The Cycle of Algorithmic Power-Testing objective → data → category → indicator/proxy → rule or model → threshold/ranking → decision → effect → contestation/correction classification ≠ nature’s final name • prediction ≠ cause • ranking ≠ value • automation ≠ the end of responsibility Problem In digital society, algorithms often function as invisible decision systems. Which result appears first in a search, which risk category receives a loan application, which application is prioritised in recruitment, which content reaches the foreground on a social platform, and who is labelled ‘high’, ‘medium’ or ‘low’ risk in an educational or administrative list—many such decisions are produced by combinations of rules, data, models, thresholds and rankings. The problem is not that calculation occurs; the problem is how visible human purposes, categories, values and institutional power remain within that calculation. The word ‘algorithm’ often acquires an aura of almost magical impartiality. If a decision comes from a computer, it is assumed that personal bias has disappeared. But a computer works only on the question it is given, reads the data made available to it, optimises the objective that has been designed, and applies the threshold or category that has, at some level, been determined by human or institutional choice. Impartial calculation and an impartial system are not the same thing. Classification becomes power because administration, markets, education, security, health and platform systems read persons or events through a limited set of categories. Categories are also necessary: without them hospitals could not diagnose, archives could not retrieve material, tax systems could not organise liabilities, and libraries could not create subject catalogues. But when a category is tied to rights, opportunity, visibility, punishment or the distribution of resources, the act of ‘naming’ itself becomes consequential. A further difficulty of algorithmic classification is that categories are often built not from the thing directly but from proxy indicators. ‘Trustworthiness’ cannot be measured directly, so income, payment history, address, network, behaviour or other indicators may be used instead. ‘Merit’ may be represented by examination scores, institution attended, years of experience, linguistic style or previous employment. Indicators can be useful, but the distance between an indicator and the quality it represents generates political and ethical questions. Another difficulty is that data from the past can become rules for the future. If the historical system was unequal, its records do not show only ‘what happened’; they also contain traces of earlier inequalities of opportunity, administrative selection, measurement bias and people who were absent from the record. If a model learns from such data, old patterns can re-enter future decisions even when no programmer has explicitly written a discriminatory rule. The power of ranking differs from classification but is closely related to it. A person need not be called ‘unqualified’; placing them very low on a list may still cause the loss of an opportunity. In search, recommendation, recruitment, advertising, assistance, news and educational material, the effect of position on visibility is extremely important. We should therefore ask not only ‘who was removed?’ but also ‘who was placed how high or low, and for what objective?’ The final difficulty is the fragmentation of responsibility. A model developer may say, ‘we only built the technology’; an institution may say, ‘the model made the decision automatically’; an official may say, ‘the system gave this signal’; a user may say, ‘this was the result on the screen’. When a decision has real effects on human beings, this chain of responsibility cannot remain obscure. Parallel Philosophy asks: who will understand the reasons for the decision, who will correct an error, where will an appeal be heard, and who will answer for the harm? Core Proposition The central proposition of this chapter is that an algorithm should be examined not merely as a mathematical sequence but as a socio-technical decision system. Code, data, categories, institutional objectives, economic incentives, legal limits, human review, user behaviour and affected communities all contribute to the result. Algorithmic power resides not only in code but in the entire system and its feedback loops. First principle: classification may be necessary, but no administrative category should be treated as the complete nature of an object or person. Categories such as ‘risk’, ‘talent’, ‘relevance’, ‘suspicion’, ‘trustworthiness’ and ‘engagement’ are purpose- dependent measurements. They have useful limits; turning them into final truths of existence is reification. Second principle: defining the objective is itself a value judgement. A system increases ‘accuracy’—but which kind of error has been treated as more costly? It increases ‘engagement’—but is engagement equivalent to user welfare? It reduces ‘risk’—but risk to whom, and over what period? An objective function may take a technical form, while the underlying question remains ethical and political. Third principle: data are not a transparent mirror of reality. Data are products of selection, measurement, absence, category formation, recording and historical institutions. Saying ‘it is in the data’ is not sufficient justification; we must ask who was recorded and who was not, under what circumstances, by what definition, with what errors, and whether the data are relevant to the present decision. Fourth principle: prediction is not evidence of causation. A model may use certain indicators to estimate the probability of an outcome, but this does not establish that those indicators cause the outcome. If policy treats prediction as causation and intervenes on that basis, confusion may result. A high correlation is not a synonym for moral fault, natural essence or causal relation. Fifth principle: error is not merely an average percentage; the distribution of error is a question of justice. Within the same overall accuracy, errors may fall unequally across groups, languages, regions, ages, genders, occupations, disabilities or social circumstances. System evaluation should therefore ask not only ‘how much error?’ but also ‘error for whom, of what kind, and with what consequences?’ Sixth principle: the affected person’s rights to know the reasons, contest the decision and obtain correction lie at the centre of algorithmic justice. Not every model can be completely simple; nevertheless, the relevant reasons for a decision, the data used, the rules applied, the uncertainty, the review route and the appeals process should be made as clear as reasonably possible. Opacity may result from technical complexity, but it is not an excuse for eliminating responsibility. Seventh principle: in high-impact decisions, automation does not end human responsibility; it redistributes it. The institution that selects the system, supplies the data, sets the threshold, applies the result and benefits from it retains obligations. ‘The machine said so’ is not a moral defence. Principal Arguments First argument — classification can be intervention as well as description. If an administration labels someone ‘high risk’, surveillance may increase; increased surveillance records more incidents; more recorded incidents then appear to confirm the high-risk label. In this way a category can itself help generate the data later presented as evidence that the category was correct. Second argument — a categorical boundary makes an artificial cut in a continuous reality. Examination scores of 59 and 60, risk scores of 0.49 and 0.50, an age of 17 years and 364 days and an age of 18 years may require different treatment for administrative purposes, but persons on the two sides of the threshold do not suddenly become morally wholly different. A threshold is an administrative convenience, not a natural chasm. गजेन्द्र ठाकु र Third argument — proxy indicators create both convenience and distortion. When a quality is difficult to measure directly, a system chooses a nearby indicator. But address can proxy income, language can proxy education, a device can proxy economic status, a network can proxy social access, and purchasing behaviour can proxy many other characteristics. Such proxies can indirectly reintroduce sensitive distinctions. Fourth argument — labels are the basis of training, but labels are not themselves uncontested truth. Labels such as ‘spam’, ‘toxic’, ‘fraudulent’, ‘appropriate’, ‘inappropriate’, ‘successful’ and ‘unsuccessful’ arise from human judgement, rules or institutional history. If label quality is poor, a model may learn a poor standard with great efficiency. Fifth argument — missing data are also meaningful. Communities that are less digitally recorded, languages with few resources, informal employment, and women or marginalised people historically excluded from public records may appear ‘less present in the data’. Less data do not mean less existence. A system must learn to read the history of absence and silence. Sixth argument — when the object being measured knows that it is being measured, behaviour may change. If schools are evaluated only by examination scores, education may narrow toward test preparation; if employees are judged by one productivity index, behaviour aimed at raising the index may outweigh actual quality. When a metric becomes a target, the meaning of the metric can change. Seventh argument — ranking is not a neutral list. Higher positions in search results, news feeds, recommendations or recruitment lists receive more attention; more attention generates more clicks, applications, sales or fame; these data may then become signals for ranking still higher. Popularity and ranking can reinforce one another through feedback. Eighth argument — an individual’s qualities are not identical with a group’s statistical pattern. A group-level probability is not a certain truth about an individual future. Moving from ‘this outcome is more common on average in such a group’ to ‘this person will certainly be like this’ can be both a statistical and an ethical error. Ninth argument — accuracy is not the sole criterion of justice. A highly accurate system may violate privacy, lack any appeals process, produce more false accusations against a particular group, or pursue an objective that is itself unjust. Technical performance and ethical legitimacy require evaluation at different levels. Tenth argument — different fairness criteria can come into tension with one another. Equal positive rates, equal error rates, equal predictive reliability, individual-level equality and historical redress do not always yield the same result. Writing the word ‘fair’ does not solve the problem; the particular conception of justice chosen, and the reasons for choosing it, should be made publicly explicit. Eleventh argument — explainability depends on the audience. An engineer may need