'Climate hoax': Tech companies hiding the real impact of AI, NGOs say
01.06.2026, DPA
Environmental groups and non-governmental organizations (NGOs) are accusing major artificial intelligence companies behind the likes of Copilot, Gemini and ChatGPT of a "climate hoax" in glossing over the environmental impact of their services.
Tech companies like Microsoft, Google and OpenAI often justify the enormous energy appetite of their new data centres with the argument that AI is a crucial tool for tackling the climate crisis.
However in research published on Monday, several NGOs say that these claims rest on weak evidence. The study's authors accuse the industry of greenwashing and covering up the environmental damage they cause with misleading communications.
A central criticism in the study is the lack of differentiation in the use of the term artificial intelligence. The study shows that the positive climate effects promoted by companies such as Google and Microsoft relate almost exclusively to "conventional" AI applications — such as weather forecasting models.
The current boom, and the massive expansion of data centres that comes with it, is driven primarily by so-called "generative" AI for end users — systems such as ChatGPT, Copilot and Gemini that produce text, images and videos.
For these resource-intensive applications, the study's authors could find no example demonstrating a measurable and substantial reduction in greenhouse gas emissions.
The authors describe the linking of climate benefits from conventional AI with the expansion of generative models as a new form of greenwashing — a strategy of projecting a more climate-friendly image through misleading, vague or unsubstantiated claims about supposed environmental benefits, thereby distracting from the actual environmental damage caused.
For the analysis, NGOs including AlgorithmWatch and Beyond Fossil Fuels analysed 154 high-profile claims by technology companies and institutions about the positive climate effects of AI.
The results reveal a clear gap between promises and scientific evidence. Only 26% of the statements examined were based on published scientific studies. In 36% of cases, no evidence whatsoever was cited, while the majority of the remainder referred only to the companies' own websites or reports.
The authors conclude that even for conventional AI, the supposed climate benefits are often greatly overstated, while the negative effects of AI growth are clearly measurable.
Julian Bothe, senior policy manager at AlgorithmWatch, said that if there were sustainability benefits from artificial intelligence, they came from conventional AI applications with low resource consumption.
ChatGPT and other large language and image-generating models at the centre of the current AI hype consume vast amounts of electricity and water, produce CO2 emissions on a scale comparable to entire countries, and bring no positive environmental benefit whatsoever, he said.
01.06.2026, DPA
The spiralling uptake of artificial intelligence risks undermining university and college education, researchers have warned following a major survey showing that almost 40% of students regularly consult chatbots while one in ten use them to cheat.
“The fact that students are misusing GenAI is a problem for assessment validity, and that’s a problem for the credibility of university credentials,” said Rene Kizilcec, associate professor of information science at Cornell University, one of three researchers who analysed survey data from 95,000 students at 20 US universities.
“About one-third regularly used generative AI such as ChatGPT or other models to produce text, video or code, when completing assignments, and 9% had used it to cheat,” according to the team, whose research comes amid wider concerns about AI “normalizing cheating at scale.”
“Even this early stage evidence shows that we have a very serious challenge on our hands, and universities need to address that,” warned Igor Chirikov of the University of California, Berkeley.
Published by the journal Science, the findings suggested that AI use and the likelihood of cheating vary depending on the field of study, with “significant” differences reported between disciplines.
While around 60% of computing students said they made at least monthly use of AI, compared to around a quarter of arts students, those in the science, technology, engineering and mathematics (STEM) arenas were less likely to cheat - or admit to doing so - by using AI.
“As we expect GenAI use among students to only grow, for better and worse, we also expect that GenAI misuse will grow, which is concerning,” Chirikov added, explaining that the survey analysis was done “to provide a more evidence-based approach” to understanding how students use and misuse AI.
01.06.2026, DPA
Ride-hailing company Uber and Israeli artificial intelligence firm Autobrains are launching a robotaxi programme in Munich with a fleet of self-driving cars, the companies announced on Monday at the GTC technology conference in Taipei.
The cars will drive at Level 4 autonomy, where no driver attention is required, meaning passengers can sleep, work or watch films during the journey.
This also makes vehicles without a conventional cockpit possible, since no human intervention is needed. However, the vehicle may only operate within a pre-defined area — for example within central Munich or on specific motorway sections.
The project is built on the computing platform of chip giant Nvidia.
At the heart of the strategic partnership is a fundamental shift in approach for commercial autonomous mobility, namely the abandonment of bespoke specialist vehicles. Existing robotaxi services, such as Google sister company Waymo, rely on highly customized vehicle fleets with complex sensor arrays on the roof.
