Nello Cristianini, Author at New �Ӱ�ԭ��

Intelligence rethought: AIs know us, but don’t think like us

Nello Cristianini — Wed, 23 Nov 2016 18:00:00 +0000

Many of us use artificial intelligence every day, often without realising it

Gilles Coulon/Tendance Floue

This article is being made free to view thanks to sponsorship from P&G

CAN a human-made creature ever surprise its creator, taking initiatives of its own? This question has been asked for centuries, from the golem of Jewish folklore to Frankenstein to I, Robot. There are various answers, but at least one computing pioneer knew well where she stood. “The Analytical Engine has no pretensions whatever to originate anything,” said Ada Lovelace, Charles Babbage’s collaborator, in 1843, removing any doubt about what a computing machine can ever hope to do. “It can do whatever we know how to order it to perform,” she added. “It can follow analysis; but it has no power of anticipating any analytical relations or truths.”

But 173 years later, a computer program developed just over a mile away from her house in London beat a master of the game Go. None of AlphaGo’s programmers can come close to defeating such a strong player, let alone the program they created. They don’t even understand its strategies. This machine has learned to do things that its programmers can’t do and don’t understand.

Far from being an exception, AlphaGo is the new normal. Engineers began creating machines that could learn from experience decades ago, and this is now the key to modern artificial intelligence (AI). We use them every day, usually without realising it.

For programmers who develop such machines, the whole point is to make them learn things that we don’t know or understand well enough to program in directly. This approach – called machine learning – has been extremely fruitful. It is the secret sauce of modern AI and has delivered recent successes (and spectacular failures) in autonomous cars, product recommendations, personal assistants, Go and more.

How can a machine learn? When I was growing up, my bicycle never learned its way home and my typewriter never suggested a word or spotted a spelling mistake. Mechanical behaviour was synonymous with being fixed, predictable and rigid. For a long time, a “learning machine” sounded like a contradiction, yet today we talk happily of machines that are flexible, adaptive, even curious.

The bicycle is a quintessential example of a non-learning machine

Getty

In artificial intelligence, a machine is said to learn when it improves its behaviour with experience. To get a feel for how machines can perform such a feat, consider the autocomplete function on your smartphone.

Event:

If you activate this function, the software will propose possible completions of the word you are typing. How can it know what you were going to type? At no point did the programmer develop a model of your intentions, or the complex grammatical rules of your language. Rather, the algorithm proposes the word that has the highest probability of being used next. It “knows” this from a statistical analysis of vast quantities of existing text. This analysis was done mostly when the autocomplete tool was being created, but it can be augmented along the way with data from your own usage. The software can literally learn your style.

The same basic algorithm can handle different languages, adapt to different users and incorporate words and phrases it has never seen before, such as your name or street. The quality of its suggestions will depend mostly on the quantity and quality of data on which it is trained. So long as the data set is sufficiently large and close in topic to what you are writing, the suggestions should be helpful. The more you use it, the more it learns the kinds of words and expressions you use. It improves its behaviour on the basis of experience, which is the definition of learning.

Note that a system of this type will probably need to be exposed to hundreds of millions of phrases, which means being trained on several million documents. That would be difficult for a human, but is no challenge at all for modern hardware.

If you feel that this is cheating, because the algorithm is not really intelligent, then brace yourself. Things get worse.

The next step up in complexity is a product recommendation agent. Consider your favourite online shop. Using your previous purchases, or even just your browsing history, the agent will try to find the items in its catalogue that have the highest probability of being of interest to you. These will be computed from the analysis of a database containing millions of transactions, searches and items. Here, too, the number of parameters that need to be extracted from the training set can be staggering: Amazon has more than 200 million customers and in excess of 3 million books in its catalogue.

Matching users to products on the basis of previous transactions requires statistical analysis on a massive scale. As with autocomplete, no traditional understanding is required – it does not need psychological models of customers or literary criticism of novels. No wonder some question whether these agents should be called “intelligent” at all. But they cannot question the word “learning”: these agents do get better with experience.

Emulating behaviour

These are just the simplest examples. Using the same or similar statistical techniques, in multiple parts of a system and at various scales, computers can now learn to recognise faces, transcribe speech and translate text from one language to another (see “Bots in translation“). According to some online dating companies, they can even find us potential love matches. In other words, they can emulate complex human behaviours that we cannot fully model. But they do it in a way that is very different from what we do.