model coefficients; an official may need the applicable rule; an affected person needs an answer to ‘why did this happen to me?’; a court needs evidentially defensible reasons; a policymaker needs system-level effects. One technical explanation is not sufficient for all audiences. Twelfth argument — there is a real tension between confidentiality and transparency, but confidentiality does not justify secrecy everywhere. Trade secrets, security or personal data may limit disclosure of some details; nevertheless, sufficient information about objectives, data categories, major indicators, errors, appeals, testing methods and institutional responsibility can often be disclosed. Thirteenth argument — human review can be merely nominal. If an official treats the system’s recommendation as automatically correct, lacks time, has no access to alternative data, or fears punishment for disagreement, the presence of the phrase ‘human in the loop’ does not amount to genuine review. Meaningful human intervention requires authority, time, reasons and alternatives. Fourteenth argument — automation bias can induce people to accept machine errors; the opposite tendency, algorithm aversion, can also occur, where one visible mistake leads to complete rejection of a useful system. The proper stance is neither blind trust nor blind distrust, but context-sensitive evidence and comparative performance. Fifteenth argument — models can become outdated over time. Language changes, markets change, disease patterns change, behaviour changes and policy changes. It is therefore improper to infer permanent validity from successful initial testing. Monitoring, re-evaluation, version history and timely retirement are required. Sixteenth argument — in competitive environments people can learn to ‘game’ a model. If a recruitment system looks for particular words, applications can be filled with those words; if a platform’s recommender rewards certain behaviour, content creators can target that behaviour. Such adaptation itself changes the meaning of what the system measures. Seventeenth argument — algorithmic inequality can be especially visible for small languages and local knowledge. Limited training material, variable spelling, mixed scripts, multiple colloquial forms and weak representation of local names can increase errors in search, speech recognition, translation, content moderation and archival indexing. A technical ‘average’ is not a synonym for multilingual justice. Eighteenth argument — the right to change a category is important. Persons and communities change through life; incorrect records can be corrected; old debts can be repaid; acquittal or exoneration may occur; names, addresses, occupations, health and education can change. If a permanent digital label blocks the temporality of life, it is both epistemically and morally defective. Nineteenth argument — algorithmic decisions can affect third parties as well. Giving one person’s content greater prominence can make another’s less visible; a person’s credit-risk classification can affect the reputation of a family or area; policing or insurance models can alter the experience of an entire neighbourhood. The broader affected community, not only the direct user, therefore matters. Twentieth argument — the question of power cannot stop at ‘who wrote the algorithm?’ Whose data are used, who owns the platform, who defines the objective, who controls the resources, who can change the model, who hears an appeal, who conducts oversight, and who has the power to withdraw the system—together these questions constitute the real map of power in algorithmic governance. Pūrvapakṣa First Pūrvapakṣa: algorithms are more impartial than human beings, so decisions should be delegated to them wherever possible. The answer is that in some tasks calculation can reduce human arbitrariness, apply the same rule consistently and increase stability. But impartiality is a property of the system, not of the computer alone. With a biased objective, poor data, an unjust proxy or a process without appeal, calculation can make an old injustice more regular and systematic. Second Pūrvapakṣa: the data are vast, so individual bias will automatically wash out. Scale does not guarantee representation. Billions of records may come from the same kind of user; historical records may themselves be unequal; incorrect labels may be repeated on a massive scale. Quantity is not a substitute for quality, relevance or justice. Third Pūrvapakṣa: if sensitive attributes are removed, discrimination disappears. Many other indicators can act as proxies for sensitive attributes. Moreover, group-related information may sometimes be needed to measure unequal impact. Blindness and justice are not the same; the question is which characteristic is used, for what purpose, and under what safeguards. Fourth Pūrvapakṣa: for transparency, simply publish the entire source code. Code can be useful, but by itself it is insufficient. Without training data, labels, objectives, versions, thresholds, institutional processes, human overrides and real-world effects, even a reader of the code receives an incomplete picture. Transparency must be multi-layered. Fifth Pūrvapakṣa: if a model is very complex, giving reasons is impossible, so the result must simply be accepted. Some systems are indeed complex; but in high-impact decisions an institution can at least provide the relevant inputs, principal reasons, threshold, uncertainty, route to alternative review and appeals