Academic Reading Articles Practice 14 Test 02

Artificial intelligence has become a routine component of recruitment, particularly in large organisations that receive far more applications than human staff can reasonably read. Many employers now rely on Applicant Tracking Systems (ATS), software platforms that store candidate data, filter resumes, and help recruiters manage the hiring pipeline. When a single vacancy attracts hundreds or even thousands of applicants, an ATS can automatically remove duplicates, detect missing requirements, and rank resumes according to job-related keywords. Vendors increasingly promote “AI-powered” modules that go beyond sorting, claiming to predict job performance or retention by analysing patterns in previous successful hires. These tools are adopted largely to reduce time-to-hire and to keep recruitment manageable at scale, but their growing influence raises questions about how decisions are made and who bears responsibility when outcomes seem unfair.

A central ethical issue is bias in the data used to train or configure these systems. If an employer’s historical hiring records reflect unequal opportunities or past discrimination, statistical models may learn patterns that replicate those outcomes. Even when a company intends to be neutral, an algorithm trained on past choices may interpret those choices as a template for “fit.” Bias can also enter through proxies—variables that correlate with protected traits without naming them explicitly. For example, a zip code can correlate with race or socioeconomic status, and a school attended can reflect long-standing educational inequalities. Likewise, gaps in a resume may correlate with caregiving responsibilities, including maternity, and could be unfairly interpreted as a lack of commitment. In such cases, removing explicit demographic fields does not necessarily prevent discriminatory effects, because the system can still infer sensitive characteristics indirectly.

Another area of concern is the practical meaning of transparency. Hiring tools are often presented as objective, yet candidates may not know that an algorithm assessed them, which criteria were prioritised, or how to challenge an outcome. Even within organisations, the logic may be difficult to explain. Some systems operate like “black boxes”: they produce a score or ranking without a clear, human-readable justification. Complexity is not the only problem. Vendors may treat model details as proprietary, limiting what employers can disclose. As a result, transparency is not simply about announcing that AI is being used; it also involves documenting what the tool was designed to do, what it cannot do, and how decisions can be reviewed when an applicant believes an error occurred.

Automation can also reshape recruiter behaviour in ways that reinforce model preferences. When candidates are presented in an ordered list, staff may spend less time exploring those placed lower, especially under time pressure. This effect, sometimes described as automation bias, can make recruiters more complacent and more likely to accept the tool’s judgement as correct. Over time, this may create a feedback loop. If the model favours conventional career paths—continuous employment, familiar job titles, and standard qualifications—then fewer non-traditional applicants are shortlisted, and the next round of training data becomes even more skewed toward those profiles. Some organisations try to counteract this by requiring human review of a wider set of applicants, or by using AI as a triage tool rather than as a decision-maker. In these designs, staff are trained to treat model outputs as one input among several, not as final verdicts.

Privacy concerns have expanded as hiring systems have moved beyond resumes. Some tools analyse video interviews, tracking speech rate, word choice, or facial movements, and then infer traits such as confidence or sociability. Others scrape online profiles or use third-party data brokers to enrich candidate records. These practices raise questions about consent and proportionality. Even if candidates agree to an online interview, they may not realise that micro-expressions or voice features are being measured, or how those measurements will be interpreted. The risk is not only intrusion but also function creep: data collected for hiring can be repurposed later for unrelated uses, such as employee monitoring or marketing. There are additional technical risks, including re-identification of supposedly anonymised data, or leakage through insecure storage and sharing between vendors and employers.

Because job contexts change, ethical use also depends on validation and ongoing checks. A model that appears accurate in one setting may perform poorly in another if the role, labour market, or applicant pool changes. Skills demanded for a job can evolve, and hiring practices can shift after organisational restructuring, creating what data scientists call drift. Without monitoring, a system that once seemed reliable can become outdated or discriminatory. For this reason, experts recommend auditing: testing outcomes to see whether certain groups experience systematically different selection rates, examining error patterns, and checking whether the model’s criteria remain relevant to the job. Auditing can also identify when a tool is being used outside the conditions for which it was validated—for example, applying a model trained on one region or job family to a very different context.