“We can expect to reap the benefits of data-driven artificial intelligence for many years to come”

Things can get more complicated. Online retailers keep track not just of purchases, but also of any user behaviour during a visit to the site. They might track information such as which items you have added to the basket but later removed, which you have rated and what you have added to your wish list. Yet more data can be extracted from a single purchase: time of day, address, method of payment, even the time it took to complete the transaction. And this, of course, is done for millions of users.

As customer behaviour tends to be rather uniform, this mass of information can be used to constantly refine the agent’s performance. Some learning algorithms are designed to adapt on the fly; others are retrained offline every now and then. But they all use the multitude of signals extracted from your actions to adapt their behaviour. In this way, they constantly learn and track our preferences. It is no wonder that we sometimes end up buying a different item from the one we thought we wanted.

Intelligent agents can even propose items just to see how you respond. Extracting information in this way can be as valuable as completing a sale. Online retailers act in many ways as autonomous learning agents, constantly walking a fine line between the exploration and exploitation of their customers. Learning something they did not know about you can be as important as selling something. To put it simply, they are curious. A similar strategy could be used by spam filters and any other software that needs to learn your preferences and predict your actions. One day soon, the appliances in your house will be interested in predicting your next action too.

Machine learning is not just about analysing past behaviour. Sometimes AIs need to deal with novel situations. How do you help a new customer? To whom do you recommend a brand new book? The trick in this case is to get machines to generalise, using information from similar customers or products. Even a customer who has never used the service before leaves a small data trail – an email address and location, for example – to get started on. The ability to detect and exploit similarities is sometimes called pattern recognition, and its importance is not limited to “cold start” situations. In fact, generalisation – detecting patterns and similarities – is a fundamental part of intelligent behaviour.

But what does it mean that two items are similar? We could describe a book by the number of pages, the language it is written in, the topic, the price, the date of publication, the author, even some index of its readability. For a customer, useful descriptors might include age, gender or location. In machine learning, these descriptors are sometimes called features or signals. We can use them to locate similar items for which we have sufficient data. The machine can thus generalise from one situation to a similar one, and make better use of its experience.

Choosing the right features is one of the critical problems in machine learning: the font used in a book might not be as useful as its price, for example. This problem becomes even more critical when we handle complex items such as images. If you compare two passport photos of yourself taken one minute apart, they will not be identical at the level of raw pixels. This is sufficient for the computer to treat them as two completely different images. We would like the computer to represent those images in a more robust way than just using pixels, so it does not get confused by small irrelevant changes in the image. Which features of an image should be used in order to recognise the same face in different photos?

This has been a surprisingly stubborn problem, made worse by the variation in lighting, position and background that can occur in natural scenes.

Programming this capability directly into a computer has proven difficult, so engineers have once more resorted to machine learning. One such method, called deep learning, is currently delivering the best results in some domains (see “Layers of learning“). As with the earlier examples, it involves using big data to adjust millions of parameters.

Under the bonnet

Unknown to users, AI systems are constantly running tests on us

Issouf Sanogo/AFP/Getty Images

Consider now that these nuts and bolts of machine learning can be applied to many parts of the same system at the same time: a search engine might use them to learn how to complete your queries, best rank the answers for you, translate a document among the search results and select which ads to display. And this is just on the surface.

“This is the second part of our Instant Expert on artificial intelligence. The first part was published in the 29 October issue”

Unknown to users, the system will probably also be running tests to compare the performance of different methods by using them on different random subsets of users. This is known as A/B testing. Every time you use an online service, you are giving it a lot of information about the quality of the methods being tested behind the scenes. All this is on top of the revenue you generate for them by clicking on ads or buying products.

While each of these mechanisms is simple enough, their simultaneous and constant application on a vast scale results in a highly adaptive behaviour that looks intelligent to us. AlphaGo learned its winning strategies by studying millions of past matches and then playing against various versions of itself for millions of further matches. An impressive feat. Nonetheless, every time we understand one of the mechanisms behind AI, we cannot help feeling a little cheated. AI systems generate adaptive and purposeful behaviour without needing the kind of self-awareness that we like to consider the mark of “real” intelligence. Would Lovelace dismiss their suggestions as unoriginal? Possibly, but while the philosophers debate, the field keeps moving forward.

