Artificial Intelligence

How It's Developing

One of the big stories highlighting AI’s development and reach was the AlphaGo program’s victory over 18-time Go world champion Lee Sedol in a five-game series in 2016. In an article in Nature, DeepMind researchers explained the development of AlphaGo's expertise through a combination of Monte-Carlo tree search (an algorithm for optimal decision making) and deep neural networks that have been trained via supervised learning, observing human expert games, and reinforced by playing games against itself. [1] A year later, in 2017, AlphaGo had another successful outing, defeating a team of five Go champions and then demonstrating a collaborative match where two teams, each composed of a human and an AlphaGo companion, played against each other – the researchers celebrated the collaborative approach as a future direction for AI where humans would work in step with Artificial Intelligence to elevate overall performance. [2]

If Google’s DeepMind AlphaGo represented the positive developments toward AI, Microsoft’s Tay artificial intelligence chat bot represented the challenges and limits of the technology. Microsoft’s research team launched the AI chatbot on Twitter, GroupMe, and Kik as a way to test and improve Microsoft's understanding of conversational language, including the nuances of teens’ online language. [3] The bot quickly began issuing offensive posts (disputing the existence of the Holocaust, referring to women and minorities with unpublishable words, and advocating genocide), partly in response to user commands for the bot to repeat users’ own statements, while also learning bad behavior as it ingested content from its social media forums. [4] Microsoft apologized for the unintended offensive tweets and tried to explain some of what happened while recognizing the pilot as part of a process for moving forward with the technology. [5] The experience with Tay could actually limit artificial intelligence development, as some technology companies have become reluctant to set conversational artificial intelligence systems free to talk with the large numbers of people needed to train them. [6]

Technology companies are finding roles for artificial intelligence in moderating online content. Facebook’s artificially intelligent language processing engine, Deep Text, applies deep learning to understand human language – the company initially pursued Deep Text to power chatbots in Messenger, to filter out spam and abusive comments from users’ Newsfeeds, and to help understand the topic area and even the content of just about anything people post on the social network. [7] The system quickly surpassed humans in flagging offensive photos, quarantining obscene content before it ever reaches users. [8] As Facebook’s use of AI has advanced, it has begun to explore artificial intelligence’s use in flagging material on the video platform Facebook Live, which requires computer vision algorithms that are fast, know how to prioritize policies, and when content should be taken down. [9] Facebook also sees opportunities for artificial intelligence to teach itself to identify key phrases that were previously flagged for being used to bolster a known terrorist group, to identify users who create fake accounts in order to spread extremist or terrorist content, or to identify users associated with clusters of pages or groups that promote extremist content. [10] In 2017, Facebook announced plans to integrate AI into a program that allowed users to flag troubling image or status posts, helping to identify posts that suggest that a user may be suicidal; Facebook partnered with organizations like the National Suicide Prevention Lifeline, the National Eating Disorder Association, and the Crisis Text Line so that when users’ posts are flagged, they can connect immediately via Messenger. [11] Facebook’s suicide prevention program evolved to the point that the technology can proactively identify a post or Facebook Live broadcast "likely to include thoughts of suicide," and send those posts to Facebook's trained reviewers who, in turn, can contact first responders. [12]

Facebook has continued to pursue AI as a tool to better understand content on the network, including its Automatic Alternative Text tool which uses deep neural networks to identify particular objects in a photo and pick out particular characteristics of the people in the photo to create a caption that a text-to speech engine can then read aloud for users with low visibility – while the system doesn’t always get images exactly correct, it is an improvement and shows the growing potential for AI to recognize and describe photos and images. [13] Facebook is also exploring how artificial intelligence can process content and make suggestions based on that content. By integrating AI into its personal assistant technology M, Facebook can suggest users book an Uber or prompt them to send money to a friend based on whatever it is the user was talking about in Messenger. [14] And Facebook developers have also used deep learning and neural networks to train its Lumos system to recognize scenes, objects, animals, places, attractions, and clothing items in photos, allowing users to search for and retrieve photos even if they themselves have not annotated them. [15]

In addition to content moderation, artificial intelligence is increasingly being used for content generation. From short films (Sunspring by the AI program Benjamin), to podcasts (Sheldon County), to short stories (Shelley), artificial intelligence systems are being used to develop creative or artistic outputs. [16] AI is also increasingly used to evaluate artistic outputs, such as a system of neural networks developed by Disney and the University of Massachusetts Boston that can evaluate short stories to predict which stories will be most popular by looking at different sections of each story and the holistic view of a story's meaning. [17]

AI is also being used to develop informational content and news reporting. Systems like IBM’s Watson are providing real-time scores, assessments, and automated video captions for a range of sporting events and cultural activities. [18] Newspapers are turning to AI to produce news coverage, such as The Washington Post’s use of its Heliograf artificial intelligence program to cover every House, Senate, and gubernatorial race on election day, freeing up reporters to focus on high-profile contests. [19]

