IELTS Academic Reading Practice Test 166

For most of history, making a tunnel meant digging by hand. Labourers cut the ground with picks, propped the roof with timber and broke hard rock with gunpowder. The work was slow and dangerous, but the real difficulty was the ground itself: as soon as a hole is opened, the earth above presses down and water seeps in, and unless the face is held steady the excavation closes up again. The invention that changed this was not a digging tool but a movable support.

The engineer who solved the problem was Marc Brunel, a French-born inventor living in London. His idea, he later said, came from the shipworm, a mollusc that bores into timber with two small shells at its head. As it scrapes away the wood, it lines the hole behind it with a hard tube, so that the tunnel is always supported. Brunel patented a similar arrangement in 1818. His shield was an iron frame divided into thirty-six compartments. A worker inside one compartment cut away a few centimetres of clay; the compartment was then closed and the frame pushed forward by jacks bearing on the finished brickwork behind it.

Work on the Thames Tunnel between Rotherhithe and Wapping began in 1825. It was to be the first tunnel beneath a navigable river. The ground below the Thames was saturated and treacherous, and the tunnel flooded twice, in 1827 and 1828, when the river broke through the roof and several men were killed. Brunel's son, Isambard Kingdom Brunel, who was in charge of the work, was nearly drowned. The tunnel opened in 1843, eighteen years after work began and at a cost far above the original estimate, and it carried pedestrians rather than the carriages for which it had been intended.

In 1870 a small circular tunnel under the Thames at Tower Hill was driven by James Henry Greathead, whose shield was round rather than rectangular. A circular shield suits soft ground, and it can be rolled forward on an iron plate. Cast-iron rings, bolted together behind it, formed a permanent lining. With compressed air to hold back the water, the shield made deep tunnelling feasible, and the City and South London Railway, opened in 1890, was the first railway to run in tunnels deep beneath a city.

Rock was a different matter. Cutting hard stone needed motors, hardened edges and great power, none of which existed in the middle of the nineteenth century. The first serious attempt was made in the United States, where a machine designed by Charles Wilson was set to work on the Hoosac Tunnel in Massachusetts in 1853. Weighing many tonnes and driven by steam, it was supposed to groove the rock so that it broke away in slabs. Its cutters wore out, the machine jammed repeatedly, and it was abandoned; the tunnel was finished with explosives.

The decisive step was taken by James S. Robbins, whose company built a machine in 1952 for a dam on the Missouri River in South Dakota. His machine carried a rotating head fitted with hardened steel discs, each pressed against the face with enormous force. As the head turned, the discs rolled across the rock, crushing the stone beneath them and cracking it between the grooves. Rock could then be removed in chips rather than ground to powder. Disc cutters of this kind remain the standard method of boring hard rock.

Soft ground demanded different solutions. In clay and sand the face collapses unless it is held up, and here the key idea came from Japan, where engineers developed the earth pressure balance machine in the 1960s and 1970s. The machine mixes the excavated soil with foam and keeps it under pressure in a chamber behind the cutterhead, so that the paste pushes back against the ground ahead. A screw conveyor then removes the paste at a controlled rate. Slurry shields use a similar principle with a liquid: a suspension of clay, usually bentonite, is pumped into the chamber and circulated out with the spoil.

The Channel Tunnel, dug between 1987 and 1991, showed what the technology could do. Eleven machines worked from the British and French coasts, boring three tunnels through chalk that was soft enough to cut but firm enough to stand without much support. They were driven forward at rates that sometimes exceeded seventy metres a day. When the headings met beneath the sea, they joined almost exactly. The British machines that bored the running tunnels were then buried where they stood, since removing them would have cost more than leaving them there.

The Gotthard Base Tunnel in Switzerland, opened in 2016, carried the technology through harder rock and greater depths. Its 57 kilometres made it the longest railway tunnel in the world, and for much of its length the workers laboured beneath more than two thousand metres of mountain. Temperatures in the rock reached more than 40 °C, and sudden inflows of water under high pressure were a constant danger.