What, then, is the road ahead? Science often progresses via paradigm shifts, and AI is probably no exception. We are now in the middle of a very fruitful paradigm – data-driven AI – and we can expect to reap its benefits for many years. We should expect increasingly autonomous cars, increasingly fluent machine translators, and ever-cleverer cameras, phones and homes.

Should we also think that this is the last word in this most ambitious of fields? Is this really the way in which we are going to be building the intelligent machines that will follow us in our journeys to the planets and beyond? I think not. The history of science suggests that we haven’t seen the last revolution in AI. But where the next paradigm will emerge from, and what it will be capable of, is impossible to predict.

Bots in translation

One of the most impressive feats of modern artificial intelligence is translating text from one language to another. Google’s tool, for example, can translate with reasonable competence between 135 written languages, from Afrikaans to Zulu.

Most machine-translation products don’t have linguistic rules programmed into them. Instead, they apply statistical techniques to large data sets.

The text to be translated is first broken into words or phrases, each of which is then compared with a table of available translations and their probabilities of being accurate. There will be multiple ways to divide up the source sentence, and multiple ways to translate each phrase. The problem is how to select from among these possible translations and assemble them into a sentence in the target language that is both grammatical and an accurate translation.

This is computationally hard, but modern computers can do it. The probabilities required for such calculations are stored in two tables containing phrases and probabilities – the first one providing a statistical model of the target language, and the second a list of all phrases in the source language and their possible translations. That is where all the knowledge of the system lies. Change some entry, and the system will behave differently. Improve the estimation of probabilities, and it will boost its performance. No understanding of the text is needed, just statistical patterns.

The source of the information in those tables is, of course, vast data sets such as the proceedings of the European Parliament, which are translated into 23 languages. Challenging? Yes. Still prone to error? Yes. Teaching us something about how humans extract meaning from sentences? Hardly.

Layers of learning

One of the buzz phrases in artificial intelligence research is “deep learning”. It sounds exotic but is actually another form of the data-driven approach that has delivered so many successes in AI in recent years. Deep learning relies on a technology called neural networks, and has proven to be a valuable tool for solving hard perceptual problems such as recognising images.

Consider two photographs of the same face. They can vary enormously. A big unresolved problem in AI is how to identify the features that remain invariant despite differences in the raw sensory data, and how to use them to recognise the photos as being of the same face.

Originally based on a loose biological analogy with the human cortex, neural networks have been developed into complex mathematical objects. The raw sensory information – for example, the pixels of an image – is processed as it flows through a network of simple “neurons”, each specialised in recognising some aspect of the data. Lower-level neurons detect simple features of the image – straight lines, say – and feed that information to higher-level neurons, so that increasingly complex properties of images can be detected.

Crucially, these networks can adapt and improve with training, so that the most informative and stable features of an image can be discovered and used by the network without being directly programmed in by a human.

In their previous incarnations, neural networks were not particularly useful, but with modern hardware and giant data sets they have found a new life, delivering the best performance in certain perceptual tasks – most notably in vision and speech. Deep learning is typically used as a component in larger machine-learning systems (see main article).

This article appeared in print under the headline “A different way of thinking”

