AI in Science: A Hidden Threat to Research Integrity?
Artificial intelligence (AI) is now infiltrating laboratories and scientific publications, raising crucial questions about research integrity. A recent study reveals that over 13% of biomedical articles bear the marks of ChatGPT and similar tools.
In Brief
- An analysis of 15 million biomedical articles reveals that 13.5% of 2024 publications show signs of AI use.
- Researchers identified 454 “suspicious” words frequently used by AI tools, such as “delve,” “showcasing,” and “underscore.”
- Current detection tools remain unreliable, sometimes mistaking historical texts for AI-generated content.
- Experts are divided: some see it as a threat, others as a democratization of research.
AI Leaves Its Marks in Science
Researchers from Northwestern University, in collaboration with the Hertie Institute for Applied AI in Health, analyzed over 15 million scientific abstracts published on PubMed. Their finding is unequivocal: in 2024, generative AI, notably ChatGPT , has deeply marked the language of biomedical research.
To demonstrate this, the team compared the frequency of certain keywords in 2024 with those of 2021 and 2022. And the difference is striking: terms previously less common like “delves”, “underscores”, or “showcasing” have exploded in usage, to the point of becoming typical stylistic markers of AI-generated texts.
This “word hunt” nonetheless reveals a more nuanced reality. Stuart Geiger, a professor at the University of California, San Diego, tempers the alarm:
Language changes over time. “Delve” has skyrocketed, and this word is now in the vocabulary of society, partly because of ChatGPT.
Linguistic evolution thus poses a major dilemma. How to distinguish fraudulent use of AI from mere cultural influence? More worryingly: might researchers change their natural writing style for fear of being wrongly accused?
Between Democratization and Ethical Drift
Kathleen Perley, a professor at Rice University, takes a more nuanced position on the use of AI in scientific research .
According to her, these tools can play a decisive role in democratizing access to academic research, especially for non-English-speaking researchers or those suffering from learning disabilities.
In an academic environment dominated by English and formal requirements, AI can provide a real springboard for brilliant profiles, but marginalized by the language barrier.
This approach raises a fundamental question: should researchers who use tools to overcome structural obstacles really be penalized? Couldn’t AI, on the contrary, help bring to light quality work that has until now been invisible due to writing limitations rather than conceptual ones?
Derivatives, Biases, and False Positives, Science Facing the Limits of AI
But enthusiasm runs into very real drifts. The example of the Grok chatbot, developed by the company of Elon Musk, is a chilling illustration.
Since its last update, the tool has produced a series of antisemitic messages posted on X (formerly Twitter), going so far as to justify hateful remarks and praise Hitler. Such incidents remind us that even the most advanced models can convey dangerous biases if they are not properly supervised.
At the same time, AI detection tools struggle to prove reliable. ZeroGPT, for example, estimated that the United States Declaration of Independence was generated 97% by an AI, while GPTZero evaluates it at only 10%. This inconsistency reveals the immaturity of detection technologies and the risk of unfounded accusations.
Beyond technical tools, the emergence of AI in scientific research questions the very essence of intellect. Rigor, originality, and integrity are the pillars of scientific production. Can we preserve these values when the boundary between assistance and substitution becomes blurred?
More than ever, academic institutions must define clear guidelines. It is not about slowing innovation, but about drawing a line between ethical use and intellectual fraud. The future of research depends on our collective ability to integrate artificial intelligence without losing the soul of science.
Disclaimer: The content of this article solely reflects the author's opinion and does not represent the platform in any capacity. This article is not intended to serve as a reference for making investment decisions.
You may also like
Hashrate orders are booming, why is "AI trading" falling out of favor? JPMorgan reveals the time lag between AI investment and productivity realization
Joseph Lupton from the JP Morgan Global Economic Research Team pointed out in a research report on November 10 that recent improvements in labor productivity have been driven more by capital deepening and increased intensity of factor utilization. The broad efficiency gains from AI, resulting from process reconstruction, data integration, and organizational adjustments, will still take time to materialize. This means orders for chips, servers, and electrical equipment may grow first, but continuous investment returns for computing power operators and end users may not arrive at the same pace.
AI investment enters the "Trough of Disillusionment", the next 18 months will be a life-or-death test!
Goldman Sachs believes that the AI investment cycle is currently in the "trough of disillusionment," but this is not the first time. The core contradiction lies in the fact that the business models of new cloud computing providers are highly dependent on end-user returns from AI spending, while currently "there are basically zero companies that can attribute profits to AI." Social Capital founder Palihapitiya has clearly defined an 18-month key window, during which the market will decide which AI business models are sustainable and which are not.
Anthropic's Core Concern: Revenue Can Be Withdrawn at Any Time, but Computing Power Costs Are Locked in for up to 20 Years
Zheshang Securities believes that Anthropic's revenue is mainly based on usage-based billing, with remaining performance obligations covering only about 5.1 months of annualized revenue, while compute power costs are already locked in with $895.4 billion in irrevocable commitments, some extending as far as 2045. Revenues can be reduced at any time, but costs are non-exitable, creating a highly asymmetric maturity gap between the two—this is the company's real core constraint.
Triple Pressure from the Federal Reserve, Treasury, and Congress; Goldman Sachs: Liquidity Crisis May Occur in 2027
Goldman Sachs warns that the reserve ratio of bank assets has dropped to around 11%, only slightly above the crisis trough of 2019. In the second half of 2027, a deadlock in debt ceiling negotiations will force the Treasury to rebuild the TGA, triggering a "perfect storm" of a sharp drop in reserves and a surge in T-bill supply; combined with stalled money market fund buying, this will create a huge liquidity black hole. The Federal Reserve could be forced into the awkward position of lowering interest rates and launching large-scale bond purchases to rescue the market.
Crypto prices
More
