The price for all the gifts of AI has long been marked in the dark

Generative AI is not only reshaping various industries but is fundamentally changing the way humans write, perceive, and think. After the release of ChatGPT 3.5, an optimistic expectation spread widely: AI will bring about "work equity." In 2023, two PhD economists from the Massachusetts Institute of Technology published empirical research in the journal Science, providing evidence for this claim: generative AI can significantly enhance the performance of low-performing employees, potentially bridging the gap between them and high-performing employees, thereby reducing inequality. The editors of Science summarized this by stating, "Weaker-skilled participants benefit the most from ChatGPT, which has important implications for future policies aimed at reducing productivity inequality through AI." However, two years later, reality does not seem to have fully followed this ideal path. In 2025, two PhD economists from Harvard University revealed a harsh truth by analyzing recruitment and employment data covering over 62 million employees and more than 150 million instances from 2015 to 2025: Generative AI is reshaping the labor market in a "credential bias" manner. Data shows that from 2015 to 2022, the employment growth curves for junior and senior positions remained largely consistent, but starting in 2023, the two began to diverge: senior positions continued to grow, while junior positions started to decline. For companies that fully embraced AI, the number of junior positions decreased by 7.7% over six quarters, while senior positions were largely unaffected and even saw slight growth. The main reason for this phenomenon is a significant reduction in hiring, rather than mass layoffs. AI has not brought about inclusive equity; instead, it has further highlighted the Matthew effect of "the rich get richer." Ctrip CEO Liang Jianzhang commented on this paper, stating, "AI will replace junior intellectual labor, exacerbating the difficulties faced by young people in education, marriage, childbirth, and the early stages of their careers." The structural changes in the labor market are just the tip of the iceberg. A deeper question arises: as AI is integrated into our workflows on a large scale, what impact is it having on human creativity itself? Is the efficiency brought by AI truly an internalization of individual capabilities? Is it shaping or even "unifying" our thoughts in ways we have yet to perceive? After individuals become overly reliant on AI, has their independent, original thinking ability been enhanced or, unbeknownst to them, weakened? Recently, the research team led by Li Guiquan at Peking University published a paper in the top sociology journal Technology in Society, which is a direct response to a series of key issues. The core of the research consists of two parts. Study one is a large-scale natural experiment that analyzes over 410,000 academic papers across all 21 disciplines before and after the release of ChatGPT-3.5, examining the real impact of AI on global knowledge production; study two is a longitudinal behavioral experiment that has been continuously tracked for months, exploring the long-term causal effects of AI on individual cognitive abilities in a laboratory environment. The research team combined techniques such as regression discontinuity design and machine learning to reveal the long-term and real impacts of generative AI on individual creativity and group homogeneity. This journal is in the JCR Q1 top tier, with an impact factor of 12.5, ranking 2nd among 271 journals in the social science, interdisciplinary category. 01 The "Collective Unconscious" of 410,000 Papers The most terrifying thing is not the noise, but the chorus of voices. Study one is a large-scale natural experiment. The research team extracted academic outputs spanning physical sciences, life sciences and biomedicine, applied sciences, social sciences, arts, and humanities from the authoritative Web of Science core database. By randomly sampling over 17,000 scholars, the team ultimately compiled all 419,344 papers authored by these scholars before and after the release of ChatGPT-3.5, creating a massive dataset to analyze the real impact of AI on global knowledge production. Illustration of the homogeneity and creativity results of academic papers before and after the release of generative AI. As shown in the figure above, before 2022, the creativity (red/blue lines) and homogeneity (gray line) of global academic output were steadily increasing. However, after the release of ChatGPT-3.5, the slopes of both curves sharply increased. That is, after the release of GPT-3.5, the academic community significantly accelerated knowledge output (creativity) while also exacerbating the homogeneity of its content at a faster rate, clearly demonstrating the "double-edged sword" effect of generative AI on knowledge production. To prove that the observed changes were caused by AI and not by coincidence, the research team employed a causal inference method known as "regression discontinuity design" (RDD) How to do it They took the release of ChatGPT-3.5 in December 2022 as a natural "time breakpoint." Whether a paper was published before or after this date involves many uncontrollable random factors for individual scholars (such as the review cycle), which approximates a randomly assigned "experimental group" (with access to AI) and a "control group" (without access to AI). Why it is reliable This "quasi-random" characteristic allows researchers to effectively isolate the interference of other long-term factors and accurately identify the causal effects brought by AI. To ensure the rigor of this method, the team also conducted a series of specialized statistical tests to confirm that scholars did not engage in large-scale "manuscript holding" or "rush publishing" strategies before and after the "breakpoint," thereby ensuring the reliability of the research results. How to quantify "creativity" and "homogeneity" indicators? After confirming the causal relationship, the research team conducted a quantitative analysis of over 400,000 papers from the dimensions of "creativity" and "homogeneity." Creativity is assessed by the "quantity" of papers published and the "quality" of the journals (JCR quartiles) in which they are published. Quantity: The total number of papers published by scholars. Quality: The JCR quartiles of the journals in which the papers are published (Journal Citation Reports Quartiles). This is an authoritative journal rating system, where Q1 represents the top 25% of journals in terms of influence in the field, and Q4 represents the bottom 25%. Homogeneity: Assessed through content similarity and language style similarity. Content similarity: Using the SBERT deep learning model, the semantics of paper abstracts are transformed into numerical "vectors," and the "cosine similarity" between these vectors is calculated to determine their similarity in core meaning. Language style similarity: Using character-level matching algorithms, phrases and sentence structures that repeatedly appear between paper abstracts are scanned and counted to measure the similarity of writing styles. A cold double-edged sword: more efficient, but also more monotonous As shown in the figure, the analysis results clearly reveal a "double-edged sword" effect. On one hand, the emergence of AI has indeed become a powerful "accelerator" for academic output: the average annual publication volume per scholar increased by 0.9 papers, and the quality of the published journals improved by an average of 6%, with this effect being particularly pronounced in fields such as technology and physical sciences. On the other hand, the increase in efficiency comes at the cost of diversity in thought and expression. Data shows that the average similarity of language styles in papers has astonishingly increased by 79% per year, while the content themes of the papers have also shown significant convergence, with the phenomenon of homogenization being most severe in physical sciences, arts, and humanities. Breakpoint regression result chart The research team from Peking University conducted a large-scale natural experiment, providing us with real-world macro evidence: generative AI is indeed a powerful "accelerator" of academic output, helping scholars produce and publish faster in better journals. However, this increase in efficiency comes at the cost of diversity in thought and expression. Global knowledge production seems to be becoming more efficient and more "monotonous" in this "great exchange." At the same time, Study One leaves a deeper question: what does this macro trend mean for each individual involved? Is the enhancement of creativity brought by AI a genuine growth in personal capability? To answer this question, the research team conducted a longitudinal behavioral experiment in Study Two, tracking participants over several months to explore the long-term causal effects of AI on individual cognitive abilities in a controlled laboratory environment. 02 The Creative Scars Left by AI Once thought succumbs to habit, it loses the possibility of creation. In fact, several laboratories have already confirmed the trends revealed by macro data through empirical research with small samples from different perspectives. For example, research from Cornell University found that AI writing assistants sacrifice cultural uniqueness, leading users' expressions to trend towards a "Western paradigm"; research from Santa Clara University also indicated that individuals using ChatGPT exhibit more semantic similarity in their creativity. Notably, a research team from the Massachusetts Institute of Technology directly observed individuals' brains using electroencephalography (EEG) technology, discovering that the brain activity levels of the group using ChatGPT were significantly lower than those of groups relying solely on their own thinking or using search engines. These studies collectively point to one conclusion: AI is enhancing efficiency at the cost of reducing cognitive input and sacrificing diversity. Illustration of participants during the EEG experiment However, most studies focus on the immediate effects of using AI, with few exploring whether the effects can persist after AI "leaves the scene," and whether its long-term negative impacts will dissipate. The research from Peking University makes a new attempt in this regard. It not only observes the immediate effects of AI during a seven-day experiment but also systematically tests the long-term consequences of AI dependence through two independent follow-up tests on the 30th and 60th days after the experiment ended. This allows us to truly see whether AI brings transferable "abilities" or a fleeting, non-internalizable "illusion." Specifically, in Study Two, the Peking University research team randomly divided 61 college students into two groups: the "AI experimental group" (which could use ChatGPT-4) and the "pure mental control group." The experimental design consists of three key stages: first, all participants do not use AI on the first day and complete a creativity baseline test; then, from the second to the sixth day, the "AI experimental group" completes daily creativity tasks with AI assistance, while the "pure mental control group" completes tasks without assistance; Finally, and most importantly, on the seventh, thirtieth, and sixtieth days, all participants must complete the final tracking test without AI assistance. Experimental design schematic To comprehensively assess "creativity," the study employed a composite task model covering multiple dimensions. These tasks include: Divergent thinking test: The classic "Alternative Uses Task" (AUT), which requires participants to come up with as many novel uses as possible for everyday items (such as "a pen"). Creative problem solving: Business scenario questions that are closer to real-world situations, such as asking participants to design innovative features for a "smart bicycle." Convergent thinking test: The "Remote Associates Test" (RAT) added during the tracking phase, which requires participants to find a word that connects three unrelated words. Insight problems: The classic "Candle Problem," which requires participants to use a box of thumbtacks, a candle, and a box of matches to fix the candle to the wall without