IELTS Academic Reading Practice Test 110

Few manufactured objects are as commonplace as the ballpoint pen, and few are so casually disregarded. Yet an instrument that costs a few pence in any stationery shop resolves a technical difficulty that defeated inventors for half a century. To write legibly, a pen must release its ink in a thin, uniform film; the ink must dry quickly enough to avoid smudging; and it must not escape when the instrument is inverted, agitated or transported inside a pocket. The fountain pen satisfied the first requirement imperfectly and the remaining two not at all.

The earliest patent for a pen using a small ball to regulate the delivery of ink was granted to John J. Loud, an American leather tanner, in 1888. His apparatus was designed for marking leather rather than paper, and the ink reached the surface through a small steel sphere held in a metal socket at the tip. As the sphere rotated, it accumulated ink from the reservoir behind it and deposited a trace on the material beneath. The idea was ingenious, but the engineering of the period could not support it.

Loud's apparatus performed poorly on paper. The ink leaked around the sphere and stained the writer's fingers, and the resulting line was coarse and irregular. Leather could tolerate such defects; a commercial letter could not. No manufacturer showed any interest, the patent lapsed, and for the following five decades the fountain pen, with its nib and its reservoir of aqueous ink, remained the standard instrument of written communication.

The difficulty was confronted again in Hungary during the 1930s by László Bíró, a journalist who resented the time consumed in filling pens and the smudges they deposited on inexpensive paper. He had observed that the ink employed in printing newspapers dried almost instantaneously because it was thick and oily rather than watery. A pen utilising ink of that consistency would require an alternative mechanism to deliver it: a narrow channel supplying a tiny ball, which would rotate across the page and leave a thin, uniform line. Bíró submitted a patent in Hungary in 1938, and his brother György, a chemist, formulated an ink that would traverse the ball without obstructing it.

The achievement of the design depends upon a delicate equilibrium. The sphere must rotate freely, so that it collects ink from the reservoir and deposits it upon the paper, yet the socket in which it is seated must be sufficiently tight to prevent the ink escaping around its perimeter. Because the interval between sphere and socket is measured in hundredths of a millimetre, the pen demands precision machining, and the ink must possess the appropriate viscosity: sufficiently fluid to be drawn into the gap, sufficiently dense not to run.

By the early 1940s the brothers had departed from Europe for Argentina, where in 1943 they set up a factory and began selling instruments that became known locally as birome. The design soon acquired a military application. The Royal Air Force ordered the pens in substantial quantities and issued them to navigators, whose duties required an implement that would write upon a chart inside a cramped, unheated aircraft. At high altitude the reduced atmospheric pressure caused the ink in a fountain pen to escape and stain the paper, whereas the sealed reservoir of a ballpoint remained unaffected. The contract was lucrative, and it financed further refinement of both the ink and the machining of the socket.

Peace brought a commercial scramble. In 1945 an American businessman, Milton Reynolds, purchased several pens in Buenos Aires, returned to Chicago and began manufacturing copies at speed. Reynolds had invented nothing; he had reproduced an existing product. His timing, however, was exemplary. The pens went on sale in a New York department store at $12.50 each, and the first consignment was reportedly exhausted within hours. Rival corporations inundated the market, the quality of the imitations declined, purchasers lost confidence and the boom collapsed.

What survived the collapse was a superior product. In 1945 the French manufacturer Marcel Bich bought a factory at Clichy, outside Paris, and after acquiring the rights to Bíró's design he launched his company's own pen, the Cristal, in 1950. It was inexpensive, its transparent polystyrene barrel permitted the user to observe how much ink remained, and a hexagonal section prevented it from rolling off the desk. Equally important, it was designed for machinery rather than craftsmen, so every component could be produced in enormous quantities at a minimal unit cost.

Inexpensive pens transformed the classroom as thoroughly as the office. By the 1960s a child could purchase a ballpoint for a few pence, and schools progressively abandoned the inkwell, the dip pen and the blotting paper associated with them. Certain teachers disapproved of the change. They contended that the new instruments encouraged a hasty, spidery hand, and that pupils who wrote rapidly, with no nib to restrain them, were forfeiting the older cursive scripts. The time allocated to formal handwriting instruction was diminished, although it was never eliminated altogether.