Today's machines are among the largest pieces of equipment ever moved on land. Heads more than fifteen metres in diameter are used for road and rail tunnels. Modern shields are steered by laser guidance, the concrete segments of the lining are erected by robotic arms inside the shield, and sensors record the pressure at the face many times a second. Yet the basic principle has not changed since 1818: hold the ground steady, cut a little at a time, and line the hole as you go.

Augmented reality, or AR, is the practice of adding computer-generated images to a live view of the physical world, so that the two appear to occupy the same space. It differs from virtual reality, which replaces the surroundings entirely; in AR the real world remains visible and the digital content is layered on top of it. A widely used definition, published by the researcher Ronald Azuma in 1997, sets three conditions: a system must combine real and virtual objects, must work interactively in real time, and must register the virtual objects in three dimensions, so that they stay fixed to the scene as the viewer moves.

The first head-mounted display was built in 1968 by Ivan Sutherland, who suspended a pair of small screens from the ceiling above the wearer's head because the optics were far too heavy to wear. The term 'augmented reality' was coined in 1990 by Tom Caudell, a researcher at the aircraft manufacturer Boeing, who was looking for a way to replace the large plywood boards on which workers checked their wiring diagrams. In 1999 Hirokazu Kato released a free toolkit that allowed anyone to track printed squares with an ordinary camera, and the field opened to universities and hobbyists. Consumer products followed more slowly: Google Glass was sold to selected testers from 2013, and Microsoft's HoloLens, a headset that displays three-dimensional images, appeared in 2016.

To place a virtual object convincingly, a headset must know where it is and which way it is pointing, and it must update that knowledge many times a second. The usual approach combines two sources of information: images from a small camera and readings from inertial sensors, the accelerometers and gyroscopes that measure acceleration and rotation. The camera loses track when the scene is plain or badly lit, and the inertial sensors drift within seconds as small errors accumulate, so software fuses the two, using the camera to correct the sensors and the sensors to bridge the gaps between camera frames. In an unfamiliar room the system may also perform simultaneous localisation and mapping, usually shortened to SLAM, building a map of its surroundings while working out its own position within that map. Early systems avoided this by placing printed markers around the space.

There are two ways of mixing computer images with the real world. Video see-through systems capture the scene with a camera and show the mixture on a screen that blocks out the surroundings, as a smartphone does. Optical see-through systems, used in most headsets, allow light from the real world to pass through a transparent combiner while the virtual image is projected into the same eye. The projection is generally channelled by a waveguide, a thin slab of glass in which light is trapped by total internal reflection and steered towards the eye by microscopic structures etched on its surface.

Two difficulties are particularly stubborn. The first is latency, the delay between a movement of the head and the appearance of the adjusted image. If the virtual object arrives even a few thousandths of a second late, the brain registers it as swimming behind reality, and designers therefore aim for a total delay of about twenty milliseconds. The second is the vergence-accommodation conflict. To see an object sharply, the eyes must both point at it and adjust their lenses, and in the real world these two responses agree because both depend on distance. In a headset the virtual image is usually projected at a fixed distance of a metre or two, while the eyes may be converging on something that appears much closer or much further away, and the mismatch can cause eyestrain, headaches and nausea.

Industry has been quicker to adopt the technology than consumers. Boeing technicians used early systems to see wiring diagrams laid over the aircraft they were assembling. Some surgeons now operate wearing headsets that show scans of the patient beside the operating table, so that they do not have to look away to a screen. Warehouse staff use smart glasses to find goods on shelves, and engineers use them for remote support. For the general public the breakthrough came in 2016, when a phone game in which animated creatures appeared through the camera made AR familiar to millions.

Field of view remains the most visible limitation. In the first generation of headsets the virtual image occupied only a small rectangle in the middle of the wearer's vision, so that digital content disappeared whenever the head turned. Weight and battery life are linked: a device powerful enough to run for a full shift needs a large battery, which adds both weight and heat, while a lighter battery means fewer hours of use. Social acceptability is a further obstacle, since a person wearing a camera on their face may be thought intrusive, and some workplaces restrict such devices on privacy grounds.

