IELTS Academic Reading Practice Test 134

In 1946 the American astronomer Lyman Spitzer published a paper arguing that a telescope placed above the Earth's atmosphere would transform astronomy. A telescope on the ground, he pointed out, looks at the sky through turbulent air that blurs starlight and absorbs much of the radiation arriving from space; ultraviolet light, in particular, never reaches the surface. An instrument in orbit would be limited only by the size of its mirror. At the time no rocket could lift such a machine, and for three decades the idea remained a theoretical ambition.

The idea began to seem practical in the 1970s. NASA, working with the European Space Agency, drew up plans for what was then called the Large Space Telescope, and in 1977 the United States Congress approved the money to build it. The project was later renamed in honour of Edwin Hubble, the astronomer who in the late 1920s had shown that the universe is expanding. Its main mirror, 2.4 metres across, had to be ground so accurately that no point on it strayed from the intended curve by more than a fraction of the wavelength of visible light.

Hubble was carried into orbit on 24 April 1990 by the space shuttle Discovery and released about 600 kilometres above the ground. That altitude places it above the great majority of the atmosphere but far below the belts of radiation that surround the planet, and it circles the Earth roughly every 95 minutes. Being close to home had one enormous advantage: astronauts could reach it. Hubble was designed from the outset to be visited, repaired and upgraded.

Success did not come immediately. The first images were disappointing, showing stars surrounded by haloes rather than as sharp points of light. The cause was a defect in the main mirror known as spherical aberration. The surface had been polished with extraordinary care, yet it was too flat at its edges by about two micrometres, roughly a fiftieth of the width of a human hair. The error originated in a measuring device used during manufacture that had been put together incorrectly, and it escaped detection because the mirror was never tested in a complete telescope before launch.

Fortunately the flaw was understood precisely enough to be corrected. In December 1993 a shuttle crew carried out five spacewalks to install a new camera and a package of corrective optics called COSTAR, which used small adjustable mirrors to cancel out the error before light reached the detectors. The images released the following January were strikingly sharp, and the new camera soon produced a celebrated picture of a galaxy with a bright core and dark bands of dust.

What makes Hubble valuable is not the size of its mirror — several telescopes on the ground have larger ones — but the quality of the view. Without an atmosphere to distort the light, it can distinguish detail about ten times finer than a comparable instrument on the ground. It can also observe in the ultraviolet, which the atmosphere blocks, and in the near-infrared. Astronomers compete for time on it, and most proposals are rejected.

Its results have ranged across almost every branch of astronomy. By measuring how fast stars and gas swirl around the centres of nearby galaxies, Hubble provided some of the strongest evidence that supermassive black holes, millions or billions of times the mass of the Sun, occupy the middle of most galaxies. In July 1994 it watched the fragments of comet Shoemaker–Levy 9 crash into Jupiter, leaving dark scars in the planet's clouds. Its deep-field images, made by pointing at apparently empty sky for many hours, revealed thousands of galaxies in each, some of them seen as they were billions of years ago.

The telescope also helped to settle the age of the universe. Measurements of how quickly galaxies are receding pointed to an age of about 13.8 billion years, and observations of distant exploding stars showed that the expansion of the universe is speeding up rather than slowing down — the finding that led astronomers to the concept of dark energy. Closer to home, its instruments have examined the atmospheres of planets orbiting other stars.

Five servicing missions visited Hubble between 1993 and 2009. The last installed two new instruments and replaced the gyroscopes that keep the telescope pointed steadily. Since the shuttle fleet was retired in 2011 no further visits have been possible, and the orbit is slowly decaying. Its successor, the James Webb Space Telescope, was launched on 25 December 2021 and observes mainly in the infrared from 1.5 million kilometres away. Hubble, however, remains a general-purpose observatory that any astronomer may apply to use.

A radiologist beginning a shift may face several hundred images: chest X-rays, mammograms, scans of the brain. Demand for such examinations is rising in every health system, while the number of trained specialists grows slowly, and in many countries the gap between the two has become a crisis. It is hardly surprising that medical imaging has become the most active testing ground for artificial intelligence in medicine. The promise is not that machines will replace radiologists, but that they might shorten the queue and reduce the number of things missed.

Most of the systems now being tested belong to a family of programs called convolutional neural networks. A network of this kind is not given rules; it is shown examples. During training, thousands of images are fed in, each accompanied by a label — the diagnosis a specialist has assigned to it. The program adjusts millions of internal settings until its outputs come close to matching those labels. Its early layers learn to respond to simple features such as edges and patches of light or dark; later layers combine these into textures, shapes and, eventually, structures that resemble tumours or haemorrhages. The trained system then reports a probability rather than a verdict: it might state, for instance, that an image has a 0.83 chance of showing disease.

The approach became practical after 2012, when a deep network entered an international image-recognition contest and beat every rival by a wide margin. Medicine, however, presented a harder problem than photographs of cats and cars, because the labels a network must learn from can be supplied only by clinicians. Assembling a large training set therefore costs both time and money. Some of the most influential datasets have been built from screening programmes, where hundreds of thousands of images were already stored alongside their eventual diagnoses. Several large collections of chest X-rays have also been released by public institutions, allowing research groups without access to hospital records to compare their methods.

The first system authorised for independent use arrived in 2018, when the US Food and Drug Administration allowed a program for detecting diabetic retinopathy, a complication of diabetes and a leading cause of blindness among working-age adults. A technician photographs the back of the eye with a special camera, and the software analyses the image and decides whether more than mild disease is present. If it is not, the patient can be given a follow-up appointment without a specialist ever seeing the picture. Where ophthalmologists are scarce, that saves an entire visit.