process. Complexity does not cancel procedural justice. गजेन्द्र ठाकु रक समानान्तर दर्शन — खण्ड २ Sixth Pūrvapakṣa: the market will automatically eliminate bad algorithms. Users may lack alternatives or information, or be locked in; harms may fall on third parties; discrimination may not even become visible to those affected. Market signals are useful, but they cannot wholly replace oversight, rights and public rules. Seventh Pūrvapakṣa: objections to algorithmic classification are fundamentally anti-technology. No. The question is not whether technology should be eliminated, but where and how it should be used. In some situations algorithmic assistance can be better, more consistent or more inclusive than human judgement. The Parallel criterion is comparative: compared with which alternative, with what effects, and with what path for correction? Eighth Pūrvapakṣa: if justice can be completely defined mathematically, political disagreement will disappear. Mathematics can clarify criteria, expose trade-offs and measure errors; but questions such as ‘which trade-off is justified?’, ‘whose harm is unacceptable?’ and ‘how much weight should historical injustice receive?’ require public ethical judgement. Equations can clarify disagreement; they cannot abolish it. Uttarapakṣa First response — treat the minimum unit of algorithmic governance not as the ‘model’ but as the ‘decision system’. Objective, data source, labels, model, threshold, interface, human role, affected rights, appeal, monitoring and correction should be examined together. This prevents the glamour of technical performance from concealing institutional defects. Second response — every high-impact system should have a clear limitation of purpose. Data collected for one purpose do not become automatically legitimate for unlimited other uses. When the purpose changes, consent, legality, relevance and risk should be reassessed. ‘The data are available’ is not a sufficient reason for use. Third response — categories and thresholds should have publicly defensible reasons. Why the category exists, where the threshold is placed, what the effects of both types of error are, and who benefits or loses if the boundary is moved should be documented. A threshold is not received from nature; it is shaped by policy judgement, and therefore requires justification. Fourth response — conduct both pre-deployment impact testing and post-deployment monitoring. Before deployment, examine data quality, subgroup performance, misuse, security, privacy, human processes and rights; after deployment, examine drift, complaints, unequal errors, unexpected effects and feedback loops. Fifth response — an appeal should not be merely a form but a real pathway of contestation. An affected person should be able to understand the decision well enough, see and correct inaccurate data, submit additional evidence, request human review, receive a time-bound response and, where necessary, obtain reconsideration by an independent level. Sixth response — fairness evaluation should examine both averages and distributions. Overall accuracy, false positives, false negatives, calibration, coverage, language or regional performance, success of appeals, severity of harm and reversibility of decisions should be examined in an appropriate combination. One number cannot stand in for an entire moral judgement. Seventh response — make documentation an institutional memory. Preserve where the data came from, which version was used, what limitations applied, what tests were performed, what changes were made, who approved them, when a model update occurred and when a rollback occurred. When digital power has no version history, tracing the source of an error becomes difficult. Eighth response — saying ‘no’ to some uses is also part of design. If the harm is extremely serious, error is unacceptable, evidence is insufficient, appeal is impossible, privacy intrusion is disproportionate or the purpose itself is unjust, building a more accurate model is not the only answer. The proper decision may be not to deploy the system, to stop it, or to restrict its use. Indian Dialogue Indian philosophical traditions contain extensive debates on class, jāti, universals, particulars, characteristics, word-meaning and sources of knowledge. It would be anachronistic to describe these debates as predecessors of modern machine learning or computational classification; yet conceptual dialogue is useful because they ask: what is the basis for placing an object in a class, what is the status of the universal, what is the relation between name and object, and how is misclassification detected? Vaiśeṣika reflection on padārthas—substance, quality, action, universal, particular, inherence and so forth—shows a philosophical attempt to organise the world. It is not directly equivalent to a modern administrative category, but it reminds us that classification carries ontological commitments: which distinctions we treat as real, which qualities we regard as decisive, and which relations we treat as stable. The Nyāya tradition’s reflection on pramāṇa can provide an important discipline for algorithmic decision-making: distinguish the basis of an inference, defeating evidence and sources of error. A model output may be a useful signal resembling an inference, but without testing whether its ‘reason’ is appropriate, whether counterexamples exist, and whether the underlying data are defective, the output does not become proof. Navya-Nyāya’s precise language