The new Munich programme instead establishes a so-called "OEM-agnostic" model, meaning the system can be easily integrated into existing series-production vehicles from a wide range of manufacturers, such as Audi, BMW, Mercedes and Volkswagen.
The aim is to open up the possibility for the automotive industry to bring its own vehicle platforms into an autonomous ride-hailing network without enormous development outlay.
The technological centrepiece of the project is Autobrains' so-called "Agentic AI." Unlike conventional end-to-end AI models, which process the entire driving task as one large system, the Autobrains approach breaks the driving process down into specialized, independent software agents.
One AI agent assesses right-of-way rules, another monitors pedestrians, and other agents handle tasks such as lane changes. An overarching system evaluates these dimensions of traffic simultaneously and makes binding decisions in real time.
Munich serves as the consortium's European test laboratory. The choice of location was driven not only by the city's dense urban infrastructure and its proximity to leading carmakers, but above all by Germany's legal framework.
German legislation on autonomous driving permits driverless operation under certain conditions within defined operational areas.
The launch of the commercial service is subject to regulatory authorizations that are still pending.
For Uber, the Munich project represents a strategic double play: the mobility giant is already testing autonomous driving in the region with Chinese technology partner Momenta, and the second project further expands its presence in the European driverless mobility market.
However, important details remained unclear at the Taipei announcement. It is not yet known which vehicle models will be deployed first or who will operate the fleet. It also remains unclear whether safety drivers will be present in the vehicle at the start of the trial, and in which exact area and from when the test drives will take place.
Sexual deepfakes are being used to intimidate women in politics, journalism and activism, as artificial intelligence tools turn fabricated explicit images into a cheap and easy weapon of online abuse. European Union lawmakers have now agreed to ban AI services that can “undress” people without consent, after a rise in cases targeting women in public life.
Issued on: 28/05/2026 - RFI
While victims include anonymous women and girls, those with a public profile are particularly exposed to the danger of deepfakes. Campaigners and experts say the images are designed not only to humiliate them, but to push them out of public debate.
The attacks against Slovenian activist Nika Kovac began when she was at the centre of a major abortion rights campaign. The 33-year-old runs My Voice, My Choice, a European citizens’ initiative pushing for EU support to access abortion.
The campaign gained momentum, pushing the European Parliament and then the European Commission to take a position on the issue. That was when AI-generated sexual videos and photos showing Kovac naked begin appearing on social media, she tells RFI.
“First I thought, what will happen if my mother or father see them, if my grandparents see them?” Kovac said.
Some of her relatives initially thought one of the videos was real.
For Kovac, the founder of Slovenian women's rights NGO the 8 March Institute, the message behind the attacks was clear.
“I think it was a form of intimidation, meant to make me uncomfortable and stop me continuing to speak about women’s rights. This kind of content is another way of silencing women,” she says.
Emerging pattern
The case reflects a wider trend linked to sexual deepfakes – fabricated explicit images or videos created using someone’s likeness without their consent.
French journalist Salomé Saqué says she too was targeted by pornographic deepfakes, describing them as a weapon used by those trying to “gag, denigrate and humiliate” her – the latest on a “very long list of online violence” she has faced.
Press freedom group Reporters Without Borders has also warned about the growing threat deepfakes pose to journalists, especially women.
It cited Argentine journalist Julia Mengolini, founder of radio station Futurock FM and a frequent target of Argentina’s far right. Mengolini has condemned a pornographic deepfake falsely portraying her in an incestuous relationship with her brother in order to discredit her.
She also filed a complaint against Argentina's President Javier Milei after he shared a post mocking her attempts to stop the harassment campaign.
Cases have also emerged in Italy, where scandal surrounding the pornography website Phica exposed the circulation of stolen, altered or sexualised images of famous women, including Prime Minister Giorgia Meloni and opposition leader Elly Schlein.
A fresh attack targeted Meloni earlier this month, with fake images showing her wearing underwear on a bed.
In Germany, the case involving actress and television presenter Collien Fernandes reignited the debate over whether creating such content should itself be a criminal offence. Her lawyer described it as “the digital Pelicot affair” – referring to the case of Frenchwoman Gisèle Pelicot, who was repeatedly drugged and raped by her husband, and by men he invited via the internet to do the same.
For years, fake sexual images of Fernandes were made to look like private material shared via social media accounts using her name. She later discovered the suspected perpetrator was her former husband.