The road to artificial intelligence: A case of data over theory

Nello Cristianini — Wed, 26 Oct 2016 18:00:00 +0000

This article is usually available only to subscribers but is being made free to view thanks to sponsorship from Ocado IN the summer of 1956, a remarkable collection of scientists and engineers gathered at Dartmouth College in Hanover, New Hampshire. Among them were computer scientist Marvin Minsky, information theorist Claude Shannon and two future Nobel prizewinners, Herbert Simon and John Nash. Their task: to spend the summer months inventing a new field of science called “artificial intelligence” (AI). They did not lack in ambition, writing in their funding application: “every aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it.” Their wish list was “to make machines use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves”. They thought that “a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer.” It took rather longer than a summer, but 60 years and many disappointments later, the field of AI seems to have finally found its way. In 2016, we can ask a computer questions, sit back while semi-autonomous cars negotiate traffic, and use smartphones to translate speech or printed text across most languages. We trust computers to check passports, screen our correspondence and fix our spelling. Even more remarkably, we have become so used to these tools working that we complain when they fail. As we rapidly get used to this convenience, it is easy to forget that AI hasn’t always been this way. At the Dartmouth conference, and at various meetings that followed it, the defining goals for the field were already clear: machine translation, computer vision, text understanding, speech recognition, control of robots and machine learning. For the following three decades, significant resources were ploughed into research, but none of the goals were achieved. It was not until the late 1990s that many of the advances predicted in 1956 started to happen. But before this wave of success, the field had to learn an important and humbling lesson. While its goals have remained essentially the same, the methods of creating AI have changed dramatically. The instinct of those early engineers was to program machines from the top down. They expected to generate intelligent behaviour by first creating a mathematical model of how we might process speech, text or images, and then by implementing that model in the form of a computer program, perhaps one that would reason logically about those tasks. They were proven wrong. They also expected that any breakthrough in AI would provide us with further understanding about our own intelligence. Wrong again. Over the years, it became increasingly clear that those systems weren’t suited to dealing with the messiness of the real world. By the early 1990s, with little to show for decades of work, most engineers started abandoning the dream of a general-purpose top-down reasoning machine. They started looking at humbler projects, focusing on specific tasks that were more likely to be solved.

Event:

Some early success came in systems to recommend products. While it can be difficult to know why a customer might want to buy an item, it can be easy to know which item they might like on the basis of previous transactions by themselves or similar customers. If you liked the first and second Harry Potter films, you might like the third. A full understanding of the problem was not required for a solution: you could detect useful correlations just by combing through a lot of data. Could similar bottom-up shortcuts emulate other forms of intelligent behaviour? After all, there were many other problems in AI where no theory existed, but there was plenty of data to analyse. This pragmatic attitude produced success in speech recognition, machine translation and simple computer vision tasks such as recognising handwritten digits.

Data beats theory

By the mid-2000s, with success stories piling up, the field had learned a powerful lesson: data can be stronger than theoretical models. A new generation of intelligent machines had emerged, powered by a small set of statistical learning algorithms and large amounts of data. Researchers also ditched the assumption that AI would provide us with further understanding of our own intelligence. Try to learn from algorithms how humans perform those tasks, and you are wasting your time: the intelligence is more in the data than in the algorithm. The field had undergone a paradigm shift and had entered the age of data-driven AI. Its new core technology was machine learning, and its language was no longer that of logic, but statistics. How, then, can a machine learn? It is worth clarifying here what we normally mean by learning in AI: a machine learns when it changes its behaviour (hopefully for the better) based on experience. It sounds almost magical, but in reality the process is quite mechanical. Consider how the spam filter in your mailbox decides to quarantine some emails on the basis of their content. Every time you drag an email into the spam folder, you enable it to estimate the probability that messages from a given recipient or containing a given word are unwanted. Combining this information for all the words in a message allows it to make an educated guess about new emails. No deep understanding is required – just counting the frequencies of words. But when these ideas are applied on a very large scale, something surprising seems to happen: machines start doing things that would be difficult to program directly, like being able to complete sentences, predict our next click, or recommend a product. Taken to its extreme conclusion, this approach has delivered language translation, handwriting recognition, face recognition and more. Contrary to the assumptions of 60 years ago, we don’t need to precisely describe a feature of intelligence for a machine to simulate it. While each of these mechanisms is simple enough that we might call it a statistical hack, when we deploy many of them simultaneously in complex software, and feed them with millions of examples, the result might look like highly adaptive behaviour that feels intelligent to us. Yet, remarkably, the agent has no internal representation of why it does what it does. This experimental finding is sometimes called “the unreasonable effectiveness of data”. It has been a very humbling and important lesson for AI researchers: that simple statistical tricks, combined with vast amounts of data, have delivered the kind of behaviour that had eluded its best theoreticians for decades. Thanks to machine learning and the availability of vast data sets, AI has finally been able to produce usable vision, speech, translation and question-answering systems. Integrated into larger systems, those can power products and services ranging from Siri and Amazon to the Google car. Researchers’ attention is now focused what fuels the engine of our intelligent machines: data. Where can they find data, and how can they make the most of this resource?