The ballpoint is now discarded rather than refilled, and its plastic casing has become an emblem of the era of disposable objects. Yet its principle — a rotating sphere, a sealed reservoir, an ink that dries upon contact — has scarcely altered in seventy years, and the ballpoint remains the most extensively used writing instrument in the world.

When a person speaks into a phone and the words appear on the screen a moment later, an elaborate chain of calculations has taken place. Speech recognition, sometimes called automatic speech recognition or ASR, is the task of converting an acoustic signal into a sequence of words. The problem is harder than it looks. Writing separates words with spaces, but speech does not: a sentence spoken at normal speed is a continuous stream of sound in which the end of one word and the beginning of the next merge together. Speakers also differ enormously in accent, speed and volume, and the same word may sound quite different in two mouths. Systems must therefore combine knowledge of sound with knowledge of language.

The first stage is to turn the pressure wave into numbers. A microphone samples the incoming sound thousands of times a second, typically 16,000 times for speech, and each sample is stored as a value representing the loudness of the signal at that instant. The stream is then divided into short frames, usually around 25 milliseconds long, which overlap so that no information is lost at the boundaries. Each frame is analysed to describe how its energy is distributed across different frequencies. The result is a compact numerical summary, a kind of fingerprint of that sliver of sound, and it is these summaries rather than the raw audio that the rest of the system works on.

The next stage is to decide which speech sounds, or phonemes, those summaries contain. English has around forty phonemes, but their acoustic realisation is far from fixed: the sound at the start of 'top' is not the sound in 'stop'. A phoneme is coloured by the sounds on either side of it, a phenomenon known as coarticulation. For decades, engineers modelled this variability statistically, using a family of mathematical tools called hidden Markov models, which describe the probability of moving from one sound to another over time. From about 2010, these were largely replaced by deep neural networks, which learn the mapping from acoustic features to sounds by examining very large quantities of transcribed speech.

Sound alone is not enough, because many different sentences would produce similar acoustics. A language model supplies the missing knowledge. It is trained on enormous collections of text and captures which sequences of words are probable: that 'I would like to book a table' is a far more likely sequence than 'I would like to book a stable'. Modern language models also consider the words already recognised in the sentence, so that an ambiguous sound can be resolved by the context that precedes it. The decoder then searches for the single word sequence that best satisfies both the acoustic evidence and the expectations of the language model.

Recently the boundary between these components has blurred. End-to-end systems take the acoustic features at one end and produce letters or words at the other, learning the whole mapping in a single network without separate acoustic and language models. Such systems need vast amounts of audio paired with accurate transcripts, but they can be adapted to a new language with less expert knowledge of its sound system. Another line of research trains networks on unlabelled audio, teaching them to predict the parts of a recording that have been hidden, and then fine-tuning them on a much smaller set of transcribed speech.

The applications have multiplied. Voice assistants on phones and in homes use recognition to answer questions and control devices; call centres use it to route customers and to search recordings; broadcasters use it to generate subtitles; and clinicians use it to dictate notes, which reduces the time they spend at a keyboard. For people with limited movement in their hands, recognition may be the main way of using a computer. On-device systems, which run on the phone rather than in a distant data centre, have become common because they respond faster and keep the audio away from the network.

Performance is nevertheless uneven. Recognition in a quiet room, close to a good microphone, can be excellent for a standard accent, but accuracy falls with background noise, distance from the microphone and reverberation in a large room. Spontaneous conversation, full of hesitations, false starts and overlapping turns, is much harder than read speech. Systems also differ in how well they serve different groups of people. A study published in 2020 tested commercial systems on American English and found substantially higher error rates for Black speakers than for white speakers, a disparity its authors linked to the data used in training.