And artificial intelligence is also being integrated into education. The IBM Foundation and the American Federation of Teachers have collaborated to build Teacher Advisor, a program that uses artificial intelligence technology to answer questions from educators and help them build personalized lesson plans. [20]

These advances are all in addition to the ways that artificial intelligence research will transform higher education and research centers. IBM and MIT have signed a 10-year, $240 million partnership agreement that establishes the MIT-IBM Watson AI Lab where IBM researchers and MIT students and faculty will work side by side to conduct advanced AI research. [21]

As artificial intelligence makes its way into more and more sectors, the dominant concern remains the potential impact it will have on the workforce. A 2018 Gallup survey found that the American public widely embraces artificial intelligence in attitude and practice, with nearly five in six Americans already using some product or service featuring AI, but most Americans recognize the technologies’ potential impact on future employment. [22] Those concerns for employment are supported by ample research. A 2018 report from PwC predicts three waves of automation – a flood of algorithms where machines handle data analysis and simple digital tasks; augmentation inundation, when repeatable tasks and the exchange of information will come to be done by humans and automated systems working together; and, finally, an autonomy tsunami, when machines and software will make decisions and take physical actions with little or no human input – with experts noting that most developed countries are already well into the first stage. [23] Still other research places AI’s development and threat in a more nuanced context. An AI Index created by researchers at Stanford University and the Massachusetts Institute of Technology, a McKinsey Global Institute report, and a National Bureau of Economic Research article by economists from M.I.T. and the University of Chicago collectively suggest that AI can likely do less now than we think, but that it will eventually do more in more sectors than we could expect, and that it will probably evolve faster than past technologies. [24] As important as the question of when, might be the question of where – geographically and in which sectors. Several reports (a 2017 study from Northwestern University and MIT and a 2017 report from the Institute for Spatial Economic Analysis at the University of Redlands) indicate that AI might have its greatest effects on cities where more jobs are routine clerical work, such as cashier and food service jobs, which are more susceptible to automation – while that could include larger cities like Las Vegas, Orlando, and Louisville, it could also include smaller cities with fewer than 100,000 people, where such jobs may have higher concentrations. [25] As routine service and clerical jobs become susceptible to automation, other industries that rely on skills in statistics, mathematics, and software development, will likely see growth or stability, as they build and improve the systems that replace traditional manufacturing and service workers. [26]

A preview of AI’s potential impact on clerical work might be available in Google Duplex. At its 2018 I/O conference Google debuted its Google Duplex AI System, which helps Google Assistant accomplish real-world tasks over the phone (book an appointment, make reservations) – initially, the system only operates in “closed domains” (exchanges that are functional, with strict limits on what is going to be said) and will have “disclosure built-in” so that a verbal announcement will be made to the person on the other end of the call. [27] Google has begun to explore options for Duplex’s use in call centers to improve call handling by giving the common but simple queries to Duplex, leaving a limited number of human workers to field more advanced call issues. [28] In a similar vein, IBM Watson and Japanese insurance company Fukoku Mutual Life Insurance introduced an AI solution that can scan hospital records and other documents to determine insurance payouts, factoring injuries, patient medical histories, and procedures administered – the system will replace 34 human insurance claim workers. [29]

Google’s Duplex is just one of several initiatives to make artificial intelligence systems that can communicate more like humans and accomplish more human tasks. IBM’s Project Debater seeks to interact and debate with people across 100 topics – the current scope of interactions are tightly constrained to a four-minute opening statement, followed by a rebuttal to the opponent’s argument, and then a statement summing up a viewpoint. [30] Amazon’s Alexa Prize competition has researchers create a chatbot using Alexa that can talk to a human for 20 minutes without messing up. [31]

Even if AI does not fully replace jobs, there is a clear desire to use AI to augment work. Google’s DeepMind has begun exploring avenues into healthcare with the creation of DeepMind Health that will create apps to help medical professionals identify patients at risk of complications and organize and prioritize admitted patients – while neither of the initial products use artificial intelligence, deep learning, or neural networks, the entry into the space indicates their longer-term interest in the technology’s deployment in this space. [32]

Through all of these developments, governments will increasingly consider the technology’s potential effects on the economy and innovation. The U.S. government has accelerated its focus on artificial intelligence, hosting a White House summit on artificial intelligence that included representatives from 38 companies (including Amazon, Facebook, Google, and Intel) to discuss how the government can fund AI research and alter regulations to advance the technology and announcing a Select Committee on Artificial Intelligence made of up the leading AI researchers in government and charged with advising the White House on governmentwide AI research and development priorities and the establishment of partnerships between government, the private sector, and independent researchers. [33] The Trump administration has pledged to release more government data that might help fuel AI research in the U.S., but what kind of data would be released or who would be eligible to receive the information remains unclear. [34]

