EigoPro 1級 読解 — The Mathematics of Unfairness

It is a comforting fiction that automated decision-making, being free of the prejudices that afflict human judgement, must therefore be impartial. In reality, machine-learning systems frequently entrench the very discrimination they are presumed to transcend, and they do so without any malicious architect. The mechanism is mundane rather than sinister: a model trained on historical data inherits the patterns embedded in that data, and where the past encoded inequity—in lending, hiring, or sentencing—the algorithm faithfully reproduces it. Worse, because such patterns are statistically predictive, the system is rewarded for perpetuating them. A résumé-screening tool that learns to favour candidates resembling a firm's incumbent workforce is, by any conventional metric, performing admirably. The deeper trouble is conceptual. We might hope to audit a system for fairness, but fairness resists a single mathematical definition. One can demand that a model's error rates be equal across groups; one can insist that a given predicted score carry the same meaning regardless of group; one can require that the proportion of favourable outcomes match across populations. Each criterion is intuitively reasonable. Yet a celebrated result in the field demonstrates that, save in trivial cases, these conditions are mutually incompatible—satisfying one mathematically precludes satisfying another. There is, in short, no neutral choice; every formalisation of fairness smuggles in a contestable value judgement about which harms matter most. This renders the customary appeal to algorithmic objectivity not merely naïve but obfuscatory. To delegate a decision to a model is not to evacuate it of politics; it is to bury the politics in code, where it escapes the scrutiny that an openly stated principle would invite. The remedy cannot be purely technical. It demands that designers concede the irreducibly normative character of their choices, and that the question 'fair to whom, and at whose expense?' be answered explicitly, before the data is ever assembled.

5問合格 60%以上普通

新規登録 または ログイン が必要です

この単元の問題を3問、解説つきで公開しています

1級 · 読解 · 全5

1

Which statement best captures the central argument of the passage?

  • Because algorithms outperform humans on conventional accuracy metrics, replacing human judgement with automated decision-making reliably reduces discrimination.
  • Algorithmic bias arises chiefly from data and the impossibility of a value-neutral definition of fairness, so the fix must be normative, not merely technical.
  • Machine-learning systems are unbiased by design and only become discriminatory when programmers deliberately encode their own prejudices.
  • The competing mathematical definitions of fairness can all be satisfied at once provided that the training data is sufficiently large and representative.
Phrase

every formalisation of fairness smuggles in a contestable value judgement

1級読解主旨把握論旨
ひとことで

偏りはデータ由来であり、公平の定義自体が価値中立ではないため、技術だけでは解決できないという論旨。

選択肢の解説Choice Analysis

Algorithmic bias arises chiefly from data and the impossibility of a value-neutral definition of fairness, so the fix must be normative, not merely technical.

正解

筆者は、悪意ある設計者がいなくても過去のデータの不公平を忠実に再現する点("faithfully reproduces it")と、公平には単一の数学的定義が存在しない点("fairness resists a single mathematical definition")の二つを柱に論を進める。結論部では救済策は "cannot be purely technical" と述べ、規範的判断を明示せよと迫る。したがってデータ起因性と定義の価値非中立性、そして非技術的対応の必要を統合した選択肢が主旨に合致する。

  • Machine-learning systems are unbiased by design and only become discriminatory when programmers deliberately encode their own prejudices.

    = 本来は無偏で、悪意あるプログラマが偏らせるだけだとする説 ━ 第1段が "without any malicious architect" と明言し、悪意なき偏りこそ問題だとしている点に反する。。

  • The competing mathematical definitions of fairness can all be satisfied at once provided that the training data is sufficiently large and representative.

    = データを大規模化すれば全公平基準を同時に満たせるとする説 ━ 第2段の "mutually incompatible" という不可能性の結果と矛盾し、データ量で解決できるとはどこにも書かれていない。。

  • Because algorithms outperform humans on conventional accuracy metrics, replacing human judgement with automated decision-making reliably reduces discrimination.

    = 精度指標で勝るゆえ自動化が差別を減らすとする説 ━ "performing admirably" でも差別を温存しうると示されており、精度の高さは無偏の証拠にならないとする本文と逆である。。

Collocation
パターン意味
entrench discrimination差別を固定化するOpaque scoring systems can entrench discrimination for decades.
free of prejudice偏見がないNo dataset is entirely free of prejudice.
mutually incompatible相互に両立しないThe two safety goals turned out to be mutually incompatible.
value judgement価値判断Ranking applicants always involves a value judgement.
delegate a decision to判断を〜に委ねるCourts are reluctant to delegate a decision to a black-box model.
例文
  • The audit revealed that the model had quietly entrenched a historical bias.

    監査により、そのモデルが過去の偏りをひそかに固定化していたことが判明した。

  • Claiming objectivity, the firm buried a political choice inside its code.

    客観性を掲げつつ、その企業は政治的選択をコードの中に埋め込んでいた。

  • Fairness, the author insists, is irreducibly normative.

    公平とは、と筆者は主張する、どうしても規範的なものなのだ。

  • High accuracy alone cannot certify that a system is impartial.

    高い精度だけでは、システムが公平だと保証することはできない。

💡Tip · 覚えるコツ

主旨設問は「両方の柱+結論」を含む選択肢を選ぶ。一段落だけを正しく要約した肢は魅力的でも主旨にならない。

2

According to the passage, why is a model rewarded for reproducing historical inequities?