Current research addresses these weaknesses. One goal is to separate and follow individual voices when several people talk at once, a problem often called the cocktail party effect. Another is to make systems work well in languages with few written resources, where large transcribed corpora do not exist. Researchers are also trying to give machines a sense of when they have misheard, so that a system can ask for clarification instead of confidently producing nonsense. One further aim is to reduce the energy such systems consume, since training a large model carries an environmental cost. Progress in all these areas depends less on clever algorithms than on the data available to train them.

Few arguments about technology are repeated as often, and examined as little, as the claim that people who have nothing to hide have nothing to fear from surveillance. The line is effective in conversation because it reframes privacy as concealment: if you are doing nothing wrong, the reasoning goes, you cannot be harmed by being watched. That description fits only a small part of what privacy protects. Most of the information people try to keep to themselves is not shameful at all. We do not announce our salary to colleagues or our medical history to neighbours, not because either is a crime, but because information given in one setting is not meant for another.

What digital systems collect is not a set of confessions but a record of ordinary behaviour. A phone notes where its owner has been, which applications were opened and for how long, and which advertisements were passed over. A browser keeps a history of pages visited, and the sites themselves pass information to advertising networks that few users have heard of. Individually these fragments look trivial. Combined, they become a portrait: patterns of sleep, religious observance, financial worry and illness can all be inferred from behaviour that nobody would describe as private in isolation.

The benefits of such collection are real, and it would be dishonest to deny them. Banks examine patterns of transactions to detect fraud within seconds. Medical researchers use large collections of patient records to identify side effects that trials are too small to reveal. Traffic data helps cities to manage congestion, and recommendation systems, however irritating, save people time. Some of the most valuable uses of personal data are ones that individuals would enthusiastically agree to if they were asked clearly. The difficulty is that the same databases that support these services also support purposes their subjects never imagined.

Consent is the mechanism that most legal systems rely on, and it is failing. The notice-and-consent model assumes that a person can weigh the consequences of a transaction and make a rational choice. In practice nobody reads the terms attached to a website, and nobody could: the documents are long, the language is impenetrable, and the choices are rarely genuine, since refusing means losing access to a service that friends and employers expect you to use. Consent obtained in these conditions is not a meaningful expression of preference but a ritual that transfers responsibility from the organisation to the individual.

The comforting assumption that removing names renders data harmless is also unsafe. Researchers have repeatedly demonstrated that supposedly anonymous records can be re-identified by matching them against other sources. A dataset that contains no names may still contain a postcode, a date of birth and a job title, and that combination is often enough to single out one person. Because the risk grows with every new dataset published, anonymisation is best understood as a temporary measure rather than a guarantee.

Legislation has begun to respond. The European Union's General Data Protection Regulation, which came into force in 2018, gave individuals rights to see the information held about them and, in some circumstances, to have it erased; it also allowed regulators to impose fines of up to four per cent of a company's global turnover. Courts have extended the principle: a ruling by the European Court of Justice in 2014 required search engines to remove links to outdated material on request. Yet enforcement has been slow, and the largest penalties have been modest beside the revenues of the firms concerned.

Critics of regulation make two serious points. The first is that compliance falls hardest on small companies, which cannot afford lawyers, while large firms absorb the cost and gain an advantage. The second is that restrictions on the use of data may slow research that saves lives. Both deserve an answer. Rules can be designed to distinguish between a start-up and a multinational, and medical research already operates under ethical approval systems that permit the use of sensitive data with safeguards. Neither objection, in my view, justifies the current arrangement, in which the burden of protecting information rests on the person least able to bear it.

A better approach begins with collecting less. Data minimisation, the principle that an organisation should gather only what it genuinely needs, and purpose limitation, the principle that information gathered for one use should not quietly be repurposed for another, are more effective than any consent form. Independent audits, and a duty to explain in plain language what is held, would give the public something that notices never do: a reason to trust. Privacy, in this view, is not a luxury for the secretive but a condition of ordinary social life, as necessary as clean air and as easily damaged by others.

The question, then, is not whether people should be able to hide, but who decides what is known about them, and for how long. A society that answers 'the highest bidder' will find that trust, once lost, is expensive to rebuild. Regulation and good design will not abolish the collection of personal data, and should not try to. They can, however, keep it within limits that the people concerned would recognise as fair.