Why It Matters

Artificial intelligence could become an invaluable tool for organizing and making accessible large collections of information. Google’s Life Tags project is a searchable archive of Life magazine photographs that used artificial intelligence to attach hundreds of tags to organize the archive. [35] Another Google project, Talk to Books, lets users type in a statement or a question and the system retrieves whole sentences in books related to what was typed, with results based not on keyword matching, but on more complex training of AI to identify what a good response looks like. [36] The Allen Institute for Artificial Intelligence, a nonprofit created by Microsoft co-founder Paul Allen, unveiled Semantic Scholar, a search engine that uses machine learning and other AI to improve the way academics search through the growing body of public research, more easily accessing research papers, targeting specific results, and revealing images, by using natural language processing algorithms and computer vision technology. [37]

As AI becomes more adept at generating content, it could further complicate users’ navigation of a complex information environment. Artificial intelligence will be able to create 3D face models from a single 2D image; manipulate facial expressions on video in real time using a human “puppet”; change the light source and shadows in any picture; generate sound effects based on mute video; and resurrect characters using old clips – and many of these effects have given rise to the “deep fakes” that manipulate video and other images. [38]

As with many other technologies, AI may become one more development that libraries help communities better understand. Facebook launched a campaign to educate people on the basics of artificial intelligence, focusing on the technology behind photo recognition, self-driving cars, and language translation. [39] In a similar way, the Urban Libraries Council (ULC) articulated a vision for libraries to serve communities by advancing algorithmic literacy while also ensuring an equitable and inclusive future by monitoring the storage, privacy, and application of data as AI technology becomes more ubiquitous.

If AI becomes a serious threat to jobs, libraries’ roles in workforce development may become even more important, but also more complicated. A compounded challenge may arise where workforce development will need to encompass not only the preparation for entry level individuals (into a market that is increasingly limited and competitive), but also solutions for a new vacuum in middle level management caused by the elimination of once plentiful entry level workers who matriculated into middle management. [40] The new workforce development demands will likely require higher-order critical, creative, and innovative thinking as well as emotional engagement, placing a greater value on the quality of thinking, listening, relating, collaborating, and learning. [41]

AI’s dependence on data sets can reinforce certain human systems, including bias. [42] Many researchers and practitioners are exploring options to address sexism and racism in AI development by curating new data sets that balance gender and ethnicity and more intentionally labeling and annotating data sets to show how the sets were collected. [43] To help change the way AI understands LGBT-related content, GLADD announced a partnership with Alphabet’s Jigsaw division to train AI with positive LGBT-related content and distinguish between phrases that are offensive to the LGBT community and those that are acceptable. [44] Coupled with efforts to change the scope and nature of data that trains AI systems are efforts to recruit women and other underrepresented groups into the field of artificial intelligence. [45]

Issues of sexism, racism, and bias are just part of the larger ethical concerns around AI. In 2017, Google launched a DeepMind ethics group to oversee the responsible development of artificial intelligence by helping developers put ethics into practice and educating society about the potential impacts of AI. [46] A 2018 report, authored by two dozen researchers from Oxford, Cambridge, OpenAI, the Electronic Frontier Foundation, Endgame, and the Center for a New American Security, focused on the potential negative effects of AI, including malicious uses of the technology. [47] While computer science programs have been required to provide students with an understanding of ethical issues related to computing in order to be accredited by ABET, a growing number of universities are launching new courses on the ethics of artificial intelligence, the ethical foundations of computer science, and other offerings that will help train the next generation of technologists and policymakers to consider the ramifications of innovations before those products are made available to the public. [48] As technologist are increasingly motivated to consider the ethical implications of their innovations, religion, philosophy, and the humanities could play an increasingly important role in the development of artificial intelligence. [49]

Many technology leaders are working to open the artificial intelligence field to make it more collaborative. Organizations like OpenAI, which was established by tech leaders like Elon Musk, Peter Thiel, and Reid Hoffman, promote a beneficial goal of advancing digital intelligence in ways that benefit humanity, free from the demand to generate financial return. [50] Facebook, Amazon, Microsoft, Google's DeepMind, and IBM are among the major partners in the Partnership on Artificial Intelligence to Benefit People and Society, which seeks to conduct open-source research and investigate globally important AI issues such as ethics and human and AI system collaboration. [51] In 2016, Apple announced plans to allow its artificial intelligence teams to publish research papers, reversing an earlier strategy to keep their research in-house, in the hopes that engaging with the larger community might allow researchers to feed off wider advances in the field. [52]

Even as artificial intelligence research has sought to become more collaborative, it has also put a strain on traditional systems of research and knowledge production and sharing. Many universities in the United States and Europe are losing talented computer scientists and artificial intelligence experts, lured away from academia by private sector offers – the shift from academic settings to the private sector has implications for not only research production and dissemination, but also the teaching and training of future generations. [53] In the United States, some technology companies have shifted their artificial intelligence operations to be closer to the universities that produce leading researchers. Facebook opened new artificial intelligence labs in Seattle and Pittsburgh after hiring three AI and robotics professors from the University of Washington and Carnegie Mellon University – in addition to advancing Facebook’s research, the professors will be better positioned to recruit and train other AI experts from those universities’ programs. [54] Still other technology companies have developed research labs with specific commitments to academic institutions – Microsoft’s Research AI unit engaged in a formal partnership with MIT’s Center for Brains, Minds and Machines. [55]