Other applications are close behind. Screening mammography, in which hundreds of images are read for every cancer discovered, is an obvious target. A large study published in 2020 reported that a deep-learning system matched or exceeded the performance of human readers on mammograms from the United Kingdom and the United States, though the authors noted that the comparison was retrospective, using stored images rather than patients in a clinic. Systems that read chest X-rays for signs of tuberculosis are being tested in countries where radiologists are fewest.

The limitations, however, are serious. Networks are prone to what researchers call shortcut learning. A model trained to detect pneumonia may learn instead to recognise the portable scanners used on the sickest patients, or the small letters one hospital prints on its images, because those features happen to accompany the diagnosis in the training data. Such a model performs brilliantly on the data it was trained on and badly anywhere else. Because hospitals differ in their equipment, their imaging protocols and their patients, the gap between laboratory results and real-world performance is often wide.

A second difficulty is that the labels themselves are unreliable. Two experienced radiologists looking at the same scan frequently disagree, and a network trained to imitate one of them inherits that person's habits along with their expertise. Rare conditions are, by definition, rare in the training data, so the cases where a second opinion is most valuable are the ones the program has seen least often. Careful curation of datasets, and the reporting of how each was assembled, matter as much as the architecture of the network itself.

There is also the question of explanation. Clinicians are understandably reluctant to act on advice they cannot interrogate, and the techniques used to show which part of an image influenced a decision have been criticised as unreliable: a map of the regions a network looked at can appear convincing while pointing at the wrong thing. Regulators have responded by asking for evidence from prospective trials, in which the software is tested on new patients and its effect on outcomes is measured, rather than retrospective accuracy alone. Such trials are expensive and, so far, uncommon.

The next generation of systems may need far fewer labels. Models trained on very large collections of unlabelled images, using a technique known as self-supervised learning, can then be adapted to a specific task with only a small set of expert annotations. Another line of work, federated learning, allows a model to be trained across several hospitals without any images leaving them, which addresses both privacy and the reluctance of institutions to surrender their data. Neither approach removes the need for careful evaluation, but both make larger and more varied training sets possible.

The concert has just ended and it is raining. Opening a ride-hailing app, you are offered a car home at two and a half times the fare you paid on the way out, and a small note explains that prices are high because many people are asking for cars. No human being decided on that figure. An algorithm, watching the number of requests and the number of drivers available in each small area of the city, raised it within seconds. Such dynamic pricing is now ordinary in airlines, hotels, electricity supply and ticket sales, yet it is the ride-hailing apps that have made it visible, and controversial, to millions of people.

The economic case for it is old and simple. When a price is fixed, a sudden rise in demand produces a shortage: cars are all occupied, would-be passengers wait, and some give up. Raising the price does two jobs at once. It persuades people whose journey is not urgent to postpone or abandon it, and it draws drivers from quieter districts towards the crowds. Airlines have long sold seats on the same flight at very different fares for precisely this reason, and electricity suppliers have begun to charge households less at times of day when the grid is under strain, so that demand shifts away from the peak.

Against this, critics raise the question of fairness. A person leaving a flooded street is not a shopper comparing brands; urgency, not wealth, has put them in the market, and a price that rises with their desperation looks like exploitation. Several American states have laws against excessive pricing during declared emergencies, and the argument that such laws protect the vulnerable is widely accepted. Willingness to pay, the critics point out, measures the depth of a wallet as much as the strength of a need.

Defenders of surge pricing reply that a cap does not conjure up cars. If the fare is held down during a storm, the queue lengthens, drivers who might have come out stay at home, and some passengers wait for hours or set off on foot. On this view a high price is not a penalty imposed on the desperate but a reward offered to whoever is willing to drive through the weather. There is force in this argument, and I accept it as far as it goes. It does not, however, go as far as the companies would like.

The first reason is information. A platform knows the demand in every street, the number of cars nearby and the fare it charged a minute ago; the customer knows only the number on the screen. When a price can be varied for each individual, using what is known about their device, their history and their location, the ordinary discipline of a visible market disappears. A second reason is competition. In many cities a handful of firms dominate the local market, and a new entrant faces costs that make it hard to survive. Where rivals use the same software to set prices, moreover, fares may drift upwards without anyone having agreed to anything.

Regulators in several countries have begun to examine pricing algorithms for this reason, concerned less with a secret agreement in a closed room than with the possibility that similar code, fed similar data, will produce similar prices. Such tacit coordination is difficult to prove and harder still to police, since no message need ever pass between the companies. The revenue from a surcharge, meanwhile, is divided between platform and driver on terms that are rarely made public.

My own position is that the practice should be permitted, because the alternative — a fixed fare and an empty street — serves nobody, but that permission should carry conditions. The first is disclosure: the total price, including the multiplier that produced it, should be shown before the customer confirms a booking, not added at the final screen. The second is a limit. When an emergency has been officially declared, a maximum fare should apply, set high enough to keep drivers working and low enough to prevent the worst of the exploitation the critics describe.

Further rules follow from the same logic. Prices should not be tailored to the individual's apparent willingness to pay; a fare that depends on the phone you own or the neighbourhood you live in is a different thing from a fare that responds to traffic. Companies should publish the rules their algorithms apply, and regulators should have the power to inspect shared pricing software for signs of tacit collusion. A plainly advertised basic service, at a fixed fare, should be available to those who need it.

None of this is an argument for leaving prices frozen. A price that moves is a signal, and signals coordinate behaviour in ways that queues cannot. The real question is not whether prices should change with demand, but who sets them, whether we can see how they are set, and where the movement must stop. A dynamic fare that is disclosed and bounded is a useful instrument; one that is hidden and unbounded is simply a charge on the unlucky.