Research is concentrating on making the optics smaller and the experience more comfortable. Holographic waveguides can bend light through a wider angle than conventional glass, widening the field of view. Retinal scanning displays, in which a tiny mirror directs a beam across the eye to draw the image directly on the retina, promise sharp pictures from a device the size of a pair of sunglasses. Eye tracking may help as well: by measuring where the wearer is looking, a system can render in full detail only the small area at the centre of gaze, saving power, and can adjust the projected distance to match the eye's focus.

For more than a decade, researchers and campaigners have travelled to United Nations meetings in Geneva with a simple message: machines should not be allowed to decide who dies. Their subject is the autonomous weapon, a system that, once activated, selects and attacks targets without a human being making the final decision. The debate is often described as a quarrel about whether to ban 'killer robots', a phrase that misleads in both directions: it suggests machines far cleverer than any that exist, and a choice between prohibition and doing nothing. My own view is that a legally binding prohibition is needed, but for a narrower category of weapon than campaigners imply.

The phrase at the centre of the discussion is 'meaningful human control', which at first sounds like a solution and turns out to be part of the problem. Everyone agrees that a human should be 'in the loop'; few agree on what the loop is. A commander who authorises a mission and then watches a screen while the system picks its targets has control of a kind, but not the kind that involves knowing why a particular person was attacked.

Meaningful control, on the most defensible account, requires a human to understand the situation in which the system will act, to foresee the range of targets it may select, and to have a real opportunity to stop it. Where any of these is missing, the human is present in name only.

Autonomy is already with us, in narrow forms. Naval close-in weapon systems such as Phalanx have for decades been able to switch to an automatic mode in which they detect, track and fire at incoming missiles without an operator's command, because no human can react quickly enough to a missile travelling at speed. Loitering munitions, which circle above an area before diving on a target they have identified, sit a little further from the same line. These examples show that the argument is not about whether to introduce autonomy but about where its limits should fall.

Those who resist a ban make two arguments. The first is precision. Sensors, they point out, do not panic or tire, and a machine may hold its aim more steadily than a frightened soldier. A system that identifies a tank correctly may kill fewer people than an artillery crew firing on a village. The claim deserves to be taken seriously, but it confuses two questions. Accuracy concerns how a target is struck; the laws of war are also concerned with who is chosen, and that is a question of judgement rather than of aim.

The second argument is that existing law is sufficient. International humanitarian law already requires armies to distinguish combatants from civilians, to avoid attacks whose civilian cost is excessive, and to take precautions. Nothing in these rules, it is said, demands a human finger on the trigger. The difficulty is accountability. If a system attacks a hospital, an inquiry must establish why, and the commander, the engineer and the officer who approved the software will each point to the others. A rule requiring a human decision at the point of attack would at least give the inquiry someone to question.

Then there is the risk that nothing is agreed. Weapons of this kind are attractive precisely because they are fast, and speed is competitive. A state that believes its adversary is developing autonomous air defence will want its own, and a state that fears it is falling behind may lower the threshold for using force. Because the software is small and the hardware is often adapted from civilian parts, controls on proliferation would be hard to enforce. The technology would not stay in the hands of responsible governments.

The proper forum for all this is the Convention on Certain Conventional Weapons, a United Nations treaty concluded in 1980 that already restricts or prohibits weapons such as those designed to blind. The convention has discussed autonomous weapons since 2013 and in 2017 established a Group of Governmental Experts, which meets in Geneva. It has not, however, been given a mandate to negotiate, and it operates by consensus, which means that any single delegation can prevent progress. Some states want a new protocol; others say that no new rule is needed.

Critics of a ban raise two practical objections: that autonomous weapons cannot be defined, and that a prohibition could not be verified. The 1995 protocol on blinding lasers banned a category of weapon that had never been used, and it did so by describing a purpose rather than a device. A prohibition could be drawn in the same way, around systems that select and attack targets without meaningful human control. The verification problem would then be smaller than for nuclear weapons, because the distinctive element is the software and the doctrine rather than a hidden material.

What, then, should be done? I would not ban sensors, nor the automatic modes that already protect ships at sea. The line should be drawn at systems that select and attack human beings without a human decision that is informed, deliberate and possible to reverse. That is a narrower prohibition than some campaigners want and a broader one than some governments will accept. But the alternative is a world in which the most consequential decision a person can make — whether to kill — is delegated by default, and no serious account of the laws of war can survive that.