of relations offers a parallel caution: instead of simply saying ‘such-and-such a property exists’, specify the limits of the relation—within what domain, under what conditions, through what delimitor, and in what context. In modern classification too, ‘risk’ or ‘relevance’ is not an absolute property; it is a measurement relative to a particular objective, time, environment and decision. Buddhist apoha theory emphasises the role of exclusion in semantic classification: a class is formed through a boundary of ‘not- other’. There is no direct identity with modern classification, but it raises a useful question: whom does the boundary of the category exclude? The ethical effect of a category falls not only on those placed inside it but also on those kept outside. Jain anekāntavāda and naya theory remind us of the limits of perspective. The lesson here is not that all classifications are equally true, but that one administrative perspective is not the whole truth about a person or object. A credit-risk classification reads one aspect of economic behaviour; it does not read the person’s agency, morality, talent, social contribution or entire future. Mīmāṃsā developed rigorous disciplines concerning language, sentence, context and the meaning of prescriptive statements. Algorithmic rules also face the risk of context collapse: the wording of a rule may remain the same while institutional circumstances change. For a rule written ‘if X, then Y’ to be properly applied, the definition of X, exceptions, purpose and jurisdiction must be clear. The essence of this Indian dialogue is that classification should be an instrument of knowledge, not a prison of existence. Examine sources of evidence, understand the limits of categories, accept counterexamples, clarify the context of relations, and avoid generalisations that imprison the multidimensional existence of persons or communities within a single administrative label. Mithila’s Parallel Perspective Mithila’s Parallel perspective will assess algorithmic power not through local pride or anti-technology slogans, but through a combined test of evidence, history, language and affected rights. As regional administration increasingly digitises land, flood management, agriculture, scholarships, social security, health, education, disaster relief and archives, the strengths and defects of classification can directly affect everyday life. In land records, errors in spelling of names, father’s or husband’s name, village, police station, khata, khasra, boundary description or OCR of older documents can affect decisions tied to a person’s rights. A documentary error should not become truth merely because ‘the system says so’. Original records, correction history, local evidence and an appeals route are necessary. गजेन्द्र ठाकु र Flood-risk classification can be useful, but the dynamics of rivers, embankments, rainfall, local drainage, elevation, roads, previous erosion and village micro-geography are more complex than a fixed label. If relief depends only on an old digital risk map, newly affected settlements may be missed. Local direct information and updated data should accompany the model. Algorithmic representation of Maithili and other regional languages requires special caution. Devanagari alongside Tirhuta, Roman transliteration, local spelling, Sanskritised or colloquial forms, and mixed Hindi–Maithili text together constitute the actual digital corpus. If language-identification or moderation systems treat only a single ‘standard’ form as correct, they may classify living linguistic diversity as error. In the digitisation of Panji records, genealogies, manuscripts and historical archives, the temporality of categories should be recorded. Mechanically projecting a modern fixed social label onto documents from the past can distort history. The term found in an archive should retain its date, source, textual variants and historical context as it was used and understood at that time. In local literary collections, recommendation algorithms may repeatedly raise popular texts and push little-read works lower. The goal of a cultural archive is not merely more clicks; the discoverability of rare texts, women’s writing, local journals, manuscripts, oral histories, old issues and lesser-known authors is also part of cultural justice. Mithila’s Parallel maxim here is: ‘Do not treat any classification as a final identity; record its source, date, purpose, boundary, exceptions and path of correction.’ Bringing local knowledge into algorithmic systems does not mean digitising local prejudice; it means giving local experience a place within a universally testable dialogue of evidence. Ten Criteria for Algorithmic Classification objective clear? → data relevant? → labels reliable? → proxy justified? → threshold reasoned? → error distribution tested? → explanation available? → human review genuine? → appeal possible? → version/correction recorded? Before saying ‘the model is correct’, ask: correct for which question, for which people, at what time, compared with which alternative, and at what cost of harm? Contemporary Applications In credit and financial decisions, algorithmic scores can be useful, but irregular income, informal employment, a rural address, lack of digital transactions or historical exclusion from banking do not automatically imply moral untrustworthiness. High- impact credit decisions require reasons, correction of inaccurate data, alternative evidence and a route to human reconsideration. In