The ethics of algorithms: Mapping the debate

The ethics of algorithms: Mapping the debate Brent Mittelstadt Patrick Allo Mariarosaria Taddeo Sandra Wachter Luciano Floridi

Tag layer of the DHSS Meta-Tagging Project. The article text is reproduced from the staged copy named in sourceDesc and is not altered; the tags are held separately in standOff so that overlapping and nested spans are preserved.

The ethics of algorithms: Mapping the debate 2016 Big Data & Society 10.1177/2053951716679679 mittelstadt2016-the-ethics-of-algorithms-mapping-t mittelstadt2016-the-ethics-of-algorithms-mapping-t.txt, the project's staged copy 87582e45f2d3957bab29d14130a68dbaa9892333 cc-by-nc ai-ethics

the paper's own words

the tag layer; a sentence written by the tagger about the paper, always anchored to a quotation but not itself a quotation

0 spans located by searching the stored text for the quote's exact words, allowing any whitespace between them the ethical concerns raised by decision-making algorithms, as they appear in the academic literature that discusses algorithms explicitly Quotation withheld because this record is not cleared for redistribution. what kinds of ethical issues algorithms raise, and which of them belong to the algorithm rather than to the technology built on it Quotation withheld because this record is not cleared for redistribution. a prescriptive six-part map of ethical concerns - three epistemic, two normative, one overarching - together with a review of the literature against it and an assessment of where the field is thin Quotation withheld because this record is not cleared for redistribution. method Quotation withheld because this record is not cleared for redistribution. the gap between how algorithms are designed and operated and how their ethical implications are understood Quotation withheld because this record is not cleared for redistribution. systematic-review systematic-review Quotation withheld because this record is not cleared for redistribution. mechanisms OUTPUT CANNOT EXCEED INPUT: a conclusion inherits the reliability and the non-neutrality of the data it was computed from, so an algorithm cannot repair evidence that was misguided before it arrived status: the informal ‘garbage in, garbage out’ principle clearly illustrates Quotation withheld because this record is not cleared for redistribution. mechanisms A LEARNING ALGORITHM REWRITES ITS OWN DECISION STRUCTURE WHILE RUNNING, so the rationale for any particular classification is not fixed at design time and cannot be read off the program Quotation withheld because this record is not cleared for redistribution. mechanisms ONCE TRAINED, CLASSIFICATION PROCEEDS WITHOUT AN OPERATOR - the structure produced by training is what new inputs meet, which is why oversight has no natural point of entry Quotation withheld because this record is not cleared for redistribution. mechanisms POPULATION KNOWLEDGE APPLIED TO AN INDIVIDUAL: correlations hold of groups while the action lands on a person, and even population-level action spills over - an insurance premium set for a sub-population is paid by each member Quotation withheld because this record is not cleared for redistribution. mechanisms VALUES ENTER THROUGH PARAMETERS: developers specify and users configure operational parameters with desired outcomes in mind, so the algorithm privileges some interests before it processes anything Quotation withheld because this record is not cleared for redistribution. mechanisms SECRECY IS COMMERCIALLY MOTIVATED, NOT INCIDENTAL: disclosure would enable gaming and threaten viability, so inaccessibility is a position taken in a conflict of interests rather than a technical limit Quotation withheld because this record is not cleared for redistribution. mechanisms CORRELATIONS ARE ACTED ON WITHOUT CAUSAL KNOWLEDGE because a sufficient volume of data is treated as credible enough to drive action - the paper's name for that lowered bar is `actionable insights` Quotation withheld because this record is not cleared for redistribution. the review covers only literature that discusses algorithms explicitly, so neighbouring fields - AI ethics, surveillance studies, computer ethics, machine ethics - are touched on only briefly Quotation withheld because this record is not cleared for redistribution. the authors accept there may be gaps in coverage for topics discussed only in relation to specific algorithm types rather than to algorithms as such Quotation withheld because this record is not cleared for redistribution. summarising every field in which algorithms are used is stated to be beyond reach, and the narrowing is justified on that ground Quotation withheld because this record is not cleared for redistribution. discussion of 'algorithm' runs into a level-of-abstraction problem: without fixing the level, parties talk past each other and the set of relevant observables is never bounded Quotation withheld because this record is not cleared for redistribution. the authors state that a mature ethics of algorithms does not yet