雅思阅读 2: The Vanishing Fingerprint of Machine Writing（机器写作的消失指纹）

改编自 The Economist 研究报道（2026年8月）。本文为雅思阅读 Section 3 难度，约 980 词。 素材来源：The Economist — "Economist study: AI writing no longer betrayed by em-dashes"

Reading Passage

A. When researchers first began to detect the fingerprints of artificial intelligence in prose, they thought they had found a reliable tell: the em-dash. As large language models were deployed across journalism, academia and marketing, critics noticed that machine-generated text was riddled with the long, sweeping punctuation mark — used instead of commas, parentheses or semicolons to inject what the models presumably regarded as rhetorical emphasis. By 2025, the em-dash had become a cultural shorthand for AI writing, the literary equivalent of a barcode stamped on every generated paragraph. A study conducted by The Economist, comparing 1.2 million words of its own journalism against text produced by four leading chatbots, has now delivered a sobering verdict: the em-dash is no longer a dependable indicator of authorship. The giveaway that once seemed so obvious has, in the space of just a few software updates, all but disappeared.

B. The study, published in August 2026, asked four models — OpenAI's ChatGPT, Anthropic's Claude, Google's Gemini and xAI's Grok — to rewrite articles that had originally appeared in The Economist, using AI-generated summaries as prompts rather than live web access. The resulting corpus of 55,940 sentences was then compared against human-written journalism from CNN, the New York Times and the Washington Post, as well as excerpts from novels published between 1950 and 2022. On the em-dash question, the results were unequivocally mixed. Only Claude continued to use the punctuation mark more frequently than human writers. ChatGPT, meanwhile, used markedly fewer em-dashes than any other writer in the comparison, human or machine. The Economist did not publish figures quantifying the size of these gaps, but the directional finding was clear: the most famous AI tell had become, at best, an inconsistent signal.

C. If em-dashes no longer betray the machine, what does? The study identified a cluster of more subtle markers. All four models displayed a marked preference for polysyllabic vocabulary — words such as significant, increasingly and consequences — alongside rarer terms like interdependence and reindustrialisation, and a fondness for scientific jargon including parameter and methodology. This tendency toward Latinate, abstract language was accompanied by what linguists call nominalisation: the conversion of verbs and adjectives into nouns, as when "improve" becomes "improvement" and "analyse" becomes "analysis". The effect is prose that sounds elevated but emotionally flattened — the stylistic equivalent of speaking in a slightly formal, slightly vague register. Gemini and Claude exhibited these tendencies most strongly, though the study did not rank the models numerically.

D. Punctuation patterns told a complementary story. Far from being over-punctuated, AI-generated text turned out to be strikingly under-punctuated: the models used fewer commas and semicolons than their human counterparts and rarely deployed parentheses. The explanation lies partly in sentence length. Chatbots construct long, unbroken statements, accumulating clauses and subordinate ideas in a single breath, and they do not interrupt their own flow with short, punchy declarations. For moments of emphasis, the models reach for a limited repertoire of rhetorical formulas: the contrastive "not X but Y", the additive "not only... but also", and the so-called rule of three, in which three items are listed in sequence. ChatGPT and Claude deployed these constructions more frequently per thousand sentences than either the other models or the human writers. The word "and", the study found, was the single most overused word across all four systems.

E. The study's findings carry an uncomfortable implication for the detection industry. Companies such as Pangram, which claims 99.98% accuracy in identifying machine-generated text and has partnered with the blogging platform Substack, rely on algorithms trained to recognise precisely these kinds of stylistic signatures. Yet as Tommie Juzek of Florida State University observed, the models are trained on human writing and human feedback — they adopt what readers find impressive and discard what they do not. Every published tell is, in effect, a public invitation for its own removal. When The Economist itself noted that detectors are "black-box algorithms that can produce false positives and give no reasons for their conclusions", it raised a question that extends well beyond the realm of punctuation: if the stylistic markers of machine authorship are systematically erased with each software update, can detection ever remain a moving target, or will it eventually become a lost cause? The answer may determine not only how we judge the authorship of an essay or a news article, but how we distinguish between human and machine in an increasing range of professional and creative endeavours.

Questions 1-5

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

List of Headings i. The commercial prospects of AI detection tools ii. What new markers have emerged in machine prose iii. Why em-dashes were originally chosen by AI companies iv. The methodology behind a large-scale comparison v. The paradox of evolving detection vi. How punctuation patterns distinguish human from machine vii. The decline of em-dash usage in human journalism

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. What did the Economist study reveal about em-dashes? A. They are now used more by humans than by all AI models. B. They remain a reliable marker only for Claude's output. C. Their frequency varies considerably between different AI systems. D. They have been replaced by semicolons in most AI writing.

6. According to the study, AI models tend to use nominalisations because they A. were trained on scientific papers from the mid-twentieth century. B. prefer abstract, formal vocabulary over concrete verbs and adjectives. C. find them easier to generate than shorter, direct constructions. D. were specifically instructed to avoid informal language.

7. What does the study say about AI punctuation? A. It uses parentheses more frequently than human writing. B. It relies primarily on short sentences separated by periods. C. It employs fewer interrupting punctuation marks than human prose. D. It favours semicolons as the main device for joining clauses.

8. What concern does Tommie Juzek raise about AI detection? A. Detection companies overstate their accuracy to attract clients. B. The models may actively remove the features detectors rely on. C. Human writers are increasingly adopting AI-style punctuation. D. The methodology used by The Economist was insufficiently rigorous.

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. The em-dash became widely associated with AI writing before The Economist's study.
10. The four AI models were given access to live web sources when rewriting articles.
11. Gemini and Claude produced the most human-like prose according to numerical rankings.
12. Pangram has acknowledged that its detection algorithms may produce incorrect results.
13. A story that won a literary prize was definitively proven to be AI-generated.

Questions 14-15

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

The study found that AI writing exhibits a preference for long, abstract vocabulary, particularly words of (14) __________ origin, and tends to convert verbs and adjectives into nouns — a process known as (15) __________.

答案与解析

← 上一篇 | 返回雅思焦点 | 下一篇 →