IELTS Academic Reading Practice Test 192

Few objects are as familiar, and as rarely examined, as the touchscreen. Millions of people tap, swipe and pinch a sheet of glass every day without wondering how it locates their finger. Yet the technology did not appear from nowhere in 2007, when the first truly successful touch-controlled smartphone went on sale. Its underlying principles were worked out in the 1960s, in laboratories more concerned with radar and air traffic control than with consumer electronics, and touch-sensitive screens were being demonstrated before most offices possessed a personal computer.

The first published description came from E. A. Johnson, an engineer at the Royal Radar Establishment in Malvern, in the west of England. In 1965 he wrote a short paper about a screen on which a user could select an item simply by pointing at it, and three years later he published a fuller account, complete with diagrams of the hardware. Johnson's motive was practical. Air traffic controllers had to identify aircraft on a radar display, and doing so with a light pen or a joystick was slow and liable to error. A screen that responded to a fingertip promised to make the task quicker and more direct.

His device exploited capacitance, the ability of a conducting surface to store an electric charge. A thin, transparent film of metal oxide, which conducts electricity while allowing light to pass, was spread across the glass. A small voltage applied to the film created an electric field around it. When a finger, which is itself a conductor, approached the surface, it changed the capacitance at that particular point, and electronics arranged around the edge of the screen could detect the change and calculate where the touch had occurred. Nothing had to be pressed and nothing moved, so the screen could not wear out as a mechanical switch does.

A parallel line of work took place at CERN, the particle-physics laboratory near Geneva. In the early 1970s Bent Stumpe, a Danish engineer, and his colleague Frank Beck were designing controls for a new accelerator, the Super Proton Synchrotron, and found the existing push-buttons and knobs inflexible. In 1973 they built a capacitive touch screen on which a grid of buttons was drawn, so that the function of each could be changed simply by redrawing it. The device was reliable and, because it had no moving parts, unusually robust in a control room. CERN did not exploit it commercially, and reports of the work attracted little attention outside particle physics at the time.

In the United States, the physicist Sam Hurst approached the problem from a different direction. Working at the University of Kentucky, he built a sensor that located a touch by measuring electrical resistance rather than capacitance. In 1971 he founded a company, Elographics, to develop it, and within a few years the firm had patented a transparent touch sensor. By the end of the 1970s it was producing screens that could be fitted over ordinary computer monitors.

A resistive screen consists of two transparent conducting layers, separated by a grid of microscopic spacers. Pressure from a finger or a stylus brings the upper layer into contact with the lower one, closing a circuit at that point, and the controller then measures voltages to work out where the contact happened.

Such screens are inexpensive and respond to almost anything: a fingernail, a gloved hand, a plastic pen. That made them suitable for factories, restaurants and cash machines, where they were widely installed from the 1980s onwards. Their disadvantages are that the extra layers reduce the brightness of the display and, in the early versions, only one point of contact could be registered at a time.

Meanwhile, other methods were being tried. The PLATO IV teaching terminal, developed at the University of Illinois and introduced in 1972, surrounded its screen with a grid of invisible infrared beams; a finger touching the glass interrupted the beams and revealed its position. Hewlett-Packard used the same principle in 1983 in the HP-150, one of the first desktop computers with a touch screen, though reviewers complained that the beams were easily broken by dust and that the screen quickly became dirty. Systems based on surface acoustic waves, in which ultrasonic vibrations travel across the glass and are disturbed by a touch, followed later.

Recognising several fingers at once is an older idea than most people assume. In 1982 Nimish Mehta, a graduate student at the University of Toronto, built an interface that could follow more than one finger simultaneously, and other laboratories produced similar demonstrations during the 1980s. Yet for two decades the commercial market was dominated by single-touch resistive screens: the IBM Simon, a bulky telephone with a stylus, appeared in 1994, and the Nintendo DS handheld console, which sold in enormous numbers, used a resistive screen from 2004. Multi-touch remained a laboratory curiosity.

The change came when projected-capacitance sensing, a refinement of Johnson's principle, was combined with a screen able to display a sharp image and a processor able to interpret gestures. That combination reached the market in 2007, and within a few years the resistive screen had been displaced from almost every new phone. The journey from a radar laboratory in Worcestershire to a pocket in every city in the world had taken more than forty years, and most of it had been spent not on inventing the technology but on waiting for the rest of electronics to catch up.