exist, and attribute that partly to the breadth of the concept itself Quotation withheld because this record is not cleared for redistribution. configuration is excluded from changing the object: tuning an algorithm to a task is said not to alter its mathematical representation or its implementation, which bounds what the map is about Quotation withheld because this record is not cleared for redistribution. algorithms that automate mundane tasks are excluded by declaration rather than by a criterion, and the class of interest is given without a precise definition Quotation withheld because this record is not cleared for redistribution. the paper marks its own normative judgements as observer-dependent rather than claiming a standard: whether evidence counts as misguided, and whether an outcome counts as unfair, are both said to depend on who is evaluating eliminates: a reader taking the map's normative labels as the authors' own ethical standard - 'misguided' and 'unfair' are marked observer-dependent, so the framework cannot be read as smuggling in a particular ethics · outcome: held Quotation withheld because this record is not cleared for redistribution. the map is declared not to derive from any particular ethical theory, which is what lets readers who disagree about ethics still use it eliminates: the objection that the map only works for readers who already share the authors' ethical theory - it is declared to derive from none · outcome: held Quotation withheld because this record is not cleared for redistribution. algorithm algorithm Quotation withheld because this record is not cleared for redistribution. actionable insights actionable insights Quotation withheld because this record is not cleared for redistribution. transparency transparency Quotation withheld because this record is not cleared for redistribution. accessibility accessibility Quotation withheld because this record is not cleared for redistribution. comprehensibility comprehensibility Quotation withheld because this record is not cleared for redistribution. traceability traceability Quotation withheld because this record is not cleared for redistribution. data analytics data analytics Quotation withheld because this record is not cleared for redistribution. practice practice Quotation withheld because this record is not cleared for redistribution. profiling profiling Quotation withheld because this record is not cleared for redistribution. information asymmetry information asymmetry Quotation withheld because this record is not cleared for redistribution. level of abstraction level of abstraction Quotation withheld because this record is not cleared for redistribution. correlations correlations Quotation withheld because this record is not cleared for redistribution. causal connection causal connection Quotation withheld because this record is not cleared for redistribution. machine learning machine learning Quotation withheld because this record is not cleared for redistribution. opacity opacity Quotation withheld because this record is not cleared for redistribution. bias bias Quotation withheld because this record is not cleared for redistribution. discrimination discrimination Quotation withheld because this record is not cleared for redistribution. responsibility responsibility Quotation withheld because this record is not cleared for redistribution. implementation implementation Quotation withheld because this record is not cleared for redistribution. privacy privacy Quotation withheld because this record is not cleared for redistribution. algorithm - THE PAPER HOLDS TWO SENSES AT ONCE AND SAYS SO. The formal sense is Hill's mathematical construct; the sense it actually maps is the public-discourse one, which includes the implementation into a technology and the configuration of that technology to a task. Its own point is that a map of the ethics cannot use the formal sense alone Quotation withheld because this record is not cleared for redistribution. algorithm - AND THE CONFLATION IS NAMED AS THE OBSTACLE, not glossed over: the same word carries a mathematical construct and 'any procedure or decision process, however ill-defined', and the paper says the mapping exercise must address that before it can begin Quotation withheld because this record is not cleared for redistribution. transparency - NOT ONE PROPERTY BUT TWO CONDITIONS THAT COME APART: information can be accessible and still not comprehensible, which is exactly the case machine learning creates Quotation withheld because this record is not cleared for redistribution. fairness - EXPLICITLY OBSERVER-DEPENDENT HERE, a generic label for whatever ethical criteria an observer brings to the action rather than a standard the paper supplies Quotation withheld because this record is not cleared for redistribution. inconclusive evidence - THE PAPER'S OWN CATEGORY, defined as the uncertainty that is inherent rather than accidental: inferential and machine-learning methods yield knowledge that is probable and inevitably uncertain, so the deficiency cannot be