雅思阅读 172: When the Algorithm Learns Our Prejudices（当算法学会了我们的偏见）

改编自 KFF / PMC / Frontiers in Digital Health。雅思阅读 Section 3 难度，约 1050 词。 素材来源：https://www.kff.org/racial-equity-and-health-policy/the-growing-use-of-artificial-intelligence-in-health-care-and-implications-for-disparities/

Reading Passage

A. Artificial intelligence is sweeping into medicine with an attractive promise. Machine-learning models can scan thousands of scans, notes and records in seconds, spotting patterns no tired clinician could hold in mind, and scaling the judgement of top specialists to clinics that have none. Used well, such systems might narrow gaps in care by bringing expensive expertise to poor and rural hospitals. But the same systems are now being shown to carry a quieter danger: they can reproduce, at speed and at scale, the very inequalities that medicine has spent decades trying to dismantle. Because an algorithm learns from the data it is given, and because medical data in most countries reflects a long history of unequal care, a model trained without care can systematically misread, under-treat or overlook patients from certain racial, ethnic or skin-colour groups. The result is not a neutral machine but a mirror held up to society — and a mirror that scales its distortions to millions of decisions.

B. The first and most obvious problem is who appears in the training data. In a widely analysed clinical dataset, roughly 70 percent of patients described themselves as white, while black patients made up 8 to 10 percent, Hispanic patients 3 to 5 percent and Asian patients 2 to 4 percent. A model trained on such material simply has fewer examples from under-represented groups, and its rules for recognising disease are calibrated mainly on the majority. In diagnostics this shows up as reduced performance on patients with darker skin, because the images and readings the model learned from were collected disproportionately from lighter-skinned patients. A skin-lesion classifier, a retinal scanner or a heart-imager can therefore be highly accurate on the textbook patient and quietly worse on everyone else — a failure that is invisible in overall accuracy scores, which are dominated by the majority group. The problem is compounded by how such systems are tested: if the validation set merely mirrors the same skewed hospital where the training data came from, the model's weakness will not show up until it is deployed in a clinic serving different patients. A model that looks 95 percent accurate on paper can be wrong far more often for the minority group it rarely saw in training.

C. The most celebrated warning came not from a picture but from a numbers game. A widely used algorithm designed to direct extra care — for example, to patients who might benefit from intensive monitoring — was trained to predict future healthcare costs. The reasoning seemed sensible: people who cost the system a lot must be sicker. But because black patients, on average, spend less on healthcare even when equally ill — owing to poorer access, distrust and under-treatment — the model equated lower spending with lower need. As a result, it systematically referred fewer black patients to the extra-care programmes they genuinely required, simply because the money they had spent, not their underlying health, was the variable it had learned. The bias was not inserted by a racist programmer; it was distilled, faithfully, from a history of unequal spending.

D. Subsequent work has shown that fixing the numbers is harder than it looks. When a model performed poorly on young black patients, its developers tried three standard remedies: retraining on equal numbers from each group, building separate models for each racial group, and explicitly feeding race into the algorithm. None of them fully repaired the predictions. In cardiac imaging, researchers traced part of the difficulty to a physical feature rather than a label: differences in body fat outside the heart shifted the appearance of the scans, so that a model calibrated on slimmer bodies interpreted black patients' images differently even after race was accounted for; cropping the image to the heart reduced but did not remove the gap. Even models trained only on black patients have still underperformed for that group, suggesting that something about the lived experience of racism — stress, distrust, unrecorded symptoms — is simply absent from the variables the models are allowed to see. A model can be perfectly fair on paper and still miss a patient whose symptoms were never fully documented, because the clinic that saw them did not have the time or the trust to record them. Bias, on this view, lives less in the code than in the history that produced the records the code was fed.

E. Reviewers now describe the problem as four interlocking causes rather than one: biased underlying datasets, the historical and systemic prejudice encoded in those data, choices made in algorithm design, and biased deployment in the real world. No single intervention, they conclude, will cure it. Diversifying training data helps, but is not enough; careful design helps, but cannot undo deployment in clinics that already treat patients unequally. The honest conclusion is that medical AI will not, by itself, eliminate health inequity. It can either widen existing gaps, by learning them uncritically, or narrow them, if it is built, tested and audited group by group and overseen by clinicians who remember that a percentage accuracy, averaged across millions, can hide the very patients the system was meant to serve. A model that is 95 percent correct on average can still be wrong almost every time for the one patient in twenty it has never properly seen — and in medicine, that one patient is the whole point of the exercise. Fairness, on this view, is not a footnote to accuracy but a separate test that must be run on every group the system serves, not merely on the average patient who dominates the data. Until that test becomes routine, the most advanced diagnostic machine in the clinic will still be, in the end, only as fair as the history behind it.

Questions 1-4

Choose the correct heading for paragraphs B, C, D and E from the list of headings below.

List of Headings i. Who is and is not in the training data ii. The ethics of medical photography iii. A famous algorithm that confused spending with sickness iv. How hospital bills are calculated v. Why even clever fixes can fail vi. The four causes and no simple cure vii. A history of nursing in America

1. Paragraph B: ____
2. Paragraph C: ____
3. Paragraph D: ____
4. Paragraph E: ____

Questions 5-8

Choose the correct letter, A, B, C or D.

5. Why might a diagnostic AI perform worse on darker-skinned patients? A. It was trained mostly on images of lighter-skinned patients. B. It cannot read images at all. C. Dark skin blocks X-rays by law. D. It is programmed to reject them.

6. What was the flaw in the widely used "extra care" algorithm? A. It used future healthcare spending as a proxy for need, which reflects unequal access. B. It predicted heart attacks accurately. C. It charged patients different fees. D. It only worked in rural hospitals.

7. What happened when developers retrained models with equal numbers and race labels? A. Bias disappeared completely. B. The remedies did not fully repair the predictions. C. The models became slower. D. They stopped using data.

8. Why did cardiac MRI models misread some patients even after race was controlled? A. Differences in body fat outside the heart shifted how the images looked. B. The machines were too old. C. Black patients refused scans. D. The software was in another language.

Questions 9-13

Do the following statements agree with the claims of the writer?

Write:

• TRUE if the statement agrees with the information
• FALSE if the statement contradicts the information
• NOT GIVEN if there is no information on this

9. In a widely analysed dataset, around 70 percent of patients were white.
10. The extra-care algorithm was explicitly designed by a racist programmer.
11. Models trained only on black patients performed as well for black patients as for white patients.
12. Reviewers identify four interlocking causes of AI-related disparities.
13. Medical AI has already eliminated racial health inequity.

Questions 14-15

Complete the summary below using NO MORE THAN TWO WORDS from the passage.

Because the extra-care model used future healthcare (14) __________ as a proxy for need, it reflected unequal access rather than underlying sickness. Reviewers stress that no single intervention will cure the bias, and that systems must be audited group by group.

The bias was not inserted by a racist programmer but (15) __________ faithfully from a history of unequal spending.

答案与解析

← 上一篇 | 返回雅思焦点 | 下一篇 →