The Alignment Problem: Machine Learning and Human Values Brian Christian W. W. Norton (2020)

Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence Kate Crawford Yale Univ. Press (2021)

Artificial intelligence (AI) permeates our lives. It determines what we learn and purchase, whether or not we get a job, mortgage, mortgage, subsidies or parole. It diagnoses ailments and underlies — and undermines — democratic processes. Two new books supply complementary visions of how society is being reshaped by those that construct, use and manage AI.

In The Alignment Problem, author Brian Christian offers an intimate view of the people making AI know-how — their goals, expectations, hopes, challenges and desolations. Starting with Walter Pitts’s work on a logical illustration of neuron exercise in the early twentieth century, he recounts the concepts, goals, successes and failures of researchers and practitioners in fields from cognitive science to engineering. Atlas of AI, from the influential scholar Kate Crawford, offers with how, virtually, AI will get into and performs out in our lives. It exhibits that AI is an extractive trade, exploiting assets from supplies to labour and data.

Both books parse how the energy of, and over, the digital world is shifting politics and social relations. They nod to different approaches to those imbalances — state management in China, or the regulatory efforts of the European Union — however give attention to the North American story. Together, they beg for a sequel: on methods ahead.

From prophecy to ubiquity

Christian tracks the evolution of AI know-how, from prophecy to ubiquity. He exhibits how researchers attempt to get AI to interpret human values akin to equity, transparency, curiosity and uncertainty, and the challenges that stand of their approach. Largely primarily based on encounters with researchers and practitioners, the e book charts a gradual, regular, advanced progress, with many lows and some unimaginable highs.

We meet people akin to Rich Caruana, now senior principal researcher at Microsoft in Redmond, Washington, who was requested as a graduate pupil to look at one thing that led to his life’s work — optimizing information clustering and compression to make fashions that are each intelligible and correct. And we stroll alongside the seaside with Marc Bellemare, who pioneered reinforcement studying whereas working with video games for the Atari console and is now at Google Research in Montreal, Canada.

Christian exhibits researchers’ rising realization that AI developments are affected by societal values — and, extra importantly, have an effect on them. They include a price and can have profound impacts on communities. At its core, The Alignment Problem asks how we will guarantee that AI methods seize our norms and values, perceive what we imply and do what we would like. We all have completely different conceptions of and necessities for what such methods ought to do. As mathematician Norbert Wiener put it in 1960: “We had better be quite sure that the purpose put into the machine is the purpose which we really desire.”

Crawford’s assortment exposes the darkish facet of AI success. It traverses the globe exploring the relationships between places and their impression on AI infrastructure. From Nevada and Indonesia, the place the lithium and tin central to semiconductors are mined at nice human and environmental value, we journey to an Amazon warehouse in New Jersey. Here, labourers bend their our bodies to the will of robots and manufacturing strains, as a substitute of the automation adapting to human tempo. In an uneasy reminder of Charlie Chaplin’s 1936 movie Modern Times, we witness the hardship of “fauxtomation” — supposedly automated methods that rely closely on human labour, akin to that of employees paid under minimal wage in data-labelling farms.

Crawford concludes with the arresting reminder that AI will not be goal, impartial or common. Instead, it is deeply embedded in the tradition and financial actuality of those that make it — largely, the white, rich males of California’s Silicon Valley.

Both books are sturdy of their exposition of the challenges and risks of present use and improvement of AI, and what units it aside from ‘classic’ computing. Reading them facet by facet highlights three core points: over-reliance on data-driven, stochastic predictions; automated selections; and focus of energy.

Data dominance

As AI researcher Joy Buolamwini feedback in the 2020 documentary Coded Bias, algorithmic decision-making is undoing a long time of progress in direction of equal rights, reifying the very prejudices it reveals to be nonetheless deeply rooted. Why? Because utilizing information to tell automated selections usually ignores the contexts, feelings and relationships that are core to human selections.

Data are usually not uncooked supplies. They are at all times about the previous, and they replicate the beliefs, practices and biases of those that create and gather them. Yet present utility of automated decision-making is knowledgeable extra by effectivity and financial advantages than by its results on people.

Worse, most approaches to AI empower those that have the information and the computational functionality to course of and manage them. Increasingly, these are huge tech companies — non-public entities exterior democratic processes and participatory management. Governments and people are the customers, not the leaders. The penalties of this shift are monumental and have the potential to change society.

So, what now? Besides efforts to debias the information and clarify selections algorithms make, we have to deal with the supply of the bias. This will likely be completed not by means of technological fixes, however by training and social change. At the identical time, analysis is required to handle the subject’s perverse dependence on correlations in information. Current AI identifies patterns, not which means.

Meticulously researched and beautifully written, these books in the end maintain up a mirror. They present that the accountable — moral, authorized and useful — improvement and use of AI will not be about know-how. It is about us: how we would like our world to be; how we prioritize human rights and moral rules; who contains this ‘we’. This dialogue can wait now not. But the key query is: how can all have a voice?