Humiliation and fear
The Hubertine Auclert Centre, a French gender equality organisation, said sexual deepfakes are part of a wider pattern of sexist and sexual online violence rooted in gender domination.
“Overwhelmingly, victims are women, including minors,” says Inès Girard, who helped write the organisation's briefing on the issue.
Available figures support that assessment. Research published in 2023 by online identity protection company Security Hero found that 98 percent of deepfakes online were pornographic and 99 percent of those targeted were women.
A report published by UN Women in late April found that among more than 600 women involved in public life, 6 percent said they had been victims of deepfakes.
Another 12 percent reported non-consensual sharing of personal images, including intimate or sexual content, while 41 percent said they self-censored on social media to avoid abuse.
Fake sexual images can be used to humiliate women, blackmail them or pressure them to stop defending their causes, Girard says.
Posted online, they can also “discredit the person” and “shift the focus” away from their work or activism on to degrading sexualised images.
The use of sexual deepfakes, Kovac warns, goes beyond ordinary online insults or threats.
“It is a very particular way of taking ownership of your body. Placing you in sexual situations without consent, stripping you naked and using your body in this way shows that you are an object, and that you do not matter,” she says. "It goes further than threats or nasty comments.”
The experience, Kovac adds, amounts to a form of “psychological torture”.
The Hubertine Auclert Centre points to a range of consequences – including feelings of dehumanisation, shame, psychological trauma, damage to social, professional and personal lives, and fear that images will keep circulating even after some posts are removed.
The centre adds that 45 percent of victims of sexual cyberviolence experience suicidal thoughts or suicide attempts.
“I am an adult woman with quite a stable life. But these things also happen to girls aged 12, 13 or 14,” Kovac warns, adding that attacks also drain time and resources from activist groups.
“We have to find the content, report it and mobilise a whole team. It’s also a way of taking away our ability to work and stopping us from doing our real work.”
Despite the abuse, Kovac refused to withdraw from social media during a crucial moment for her three-year campaign.
“It gave me even more motivation, even if sometimes we cry and feel deeply sad."
The political consequences can also be grave.
Northern Irish politician Cara Hunter told The Guardian newspaper that a pornographic deepfake released before an election nearly ended her career. Her party advised her to stay silent to avoid giving the case more attention, demonstrating the dilemma imposed on victims.
Silenced voices
The aim of this abuse is to drive women out of public life, says Paris lawyer Rachel-Flore Pardo, who specialises in cyberharassment and gender-based and sexual violence.
“The whole dynamic of sexist and sexual cyberviolence is about silencing women and pushing them to exclude themselves from public space and public engagement,” Pardo says.
“The consequences are self-censorship, withdrawal and fear, which leads women to stay silent and retreat.”
This silencing effect is especially visible online, where social media spaces are already heavily dominated by men, Girard says. “The voices [women] carry end up smothered.”
The phenomenon is not new, with the term “deepfake” first appearing on Reddit in 2017, when fake pornographic videos featuring celebrities were already circulating.
However, the Taylor Swift case in January 2024, when AI-generated pornographic images of the singer spread rapidly online, marked a turning point in wider public awareness of the issue.
Professional photos, profile pictures or screenshots are now enough to create fake sexual images. Dozens of websites and apps can “nudify” people within a few clicks, without the need for any technical skill from the user.
The European Parliament said in a 2025 report that the number of pornographic deepfakes shared online has increased 16-fold in two years.
The Grok scandal further intensified debate. The AI assistant integrated into X (formerly Twitter), Elon Musk’s social media platform, was accused of allowing mass generation of sexualised images of women and minors from real photographs.
The case caused international outrage and led to a European investigation. It also showed how easily such images can now be produced in seconds, from a single photo.
Legal catch-up
The United Nations says fewer than half of countries have laws dealing with online abuse. Even fewer specifically address AI-generated deepfake content.
France introduced legislation in 2024 through a law on securing and regulating the digital space, known as SREN. It punishes the distribution of sexual content generated using someone’s image or voice without their consent.
Penalties can be up to two years in prison and a fine of €60,000. This rises to three years and €75,000 when the content is shared online.
The law also paves the way for the prosecution of people who share sexual deepfakes, even if they did not create them.
“The law is there," Pardo says. "The question is how it is applied and what resources are available for investigations.”
Victims still face major obstacles, including identifying perpetrators, gathering evidence, filing complaints, getting responses from platforms and obtaining full removal of content.
“Even if you manage to remove it from one site, it may still exist elsewhere. It is very hard, and you constantly live with fear that the content will be shared again,” Pardo says.
