt of extra virgin olive oil every time it runs out of juice.
Since you are also interested in quantum computing I’ll comment there too.
Quantum computing is at a particularly exciting and maybe scary moment. If we can build large-scale quantum computers, they would be highly useful for a variety of problems, from code-breaking (Shor’s algorithm), to drug discovery (quantum simulation), to machine learning (quantum computers could find patterns in data that can’t be found by classical computers).
Over the past two decades, quantum computers have progressed from relatively feeble devices capable of performing a few hundred quantum logic operations on a few quantum bits, to devices with hundreds or thousands of qubits capable of performing thousands to tens of thousands of quantum ops.
That is, we are just at the stage where quantum computers may actually be able to do something useful. Will they do it? Or will the whole project fail?
The primary technological challenge over the next few years is to get complex superconducting quantum circuits or extended quantum systems such as ion traps or quantum optical devices to the point where they can be sufficiently precisely controlled to perform computations that classical computers can’t. Although there are technological challenges of fabrication and control involved, there are well-defined paths and strategies for overcoming those challenges. In the longer run, to build scalable quantum computers will require devices with hundreds of thousands of physical qubits, capable of implementing quantum error correcting codes.
Here the technological challenges are daunting, and in my opinion, we do not yet possess a clear path to overcoming them.
“The primary technological challenge over the next few years is to get complex superconducting quantum circuits or extended quantum systems such as ion traps or quantum optical devices to the point where they can be sufficiently precisely controlled to perform computations that classical computers can’t.”
Quantitative futurist, Founder of the Future Today Institute, Professor of Strategic Foresight at New York University Stern School of Business, and the author, most recently, of The Big Nine: How the Tech Titans and Their Thinking Could Warp Humanity
The short answer is this: We continue to create new technologies without actively planning for their downstream implications. Again and again, we prioritise short-term solutions that simply never address long-term risk. We are nowists. We’re not engaged in strategic thinking about the future.
The best example of our collective nowist culture can be seen in the development of artificial intelligence. We’ve prioritised speed over safety, and longer-term strategy over short-term commercial gains. But we’re not asking important questions, like what happens to society when we transfer power to a system built by a small group of people that is designed to make decisions for everyone? The answer isn’t as simple as it may seem, because we now rely on just a few companies to investigate, develop, produce, sell, and maintain the technology we use each and every day. There is tremendous pressure for these companies to build practical and commercial applications for AI as quickly as possible. Paradoxically, systems intended to augment our work and optimise our personal lives are learning to make decisions that we, ourselves, wouldn’t. In other cases – like warehouses and logistics – AI systems are doing much of the cognitive work on their own and relegating the physical labour to human workers.
There are new regulatory frameworks for AI being developed by the governments of the US, Canada, EU, Japan, China, and elsewhere. Agencies like the US-based National Institute of Standards and Technology are working on technical standards for AI, but that isn’t being done in concert with similar agencies in other countries. Meanwhile, China is forging ahead with various AI initiatives and partnerships that are linking together emerging markets around the world into a formidable global network. Universities aren’t making fast, meaningful changes to their curricula to address ethics, values and bias throughout all of the courses in their AI programs. Everyday people aren’t developing the digital street smarts needed to confront this new era of technology. So they are tempted to download fun-looking, but ultimately suspicious apps. They’re unwittingly training machine learning systems. Too often, they are outright tricked into allowing others to access untold amounts of their social, location, financial, and biometric data.
This is a systemic problem, one that involves our governments, financiers, universities, tech companies and even you, dear Gizmodo readers. We must actively work to create better futures. That will only happen through meaningful collaboration and a global coordination to shape AI in way that benefits companies and shareholders, but also prioritises transparency, accountability and our personal data and privacy. The best way to engineer systematic change is to treat AI as a public good.
“Everyday people aren’t developing the digital street smarts needed to confront this new era of technology... Too often, they are outright tricked into allowing others to access untold amounts of their social, location, financial, and biometric data.”
University Distinguished Professor, Chicago-Kent College of Law, Illinois Institute of Technology, whose work focuses on the impact of technologies on individuals, relationships, communities, and social institutions
Technologies from medicine to transportation to workplace tools are overwhelmingly designed by men and tested on men. Rather than being neutral, technologies developed with male-oriented specs can cause physical harm and financial risks to women. Pacemakers are unsuited to many women since women’s hearts beat faster than men’s and that was not figured into the design. Because only male crash test dummies were used in safety ratings until 2011, seat-belted women are 47% more likely to be seriously harmed in car accidents. When men and women visit “help wanted” websites, the technological algorithms direct men to higher-paying jobs. Machine learning algorithms designed to screen resumes so that companies can hire people like their current top workers erroneously discriminate against women when those current workers are men.
Women’s hormones are different than men’s, causing some drugs to have enhanced effects in women and some to have diminished effects. Even though 80% of medications are prescribed to women, drug research is still predominantly conducted on men. Between 1997 and 2000, the US Food and Drug Administration pulled ten prescription drugs from the market, eight of which were recalled because of the health risks they posed to women.
On the other hand, some treatments may be beneficial to women, but never brought to market if the testing is done primarily on men. Let’s say that a drug study enrols 1000 people, 100 of whom are women. What if it offers no benefit to the 900 men, but all 100 women are cured? The researchers will abandon the drug, judging that it is only 10% effective. If a follow-up study focused on women, it could lead to a new drug to the benefit of women and the economy.
Workplace technologies also follow a male model. Female surgeons in even elite hospitals have to stack stools on top of one another to stand high enough to undertake laparoscopic surgeries. Their lesser hand strength causes them to have to use both hands to operate tools that male surgeons operate with one, leading female surgeons to have more back, neck and hand problems than men. Nonetheless, the patients of female surgeons do better than those of men. Imagine the health gain to the patients (and their female surgeons) if technologies were designed to accommodate women as well as men.
Female fighter pilots wear g-suits designed in the 1960s to fit men. These too-large suits do not provide adequate protection for women against g-forces, which can lead to a sudden loss of colour vision or a full blackout as blood begins to rush from their brain. The zippers generally don’t unzip far enough to comfortably fit the female bladder device, which causes some female pilots not to drink before missions, potentially leading to blackouts from dehydration. Other military equipment poses safety and efficacy risks to women. Designing with women in mind – such as the current work on exoskeletons – can benefit both female and male soldiers by providing protection and increasing strength and endurance.
I’d like to see the equivalent of a Moon Shot – a focused technology research programme – that tackles the issue of women and technology. Innovation for and by women can grow the economy and create better products for everyone.
“Innovation for and by women can grow the economy and create better products for everyone.”
Featured image: Illustration: Benjamin Currie (Gizmodo)