“Modern artificial intelligence is a brilliant and powerful technology, but also a fundamentally disruptive one“

One important step has been to recognise that valuable data can be found freely “in the wild”, generated as a byproduct of various activities – some as mundane as sharing a tweet or adding a smiley under a blog post. Engineers and entrepreneurs have also invented a variety of ways to elicit and collect additional data, such as asking users to accept a cookie, tag friends in images, rate a product or play a location-based game centred on finding monsters in the street. Data became “the new oil”. At the same time as AI was finding its way, we developed an unprecedented global data infrastructure. Every time you access the internet to read the news, do a search, buy something, play a game, or check your email, bank balance or social media feed, you interact with this infrastructure. It isn’t just a physical one of computers and wires, but also one of software, including social networks and microblogging sites. Data-driven AI both feeds on this infrastructure and powers it – it is hard to imagine one without the other. And it is hard to imagine life without either of them.

Challenges ahead

This is what makes modern AI a brilliant and powerful technology, but also a fundamentally disruptive one. The unified data infrastructure is not like any medium invented before. Unlike the copper cables that used to connect people in the telegraph or telephone age, it takes a keen interest in our actions. The medium looks back at us, anticipating our moves, guessing our intents, often trying to serve us better and sometimes to influence us. This gives a whole new meaning to the claim, made by the 1970s communications theorist Marshall McLuhan, that a medium can never be neutral. The challenges AI might present us with include surveillance, discrimination, persuasion, unemployment and possibly even addiction. Are we prepared? Intelligent machines need to collect data – often personal data – in order to work. This simple fact potentially turns them into surveillance devices: they know our location, our browsing history and our social networks. Can we decide who has access, what use can be made of the data, or whether the data gets deleted for ever? If the answer is no, then we don’t have control.

Driverless cars “see” better than ever, but mistakes can’t be ruled out

Jeff Swensen/The New York Times/Redux / eyevine

AI’s capability to make predictions is useful for insurance, loans and policing. But the quality of those predictions will depend on subtle design choices and on the way the information used to train it is collected, which creates a very real risk of implicit and unintended discrimination. A recent , for example, claims to have uncovered a bias that would disadvantage African Americans in the software used in many US courts to make parole decisions. Another case has been reported where different job ads were targeted at different ethnic groups. Both starkly illustrate the unintended effects of the complex interaction between algorithms and data. Another concern is persuasion. The business model of many AI companies is advertising, which means getting people to click on specific links. Research on how to steer users is well under way. The more the machines know about us, the better the job they can do of nudging us. Predictive interfaces might even induce addiction in vulnerable users, by actively rewarding them with the juiciest content that the web has to offer. This is something that needs to be carefully studied. Employment will be affected too, as AIs learn from us (quite literally) how to do certain jobs, either because they watch how we do them, or because we are paid to generate their training data. The emergence of internet crowdsourcing allows businesses to automatically outsource micro-tasks that require human intelligence, by posting them on websites or apps where workers can choose the tasks they want to accept. In a way this works just like Uber, but for tasks other than driving, and is mediated by a computer system. Typical tasks would include transcribing handwriting or labelling images.

AI will increase the automation of warehouses like Amazon’s

Ralph Freso / Reuters

This also creates a workforce directly managed via computers, and defines a set of tasks that are the ideal candidate for automation. Indeed, many of those task-workers are actually generating or annotating the data being used to train their AI replacements. At the same time, we can expect many call centres and warehouses to be increasingly automated within a decade. I do not believe that we yet have the legal and cultural tools to handle these and many other challenges. Who do we turn to if an intelligent algorithm denies us parole, medical treatment or a diploma? Are we prepared for our character and trustworthiness to be ranked just like our credit history, as some countries are proposing? Do we want the state to have access to our online activities and knowledge of our preferences? Do we want our children to spend their online time in the company of persuasive machines, designed to steer their behaviour in a given direction? What happens to society if large numbers of people are put out of work?

“Coming soon: the second half of our Instant Expert on artificial intelligence, in which Nello Cristianini delves deeper into the technology that allows machines to learn“

Artificial intelligence has come a long way from its early days in academic laboratories. It is now being integrated into our lives, and promises to improve them. We might not call it AI once it is deployed, but we can expect benefits in fields ranging from healthcare to transportation, from communications to schooling. And research is not slowing down. The machine-learning paradigm has been effective in addressing many areas like vision and speech processing, and it is likely that future AI will also find a way to integrate some top-down reasoning methods descended from earlier approaches. What will come after that may surprise us again. As our AI efforts continue to open up new possibilities, we can imagine seamless conversations with machines, fluent real-time translation of speech, and many useful ways to automate our houses and cars. But we might want to resist the temptation to introduce AI into as many domains as possible, at least before the cultural and legal framework evolves. Widespread adoption of AI brings remarkable opportunities, but also potential risks. Contrary to popular belief these are not existential risks to our species, but rather a possible erosion of our privacy and autonomy. So as we finally enjoy the benefits of six decades of research in AI, with machines joining us in our everyday lives, we should celebrate – but also tread carefully.