  • Because larger and more recent datasets automatically override the biases inherited from the past.
  • Because its designers deliberately tune it to favour the firm's existing workforce over outside applicants.
  • Because the discriminatory patterns it inherits from past data are statistically predictive, so reproducing them improves measured performance.
  • Because regulators grant higher compliance scores to models that mirror established industry practice.
Phrase

because such patterns are statistically predictive, the system is rewarded for perpetuating them

1級読解詳細把握
ひとことで

過去の偏ったパターンが統計的に予測力を持つため、それを再現すると性能指標が上がる。

選択肢の解説Choice Analysis

Because the discriminatory patterns it inherits from past data are statistically predictive, so reproducing them improves measured performance.

正解

第1段は "because such patterns are statistically predictive, the system is rewarded for perpetuating them" と明示する。さらに既存社員に似た候補を優遇するツールが "by any conventional metric, performing admirably" とされ、差別の温存が指標上は好成績になる構造を説明している。つまり報酬は統計的予測力に由来する。

  • Because its designers deliberately tune the model so as to favour the firm's existing workforce over outside applicants.

    = 設計者が意図的に既存社員優遇に調整したから ━ 本文は "without any malicious architect" と意図的調整を否定しており、報酬は意図ではなく統計に由来する。。

  • Because regulators grant higher compliance scores to models that mirror established industry practice.

    = 規制当局が業界慣行に沿うモデルを高評価するから ━ 規制当局による加点という記述は本文に一切なく、評価は "conventional metric" 即ち性能指標を指している。。

  • Because larger and more recent datasets automatically override the biases inherited from the past.

    = 新しく大きなデータが過去の偏りを自動で打ち消すから ━ 本文はむしろデータが過去の不公平を "faithfully reproduces" すると述べ、自動で打ち消すとは言っていない。。

Collocation
パターン意味
statistically predictive統計的に予測力のあるPostal code proved statistically predictive of default risk.
be rewarded for〜したことで報われるThe model is rewarded for minimising visible error.
historical data過去のデータThe system was trained on decades of historical data.
by any conventional metricどの従来指標で見てもBy any conventional metric, the classifier excels.
perpetuate a patternあるパターンを永続させるAutomation can perpetuate a pattern no one intended.
例文
  • Variables that correlate with race may be statistically predictive yet ethically toxic.

    人種と相関する変数は統計的に予測力を持ちうるが、倫理的には有害でありうる。

  • The screening tool was rewarded for echoing past hiring decisions.

    その選考ツールは過去の採用判断を反復したことで高く評価された。

  • A high score on the usual metrics concealed a deep injustice.

    通常の指標での高得点が、深刻な不公正を覆い隠していた。

  • Predictive power and fairness can pull in opposite directions.

    予測力と公平性は逆方向に引き合うことがある。

💡Tip · 覚えるコツ

「なぜ報われるか」型は因果を示す because 節を探す。意図(malicious)系の肢は本文の否定語と照合して切る。

3

What does the passage claim about the three fairness criteria it lists?

  • Each seems individually reasonable, but a well-known result shows they cannot all be satisfied together except in trivial cases.
  • All three can be met simultaneously once a model's error rates are equalised across the relevant groups.
  • They are interchangeable formulations of the same underlying principle and therefore never conflict.
  • Only the criterion requiring equal proportions of favourable outcomes is mathematically defensible.
Phrase

save in trivial cases, these conditions are mutually incompatible

1級読解詳細把握
ひとことで

各基準は妥当に見えるが、自明な場合を除き同時には満たせないという有名な結果が示される。

選択肢の解説Choice Analysis

Each seems individually reasonable, but a well-known result shows they cannot all be satisfied together except in trivial cases.

正解

第2段は三つの基準を列挙し、いずれも "intuitively reasonable" だと認める。しかし "save in trivial cases, these conditions are mutually incompatible" とあり、一つを満たすことが別の一つを数学的に排除する("satisfying one mathematically precludes satisfying another")。よって個別の妥当性と同時成立の不可能性を併記した肢が正しい。

  • All three can be met simultaneously once a model's error rates have been equalised across the relevant groups.

    = 誤差率を均等化すれば三つとも同時に満たせる ━ 本文は同時成立が "mutually incompatible" だと述べており、一基準を満たすと他が排除される。。

  • They are interchangeable formulations of the same underlying principle and therefore never conflict.

    = 三基準は同一原理の言い換えで衝突しない ━ 本文は明確に互いに非両立だとし、衝突しないどころか一方が他方を妨げると説明している。。

  • Only the criterion requiring equal proportions of favourable outcomes is mathematically defensible.

    = 好結果の比率均等の基準だけが擁護可能 ━ 本文は三基準とも "intuitively reasonable" とし、特定の一つだけを擁護可能とは述べていない。。

Collocation
パターン意味
a celebrated resultよく知られた(名高い)結果A celebrated result rules out a perfectly fair classifier.
save in trivial cases自明な場合を除いてSave in trivial cases, the equation has no solution.
error rate誤り率The error rate differed sharply between the two groups.
preclude doing〜することを妨げるTime constraints precluded a full review.
across groups集団間でOutcomes must be compared across groups.
例文
  • The impossibility theorem dashed hopes of a single fairness metric.

    その不可能性定理は、単一の公平指標への期待を打ち砕いた。

  • Equal error rates and equal positive rates rarely coexist.

    誤り率の均等と陽性率の均等が同時に成り立つことはまれだ。

  • Each definition looked reasonable until they were combined.

    各定義は組み合わせるまでは妥当に見えた。

  • Trivial datasets aside, the constraints clash.

    自明なデータを別にすれば、その制約同士は衝突する。

💡Tip · 覚えるコツ

列挙+逆接(Yet/However)の構造では、逆接の後ろに筆者の主張が来る。複数項目を一括評価する設問は逆接節を要約する。

残り2問は、解いて確かめられます

5問すべてに、この形式の解説が付いています。 間違えた問題は記録され、あとから復習できます。

無料で登録して続きを解く

インストール不要・カード登録なし・いま全機能無料