Regulation is developing unevenly across jurisdictions. In some places, employers face specific obligations when using high-risk automated decision systems, such as conducting impact assessments or allowing candidates to request explanations. In other regions, oversight relies mainly on general anti-discrimination law, which can be difficult to apply when decision-making processes are opaque. Many companies also adopt voluntary standards, such as documenting data sources, conducting bias tests, and establishing review committees. However, voluntary measures vary in quality and may be implemented inconsistently, especially when hiring is outsourced to vendors. Stronger frameworks typically combine legal obligations with internal organisational processes, including clear responsibility for model use, independent review, and accessible channels for candidates to raise concerns.

Finally, debate continues over what “fair” hiring should mean in practice. Some define fairness as equal treatment: applying the same criteria to all candidates in the same way. Others emphasise equal opportunity, recognising that candidates start from unequal circumstances and may need barriers to be considered when evaluating merit. In this debate, AI is sometimes presented as a tool that can reduce arbitrary decisions and expand access by standardising screening. Yet the same technology can entrench inequality if organisations treat output scores as unquestionable truths, or if models are trained on patterns that reflect past exclusion. In effect, AI does not settle the ethics of hiring; it makes the underlying choices more scalable, and therefore more consequential.

A
The “gig economy” refers to labour markets in which work is broken into discrete tasks and matched to workers through digital platforms, often on an on-demand basis. In a traditional employment contract, the firm typically offers a continuing relationship in exchange for a degree of worker availability, and it absorbs a share of risk through predictable wages and benefits. Platform-based work reverses many of these assumptions. Individuals may piece together income from multiple clients and multiple apps, with no single organisation guaranteeing hours or progression. Advocates present this as a route to flexibility and entrepreneurship, particularly for those who need to work irregular schedules. Critics, however, describe a growing precariat: workers who are nominally free but whose incomes and conditions are volatile, and whose exposure to market shocks is more direct than in standard employment.

B
Technological infrastructure has been crucial to this transformation. Smartphones, digital payments, and GPS have enabled platforms to coordinate large workforces at very low transaction cost, matching supply and demand in real time. Yet the same tools that increase efficiency can also create a form of algorithmic management. Instead of a supervisor giving instructions, the app allocates tasks, sets pay rates, and evaluates performance. Many platforms add gamification features—streaks, badges, acceptance-rate targets, and time-limited bonuses—that nudge workers toward behaviours favoured by the system. GPS tracking supports rapid dispatch, but it also allows continuous monitoring, and workers may feel that their autonomy is narrower than the “independent” label suggests. Ratings systems, presented as consumer feedback, can become a lever of discipline: a low score may reduce future offers, making workers reluctant to refuse inconvenient jobs or contest unfair demands.

C
Legal classification is among the most contested issues. Many platforms treat workers as independent contractors rather than employees, a distinction with major implications for minimum wage protections, paid leave, collective bargaining rights, and employer contributions to social insurance. Platform companies argue that contractor status preserves the freedom to choose hours and to work across multiple services. Opponents counter that misclassification can occur when a platform controls core aspects of the job—pricing, access to customers, performance standards, and penalties—while avoiding obligations that accompany employment. The debate is not merely semantic: it determines whether risk and responsibility sit primarily with the firm or with the individual. In practice, some workers may value contractor status, but others argue that the imbalance of power makes “independence” largely nominal.

D
The worker experience is therefore highly uneven. For some, gig work offers a low barrier to entry, a way to supplement income, and a sense of control over when to log in and out. For others, the reality is an income stream shaped by unpredictable demand, opaque pay calculations, and unpaid waiting time between tasks. Financial risk is often shifted onto the worker. In delivery or ride-hailing, costs such as fuel, vehicle maintenance, insurance, and depreciation can be substantial, yet they are typically not reimbursed. When illness, injury, or family responsibilities interrupt work, there may be no sick pay to buffer the loss. Information asymmetry can compound these problems: the platform may see the full distribution of demand and pricing, while workers see only what the app reveals, making it difficult to assess whether a given task is worthwhile.

E
Consumers have benefited from rapid, app-based convenience and, in some sectors, lower prices. However, these gains are not always produced by underlying efficiency alone. Many platforms have relied on venture capital subsidies to expand quickly, using aggressive incentives and temporarily reduced fees to grow market share. Such strategies can resemble predatory pricing: prices are kept artificially low to outcompete rivals, even if the model is not yet profitable. Once a market consolidates, the balance of power can shift. Reduced competition may allow platforms to raise fees charged to consumers or to restaurants and suppliers, while simultaneously lowering pay rates or increasing performance demands on workers. In this sense, the apparent affordability of platform services can depend on an ongoing “race” funded by external capital, followed by a later phase in which the platform seeks to extract value from both sides of the market.