engineered away Quotation withheld because this record is not cleared for redistribution. inscrutable evidence - defined by a REASONABLE EXPECTATION THAT IS NOT MET: the link between data and conclusion ought to be intelligible and open to scrutiny, and given how algorithms operate that is not automatically satisfied Quotation withheld because this record is not cleared for redistribution. misguided evidence - defined from a limit shared by ALL data processing rather than anything specific to algorithms: the output can never exceed the input Quotation withheld because this record is not cleared for redistribution. transformative effects - the category for harms with NO IDENTIFIABLE VICTIM AND NO OBVIOUS FAILURE: effects that look ethically neutral because nothing was broken, yet change how the world is conceptualised and organised Quotation withheld because this record is not cleared for redistribution. critiques → the treatment of distinct epistemic and normative concerns as a single cluster, so that fixing one is taken to address the rest Quotation withheld because this record is not cleared for redistribution. critiques → the assumption that better evidence reduces discrimination - the paper holds that it can be used to discriminate more efficiently Quotation withheld because this record is not cleared for redistribution. predicts → the expectation that increasingly better algorithmic tools will eliminate the ethical problems of algorithms Quotation withheld because this record is not cleared for redistribution. critiques → the naive treatment of transparency as a panacea for ethical issues arising from new technologies Quotation withheld because this record is not cleared for redistribution. qualifies → the assumption that a more discerning algorithm is the ethically preferable one Quotation withheld because this record is not cleared for redistribution. correlations —distinct-from→ causal connection correlations causal connection strength of the relation: 0.95 Quotation withheld because this record is not cleared for redistribution. accessibility —part-of→ transparency accessibility transparency strength of the relation: 0.95 Quotation withheld because this record is not cleared for redistribution. comprehensibility —part-of→ transparency comprehensibility transparency strength of the relation: 0.95 Quotation withheld because this record is not cleared for redistribution. transparency —contrasted-with→ privacy transparency privacy strength of the relation: 0.9 Quotation withheld because this record is not cleared for redistribution. machine learning —causes→ opacity machine learning opacity strength of the relation: 0.9 Audit 2026-09-28: DOUBTFUL - read against its span and kept, but the span says interpretability problems indicate opacity; causes is loose. Quotation withheld because this record is not cleared for redistribution. implementation —part-of→ algorithm implementation algorithm strength of the relation: 0.9 Audit 2026-09-28: DOUBTFUL - read against its full span and kept, but the span reports LAY USAGE of 'algorithm' as including implementation; part-of is loose. Quotation withheld because this record is not cleared for redistribution. data analytics —is-a→ practice data analytics practice strength of the relation: 0.9 Quotation withheld because this record is not cleared for redistribution. Three epistemic and two normative kinds of ethical concern arise from the use of algorithms. Quotation withheld because this record is not cleared for redistribution. solving a problem at one level of the map leaves the other kinds untouched Quotation withheld because this record is not cleared for redistribution. the concerns are interdependent, and two links were already present in the literature: bias with discrimination, and opacity with the attribution of responsibility Quotation withheld because this record is not cleared for redistribution. Even strong correlations or causal knowledge may concern only populations, while actions are aimed at individuals. Quotation withheld because this record is not cleared for redistribution. algorithm algorithm Quotation withheld because this record is not cleared for redistribution. transparency transparency Quotation withheld because this record is not cleared for redistribution. data analytics data analytics Quotation withheld because this record is not cleared for redistribution. actionable insights actionable insights Quotation withheld because this record is not cleared for redistribution. traceability traceability Quotation withheld because this record is not cleared for redistribution. accessibility accessibility Quotation withheld because this record is not cleared for redistribution. comprehensibility comprehensibility Quotation withheld because this record is not cleared for redistribution.

The full text of this article is not redistributable, so it is not included here. The evidence sentences are withheld with it; the tag layer and its grounding status remain. The article itself is at https://doi.org/10.1177/2053951716679679.