letting the wax drip onto the table. To ensure the scientific rigor of the assessment, the study adopted the "gold standard" in the field—the Consensus Assessment Technique (CAT). Multiple expert judges independently scored thousands of creative outputs (including divergent thinking tasks and complex problem solutions) on various dimensions such as novelty, practicality, and flexibility under "double-blind" conditions, where they were unaware of the grouping and research objectives. The high data consistency (inter-rater reliability ICCs > 0.90) ensured the scientific and fair nature of the assessment results. The measurement method for homogeneity in Study Two used the exact same technical approach as Study One, ensuring consistency in evaluation standards between the two studies. Creativity: The impact of ChatGPT on AUT and problem-solving innovation tasks Homogeneity: The impact of ChatGPT on content homogeneity and language style homogeneity in AUT The experimental results clearly reveal a harsh asymmetry: The enhancement of creativity is temporary and unsustainable: During the AI usage phase (Days 2-6), the "AI experimental group" significantly outperformed the "pure brainpower group" in various creativity metrics. However, once AI was removed, this advantage disappeared instantly. From Day 7 to Day 60, there were no significant differences in creativity performance between the two groups. Alarmingly, in the convergent thinking test on Day 60, participants in the experimental group performed even significantly worse than the control group that had never used AI. What AI brought was not a transferable "ability," but rather an uninternalizable "illusion." The homogenization of thought is long-term and leaves "creative scars": in contrast to the fleeting enhancement of creativity, the homogenization of thought exhibits remarkable "stickiness." Even two months after ceasing the use of AI, the output content of the "AI experimental group" still shows significantly higher similarity to each other in both semantics and language style compared to the control group. This longitudinal tracking study provides direct causal evidence confirming the long-term impact of AI on individual creativity. What AI may bring is merely an uninternalizable "illusion of creativity," while the convergence of thought it leaves behind may become a difficult-to-eradicate "creative scar," persisting in our cognitive and expressive habits. 03 If the World Lacks New Ideas This is the best of times, and the worst of times. The conclusion of this study from Peking University is not to abandon AI entirely in the AI era. On the contrary, it aims to remind us that we must consciously understand and respond to the profound impact of long-term reliance on AI on individual thinking and cognitive habits. The "homogenization" trend revealed in the study is underpinned by profound principles of cognitive science: AI's output can easily create a powerful "anchoring effect" on users. When AI quickly generates an answer or framework that "looks pretty good," our thinking becomes "anchored" to this initial proposal, making it difficult for subsequent thoughts and creations to deviate significantly, leading to convergence of thought at the group level. In July of this year, when Jensen Huang was interviewed by CNN, he made a calm judgment: "If the world lacks new ideas, then the productivity gains brought by AI will translate into unemployment." As generative AI is used continuously, the information on the internet and the human knowledge base are becoming more homogeneous at an unprecedented speed. The research from Peking University confirms with cold data that this trend is real. If society can continuously generate new ideas, AI will translate into more diverse employment opportunities; if it only repeats old tasks, AI can complete them in seconds. AI amplifies creativity while accelerating the exit of those suffering from "idea exhaustion." 04 How to Maintain Sharp Thinking in the AI Era AI reduces our workload, but we need to establish a thinking system that can think deeply, interact with AI, describe the problems we want AI to solve, reason about the problems, and also judge whether AI has answered the questions correctly; we need to have dialectical thinking. — Jensen Huang As individuals in the AI era, how should we position ourselves? How can we enjoy the convenience of AI while avoiding falling into a desert of creativity? Based on the insights from the research, here are some specific action suggestions: Treat AI as a "thought sparring partner": Consider it an tireless "thought sparring partner" that can provide infinite perspectives. Use it for brainstorming, generating multiple possibilities, and challenging your inherent assumptions. However, the final selection, deepening, decision-making, and accountability for the results must be yours. Deliberate practice of "cognitive friction": The most effective way to combat the "anchoring effect" is to actively create "cognitive friction." Do not easily accept the first answer given by AI. Deliberately refute it, look for its logical flaws, and question aspects it may not have considered. This practice of critical thinking is key to maintaining our independent thinking ability. Set "no AI time": Just as we need to exercise regularly to prevent muscle atrophy, we also need to regularly let our brains exercise without AI assistance. Designate a regular "no AI time" each week to think, plan, and create using the most primitive paper and pen or a blank document. This deliberate "cognitive decluttering" ensures that our brain's core creative and reasoning abilities do not degrade in comfort. Author of this article: Hanqing, Source: Tencent Technology, Original title: "The price of all AI's gifts has long been marked in the dark | Interpretation of Peking University's latest paper" Risk warning and disclaimer The market has risks, and investment requires caution. This article does not constitute personal investment advice and does not take into account the specific investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article are suitable for their specific circumstances. Investment based on this is at your own risk