recruitment, when using keywords, résumé ranking, skill tests or video analysis, decisive questions include how the training label ‘successful employee’ was constructed, whether the former employee group is representative, what the effect is on disability or linguistic difference, and whether candidates can contest the result. Efficiency should not oppose justice of opportunity; both should be designed together. In education, early-warning systems can help direct support to students, but a ‘dropout risk’ label can become stigmatising and produce the opposite effect. A score should assist a teacher’s judgement, not become a sentence of fate. Students and families should have clear routes for data correction, contextual explanation and support. In healthcare, triage or diagnostic-support systems should not be assessed only by average accuracy. Performance can vary across age, sex, disease prevalence, equipment, hospitals, languages and populations. In clinical decisions, model output should be one part of the evidential chain, read together with patient history, tests, clinical judgement and uncertainty. Fraud detection can be useful in public-benefit and social-security systems, but false positives can block assistance to poor families. Because harms are unequal, thresholds, prior notice, reasons, rapid human review, restoration of withheld payments and grievance routes should be designed accordingly. Efficient administration does not mean administration without appeal. Risk systems used in policing or security require particular caution about feedback loops. Increased surveillance can generate more recorded incidents; those incidents can then become the basis for higher risk classifications. Where effects on places, communities or individuals are severe, independent testing, legal limits, data quality and public accountability should be especially stringent. In social-media moderation, classification depends on linguistic and cultural context. Satire, quotation, news reporting, criticism of hateful speech, reclaimed language and direct harassment can use the same words while meaning very different things. Automated signals can help, but appeals, contextual review, policy transparency and local linguistic expertise are necessary. In search and recommendation, ‘relevance’ is itself multi-valued—immediate popularity, source quality, novelty, diversity, user interest, public interest, local language and safety. A single hidden objective can reshape the visibility of an entire knowledge environment. Giving users some control, general reasons for ranking and alternative discovery paths can increase epistemic plurality. In generative artificial intelligence, classification operates at many hidden levels—selection of training material, safety labels, retrieval ranking, token prediction, moderation and preference tuning. A fluent answer can conceal the history of these selections. Users should remain aware of source checking, uncertainty, version awareness and the need for independent evidence in important decisions. In digital archives and cultural heritage, classification creates the map for future research. What has metadata can be found; what lacks it may become invisible. Multilingual titles, alternative names, date ranges, provenance, original collection numbers, textual variants, scripts and multiple subject entries are therefore part of long-term epistemic justice. In administrative dashboards, red, yellow and green colours can accelerate decision-making, but colour is not evidence. Definitions of every indicator, data freshness, confidence, missingness and thresholds should be available. Visual simplicity should help policymakers; it should not hide complex reality. The purpose of an algorithmic audit is not merely to obtain a certificate once. High-impact systems require continuing re- evaluation because data change, new uses arise, drift occurs, complaints accumulate, adversarial behaviour develops and society adapts. Without independence, necessary access, repetition and corrective consequences, an audit can become a mere formality. Chapter Conclusion Algorithms are powers of calculation, but in social life their influence grows through the power of classification. Systems that categorise people, events, languages, risks, merit, truthfulness or relevance do not merely read the world; they also affect the distribution of opportunity, visibility, resources and surveillance. Philosophical examination of algorithmic governance must therefore move beyond code to institutional power. Parallel Philosophy treats algorithms as neither demons nor gods. Calculation can provide consistency, scale, pattern recognition and assistance; human beings can be arbitrary, fatigued and inconsistent. But technical assistance becomes just only when objectives are reasoned, data are relevant, categories are limited, error distributions are visible, appeals are real and responsibility is clear. The ethical limit of classification is that no score is the whole person; no prediction is destiny; no ranking is the final order of value; no proxy is a complete substitute for the real quality; and no automation frees an institution from responsibility. The more consequential the decision, the stronger the rights to reasons and contestation must be. Parallel maxim of Chapter 86: ‘The power of an algorithm lies not only in its calculation, but in whom it places in which category, whom it raises, whom it lowers, and who can contest an error when one occurs.’