F
Regulatory responses reflect different priorities and legal traditions. Some jurisdictions have introduced intermediate categories between employee and contractor, granting partial protections such as minimum earnings floors, limited benefits, or collective representation without full employment status. Other governments pursue stricter enforcement of existing labour law, arguing that many platform workers already meet tests for employment and should be reclassified accordingly. A third approach focuses on transparency rules: requiring clearer pay calculations, disclosures about algorithmic decisions, or access to appeals processes when accounts are deactivated. These divergent policies can also encourage regulatory arbitrage, where platforms adjust operations to exploit gaps between legal regimes. Policymakers therefore face trade-offs between preserving innovation and flexibility on the one hand and preventing a systematic erosion of labour protections on the other.

G
The gig economy also interacts with inequality in complex ways. Platforms can provide entry opportunities for migrants, students, or those excluded from formal employment, and the ability to start quickly may be valuable when alternatives are limited. Yet gig work can also lock workers into low-wage segments with limited progression, especially when training and career ladders are absent. Reputation systems intensify this dynamic. Because ratings affect job allocation, early low ratings can have long-lasting effects, effectively blacklisting workers from better tasks or from the platform altogether. Small shocks—an unfair complaint, a technical glitch, a single bad week—can therefore shape long-term earning potential. Meanwhile, where healthcare, pensions, or unemployment support are tied to traditional employment, a shift toward task-based work can make basic protections harder to access, increasing insecurity for those who depend on platform income.

H
Looking ahead, the gig economy’s impacts will depend on how technology, markets, and policy evolve. Automation could reduce demand for certain forms of task-based labour, while new platforms may expand gig models into professional services such as design, tutoring, or consulting. The central tension is whether societies can preserve flexibility while ensuring fair pay, meaningful voice, and minimum protections. Some view gig work as a stepping stone, allowing rapid entry and experimentation; others see it as the institutionalisation of a permanent secondary labour market. The outcome is unlikely to be determined by technology alone. It will be shaped by how rules are written, how platforms design incentives, and how workers and consumers respond to the costs and conveniences that on-demand labour makes possible.

A
Early accounts of the social web often carried a utopian undertone: by lowering the cost of publishing and connecting, platforms would diversify public debate, weaken gatekeeping, and expose users to a wider range of perspectives. In practice, the same infrastructures that amplify marginal voices can also intensify fragmentation. When information flows are mediated by feeds, recommendations, and social graphs, the architecture of exposure changes. Users no longer share a common front page; instead, they inhabit personalised streams shaped by interaction histories and network ties. This shift matters because polarisation is not only disagreement over policy. Researchers increasingly distinguish “affective polarisation”—the growth of distrust and dislike between groups—from ordinary ideological difference. Digital systems can amplify affective polarisation by shaping what people see, who they encounter, and how quickly emotionally charged interpretations of events are circulated and rewarded.

B
A common starting point is to clarify terms that are frequently conflated. A “filter bubble” refers primarily to algorithmic selection: content is ranked and recommended according to inferred preferences, so that exposure becomes passively customised even when users do not consciously curate it. Echo chambers, by contrast, involve active social sorting. People follow accounts that confirm their views, join ideologically aligned communities, and mute or block dissenting voices. In other words, algorithmic filtering can narrow what is visible, while user-driven choices can narrow who is socially present. The most consequential cases arise when both mechanisms interact. The platform’s recommender learns from prior engagement, while the user’s selective network supplies a steady stream of reinforcing cues. The result can be epistemic closure, in which alternative interpretations are not merely absent but pre-emptively discredited as illegitimate.