Every rocket that leaves the ground is a compromise. It must carry not only its payload but also the propellant that will lift it. The reason is that the exhaust from a chemical engine leaves the nozzle at only three or four kilometres per second, a speed fixed by the energy released when the propellants burn. Raising a spacecraft's velocity by several kilometres per second therefore demands an enormous quantity of fuel. Ion propulsion takes the opposite approach. Instead of releasing energy from chemical bonds, it uses electricity to hurl charged atoms out of the back of the spacecraft at thirty kilometres per second or more.

To compare engines, engineers use a quantity called specific impulse, which expresses how much push a given mass of propellant can supply and is measured in seconds. The best chemical propellants achieve around 450 seconds; gridded ion thrusters reach 3,000 or more. A higher specific impulse means that a spacecraft can accomplish the same change of velocity with far less propellant, which is why ion engines appeal for long missions and for satellites that must hold their position for fifteen years. The catch is that the exhaust is extremely thin. The ions are so few, and their combined momentum so slight, that the thrust of a typical engine is measured in tens of millinewtons, about the weight of a few sheets of paper.

A gridded ion thruster works in three stages. First, a gas — usually xenon, a heavy noble gas — is fed into a chamber and bombarded with electrons from a cathode, which knock electrons off the atoms and leave them positively charged. Second, the ions drift towards a pair of perforated plates, or grids, held at different voltages. The first grid, at a high positive voltage, draws them out of the chamber; the second, at a negative voltage, accelerates them to about thirty kilometres per second as they pass through its holes. Third, a neutraliser releases electrons into the exhaust so that the spacecraft does not build up an unwanted electric charge. Xenon is chosen because it is heavy and chemically inert.

The Hall-effect thruster achieves something similar by a different route. Here a magnetic field is arranged so that electrons cannot travel straight to the anode but are forced to circulate in a ring, and this cloud of trapped electrons creates an electric field along the axis of the engine. Ions passing through that field are accelerated without any grids at all. The design is simpler and can produce more thrust for its size, though its exhaust is slower and its efficiency somewhat lower than that of a gridded engine. Hall thrusters are now the workhorses of the commercial satellite industry, used to hold a spacecraft in its allotted position and, increasingly, to raise it there from a lower orbit after launch.

Other designs are emerging. Electrospray thrusters accelerate charged droplets of an ionic liquid rather than individual atoms, and produce thrusts of a fraction of a millinewton; they are being developed for CubeSats, satellites not much larger than a shoebox. Engineers are also experimenting with propellants. Xenon is expensive and must be stored under pressure in heavy tanks, whereas iodine can be kept as a solid and vaporised when needed. Argon, though less efficient, is cheap and plentiful.

The idea is old. Robert Goddard sketched the electrostatic acceleration of charged particles in his notebooks as early as 1906, and Hermann Oberth discussed electric propulsion in a book published in 1929. It was not until 1964, however, that an ion engine operated in space, on a small NASA test spacecraft called SERT-1. The technology reached a wider public in 1998, when Deep Space 1 set off to visit an asteroid and then a comet with a single ion engine that ran for many thousands of hours. Dawn, launched in 2007, used three such engines and a large supply of xenon to become the first spacecraft to orbit two separate bodies, Vesta and Ceres. BepiColombo, on its way to Mercury, relies on solar-electric propulsion.

The limitations are real. Thrust is proportional to electrical power, and power in space is scarce: Dawn's solar arrays generated more than ten kilowatts near Earth but barely one kilowatt at Ceres, and its engines were fired only one at a time. Because the push is so feeble, ion engines cannot lift a spacecraft off the ground; they are useful only after a rocket has placed the payload in orbit. They also work slowly, so a manoeuvre that a chemical engine would complete in minutes may take months of continuous firing. The grids erode under bombardment by ions, which limits an engine's working life, and the propellant, though used sparingly, eventually runs out.

Research therefore concentrates on power and propellant. Engineers are designing thrusters of a hundred kilowatts and more, intended to haul cargo slowly but cheaply between Earth and the Moon or Mars. Others are working on air-breathing engines that would collect the few gas molecules remaining at the top of the atmosphere and use them as propellant, allowing a satellite to stay in very low orbit for years without carrying fuel. VASIMR, a design that heats plasma with radio waves and expels it through a magnetic nozzle, has been tested in vacuum chambers on the ground but has yet to fly. What all these projects share is the same bargain: patience in exchange for mass.