The winters of AI discontent

Emergent technologies are often subjected to hype cycles, sometimes due to speculative bubbles inflated by excessive investor expectations. Some examples are railway mania in the UK in the 1840s and the dot-com bubble in the 1990s. Artificial intelligence is perhaps unique in having undergone several hype cycles in a relatively short time. Its slumps of optimism even have a specific name: AI winters (see timeline below). The two major winters occurred in the early 1970s and late 1980s. Both were caused largely by the withdrawal of public funding as progress stalled. AI is now in a renewed phase of heightened optimism and investment. Unlike in previous cycles, however, AI today has a strong – and increasingly diversified – commercial revenue stream. Only time will tell whether this turns out to be a bubble.

1950 Alan Turing publishes the seminal paper ““. Its opening sentence is “I propose to consider the question, ‘Can machines think?'” 1956 The term “artificial intelligence” is coined at a workshop at Dartmouth College 1959 Computer scientists at Carnegie Mellon University create the General Problem Solver (GPS), a program that can solve logic puzzles 1973 The first AI winter sets in as funding and interest dry up 1975 A system called MYCIN diagnoses bacterial infections and recommends antibiotics using deduction based on a series of yes/no questions. It was never used in practice 1987 Second AI winter begins 1989 NASA’s AutoClass computer program discovers several previously unknown classes of stars 1994 First web search engines launched 1997 IBM’s Deep Blue beats world champion Garry Kasparov at chess 1998 NASA’s Remote Agent is first fully autonomous program to control a spacecraft in flight 2002 Amazon replaces human product recommendation editors with an automated system 2007 Google launches Translate, a statistical machine translation service 2009 Google researchers publish an influential called “The unreasonable effectiveness of data”. It declares that “simple models and a lot of data trump more elaborate models based on less data” (IEEE Intelligent Systems, vol 2, p 8) 2011 Apple releases Siri, a voice-operated personal assistant that can answer questions, make recommendations and carry out instructions such as “call home” 2011 IBM’s supercomputer Watson beats two human champions at TV quiz game Jeopardy! 2012 Google’s driverless cars navigate autonomously through traffic 2016 Google’s AlphaGo defeats Lee Sedol, one of the world’s leading Go players

You win some…

One of the most celebrated successes of machine learning (see main story) came earlier this year when an algorithm called AlphaGo defeated South Korean master Lee Sedol at the game Go – something none of its programmers could come close to doing themselves. AlphaGo combined various machine-learning methodologies to analyse databases of more than 30 million Go moves, as well as playing thousands of games against itself. A similar strategy earlier allowed IBM’s Watson supercomputer to win at the TV quiz game Jeopardy! (pictured above). Given the right data, it seems that machines can improve their intelligence a great deal. But we should remember that machine learning is a statistical exercise, and therefore it can always fail. In recent years we have also seen some blunders caused by machine learning. Last year Google apologised after one of its products automatically labelled photos of two black people “gorillas”; this year Microsoft had to withdraw a conversational bot called Tay because it had learned offensive language. In both cases it was not a failure of the algorithm, but of the training data that had been fed to it. This year also saw the first fatality linked to a “driverless” car, when a driver put a Tesla on autopilot and it failed to detect a trailer on the road. The conditions were unusual, with a white obstacle against a light sky, and the computer vision system simply made a mistake. I do not expect it to be the last one as many companies move into that market. On the other hand, there are countless stories that do not end up in the news, simply because the AI systems are doing their work as expected. They include search engines, online shops and semi-autonomous cars. As we entrust machines with increasingly sensitive decisions, we need to pay careful attention to the kind of data we feed to them. It is not just the technology, but its deployment in our everyday life, that needs better understanding.

This article appeared in print under the headline “Intelligence reinvented”