C
To understand why polarisation can accelerate, it is necessary to examine the incentives built into ranking systems. Most large platforms optimise feeds for engagement because time-on-platform and repeated visits generate advertising revenue. Engagement is not a neutral metric: content that triggers strong emotions tends to attract immediate reactions, comments, and resharing. Anger, fear, and moral outrage are particularly “sticky” because they invite rapid judgement and collective signalling. Posts that frame public issues as battles between villains and victims often outperform nuanced explanations that require attention and uncertainty. Importantly, this dynamic does not require deliberate intent by platform designers to produce division. It can emerge as a by-product of optimisation goals: the algorithm promotes what reliably holds attention, and in many contexts outrage is a powerful attention magnet. Over time, repeated exposure to high-arousal content can shift the perceived “normal” tone of debate, making moderation appear weak and complexity appear evasive.

D
At the same time, evidence about how sealed people’s information environments actually are remains mixed. The popular image of users trapped in perfectly homogeneous echo chambers is difficult to sustain empirically across entire populations. Many people still encounter cross-cutting views through friends, family, colleagues, and mainstream news outlets that circulate widely. Paradoxically, some studies report that the most active political users are also among the most exposed to opposing content, simply because their high activity increases the range of material they see. Yet exposure does not necessarily produce understanding. Instead, it may become hostile exposure: people “hate-read” rival viewpoints in order to mock them, harvest examples for their own side, or confirm negative stereotypes about the out-group. In this scenario, polarisation is fuelled not by isolation alone but by antagonistic contact that hardens identities and frames disagreement as evidence of moral failure.

E
Group-based identity helps explain why the same piece of information can be processed so differently across communities. Social Identity Theory suggests that individuals derive self-esteem and meaning from group membership, and they therefore protect in-group status while derogating out-groups. Online spaces often intensify this tendency because they reward visible loyalty signals: shared slogans, insider humour, and repeated narratives that mark belonging. When political positions become fused with identity, disagreement can feel like an existential threat rather than a contestable claim. Discussion shifts from persuading outsiders to performing for insiders, with certainty and rhetorical aggression earning social rewards. Under these conditions, confirmation bias becomes socially reinforced. People may not simply prefer confirming information; they may be applauded for it. The same architecture that builds community can therefore reduce openness to revision, because changing one’s mind risks social costs within the group.

F
Misinformation and conspiracy narratives can thrive in such environments, but not only because false claims are irresistibly persuasive. A key mechanism is repetition. Repeated assertions can generate familiarity, and familiarity can be misinterpreted as truthfulness—an effect sometimes discussed in relation to the “illusory truth” phenomenon. Social reinforcement compounds this: when a claim is endorsed by trusted peers, it feels credible even when external evidence is weak. Coordinated actors can exploit these dynamics by targeting niche communities with tailored messages, but misinformation also spreads organically when issues are complex and institutional trust is fragile. In communities where official sources are assumed to be biased, corrective information can be reinterpreted as further proof of conspiracy. The result is not merely the presence of falsehood, but a breakdown in shared standards for what counts as reliable knowledge.

G
Platforms have experimented with interventions, but results are uneven and politically contested. One approach is to introduce friction: prompts that ask users to read an article before resharing, warnings about potentially misleading links, or small pauses designed to disrupt impulsive diffusion. These measures can be framed as “speed bumps” for virality, attempting to slow the outrage loop without banning content outright. Other interventions modify recommendation systems to reduce the reach of extreme material, yet such changes can lower user satisfaction and trigger accusations of censorship or viewpoint discrimination. Fact-checking labels can help in some contexts, particularly when users are uncertain, but they can also provoke partisan backlash among audiences who interpret labels as ideological signalling. In highly polarised settings, the intervention itself becomes politicised, and trust in the referee is as contested as the content being refereed.

H
Ultimately, echo chambers and polarisation are best understood as sociotechnical phenomena rather than as mere software glitches. Algorithms shape exposure, but they operate within existing divisions, media ecosystems, and incentives for political entrepreneurs who benefit from mobilising outrage. Likewise, user choice matters, but choices are made within interfaces that reward rapid reaction and public signalling. Long-term solutions therefore likely require multiple layers: platform design that discourages high-arousal amplification, governance that increases transparency around recommendation and moderation practices, and media literacy that improves users’ judgement under conditions of information overload. Beyond platforms, institutional reforms that rebuild trust—through accountability, competence, and fairness—may be necessary to reduce the demand for conspiratorial explanations. The challenge is to limit harmful polarisation while preserving legitimate disagreement, recognising that democratic debate requires conflict but not permanent hostility.