Artificial intelligence has produced an unusual argument. On one side are researchers who warn that systems of growing capability may eventually pursue goals of their own, at the expense of the people who built them, and that the outcome could be catastrophic. On the other are critics who regard such warnings as a distraction from harms that algorithms are already causing: biased decisions, unwarranted surveillance, the concentration of power in a handful of companies. Public debate tends to present these as rival positions, between which a sensible person must choose. I believe that framing is mistaken, and that it has distorted both the research agenda and the laws that are now being written.

The technical core of the first position is the alignment problem. A learning system is given an objective — a mathematical score to maximise — and it becomes extremely good at maximising it. What it cannot do is infer the objective its designers had in mind. Suppose a household robot is rewarded for the number of dust particles it collects. Nothing in that reward tells it not to scatter dust from a cupboard onto the floor in order to collect it again. Computer scientists call this reward hacking, and it is a special case of a general truth known as Goodhart's law: once a measure becomes the target, it ceases to be a good measure.

Harms of this kind are not hypothetical. Systems that rank job applicants, assess the risk that a defendant will reoffend or allocate social security payments have all been deployed, and several have been shown to reproduce the prejudices present in their training data; one large technology company abandoned an experimental recruitment tool after finding that it downgraded CVs from women. A model's reasoning is usually opaque even to its creators, so a rejected applicant cannot be told why the decision went against them, and a regulator cannot inspect the logic that produced it. Meanwhile the computing power needed to train the largest systems is concentrated in a few firms, which gives those firms an influence over public life that no electorate has agreed to.

The second position rests on a different claim: that as systems become more capable, a small misalignment will matter more, until eventually it matters absolutely. A system pursuing almost any goal, the argument runs, has reason to acquire resources, to prevent anyone from switching it off and to resist changes to its objective, because these are useful whatever the goal happens to be. Nobody could test such a scenario in advance. We would have only one attempt, and the attempt would be made on systems whose behaviour we do not fully understand. If the potential harm is unlimited, the argument concludes, even a small probability deserves serious attention — a form of reasoning familiar from insurance.

This reasoning is difficult to fault in the abstract, and almost impossible to assess in practice, because it depends on forecasts that nobody can check. Predictions about the date at which machines will match human performance in some general sense have a poor record; the experts disagree with one another, and the disagreement has not narrowed as the technology has advanced. A claim about a risk that has not yet materialised, made by people who cannot demonstrate the mechanism, is not the kind of proposition that evidence can settle. That does not make it false. It makes it a matter of judgement, and judgement should be argued for rather than asserted.

None of this means the warnings should be dismissed. The research they have provoked is valuable whatever one thinks about the timeline: methods for testing a system's behaviour before release, for understanding what is happening inside a neural network rather than merely observing its outputs, and for keeping a human being in a position to intervene. Interpretability, as the second of these is called, is already helping engineers to find the reasons for errors in translation software and medical imaging. Work on oversight has begun to shape how companies document their models. The mistake is not to study catastrophic risk; it is to let that study set the whole agenda.

Governance, by contrast, has moved slowly and unevenly. The European Union's AI Act, agreed in 2023 and brought into force in stages, classifies applications by risk and imposes the heaviest obligations on those used in hiring, education, policing and essential services. The United States has relied mainly on voluntary frameworks. In November 2023, at a summit held at Bletchley Park in England, governments and companies signed a declaration acknowledging the dangers, and an earlier open letter from technology leaders had called for a pause in training the largest models. Voluntary gestures of this kind are welcome, but they cannot bind anyone. Only mandatory evaluation, disclosure and, where harm occurs, liability can do that.

My own view is that the gravest danger is less exotic than either camp suggests. It is not a machine that decides to destroy us; it is an institution that adopts a system it does not understand, for reasons of economy, and discovers too late that the system cannot be audited, corrected or switched off. That danger is with us now, and it grows with every deployment. The long-term question deserves a serious, well-funded and independent community of researchers. It does not deserve to monopolise the argument. The test of policy in this field should be whether a decision can be explained to the person it affects, and whether anyone can be held responsible when it goes wrong.