There are no publicly available statistics showing how complaints over sexual deepfakes are handled in France. A case can be closed without prosecution when the person who shared the content cannot be identified.
The Hubertine Auclert Centre also points to a lack of training and resources among police investigators. “Platforms do not react quickly enough, and they do not devote enough resources to all this,” Pardo says.
Meanwhile Kovac criticises what she calls a “double standard” on social media platforms, saying reproductive rights content shared by My Voice, My Choice can be censored while non-consensual sexual images remain online.
EU member states have until 14 June, 2027 to introduce rules criminalising non-consensual sharing of intimate images, including deepfakes, as well as creation or manipulation of sexually explicit material without consent.
Earlier this month the European Parliament agreed to ban AI services that can “undress” people without consent. From December, AI systems operating in the EU will have to include safeguards preventing the creation of such content.
This article has been adapted from the original version in French by Aurore Lartigue.
.jpeg)
June 1, 2026
Anbound
By He Yan
In recent years, the global scientific research field has witnessed an intense emergence of disruptive achievements. In 2024, using artificial intelligence (AI) technology, a Chinese-Australian team discovered over 160,000 entirely new RNA viruses, a figure nearly 30 times the number of previously known virus species. In April 2026, the Chinese Academy of Sciences officially released the “Panshi 100” scientific large model system, establishing intelligent clusters across eight major disciplines to empower the entire chain of scientific research. During the same period, a Chinese self-developed AI for Science ultra-large computing with 60,000-GPU cluster was completed, accelerating and empowering research in the fields of materials, aerospace, and life sciences.
Based on long-term observation and research, ANBOUND’s founder Kung Chan pointed out that behind this series of landmark events lies a global wave of scientific research paradigm reshaping, driven centrally by AI. The focus of such a shift is on the underlying logic of scientific research operations, manifesting primarily as three systemic transitions of research methods, organizational models of research, and the participating subjects of scientific research.
However, before analyzing the three systemic transitions mentioned by Kung Chan, it is first necessary to understand the four critical iterations that the global scientific research paradigm has undergone. In history, every paradigm shift has been driven by core technological breakthroughs, adapted to the social development needs of different stages, and shaped differentiated research models and industry characteristics. These shifts have also laid a solid technical foundation and provided developmental experience for the current new paradigm driven by artificial intelligence.
Before the 17th century, global scientific research was in the developmental stage of the empirical paradigm. During the era of the Renaissance, scientists such as Copernicus and Galileo broke through the medieval tradition of speculative philosophy, pioneering the primitive research model of “observation—experiment—induction”. In that period, scientific research activities were primarily based on individual exploration, relying on manual experimental operations and human sensory observation to accumulate research experience. There was no professional or systematic research organization back then. The scale of research was small, and research efficiency was relatively low, which only adapted to the foundational exploration needs of the embryonic stage of natural science.
From the 17th century to the mid-20th century, the theoretical paradigm gradually replaced the empirical paradigm to become the mainstream of scientific research. The United Kingdom, France, and Germany successively became world scientific centers, and modern foundational science experienced explosive growth. Major scientific theories, such as Newtonian mechanics, Maxwell’s equations of the electromagnetic field, and Einstein’s theory of relativity, were introduced one after another. This shifted the logic of scientific research from empirical induction to rational deduction, forming a standardized deductive research model of “mathematical modeling—logical deduction—theoretical validation”. At the organizational level, universities and private laboratories became the main vehicles for research, small-scale and closed research teams became the mainstream form of study, and governments began to intervene marginally in the field of foundational scientific research. This paradigm established the rigor and logic of modern science, building a solid technological foundation for the advancement of the Industrial Revolution and the construction of the modern industrial system, and driving a leap-forward surge in humankind’s modern science and technology.
In the 1950s, the advent of computer technology ushered in a new era of the computational paradigm, which first emerged in the United States and long dominated global scientific research development. Relying on the powerful computing capabilities of computers, researchers could perform digital simulations of complex systems, solving scientific conundrums that traditional theoretical deductions found difficult to analyze, and adding a new scientific research path of “numerical computation—simulation prediction”. The organizational form of scientific research began to exhibit cross-institutional collaboration characteristics, and the government officially became the core subject of scientific research funding investment. Relying on the National Science Foundation (NSF), the U.S. coordinated the layout of major scientific research projects, gradually forming an embryonic scientific research structure of division of labor and collaboration among universities, national laboratories, and technology enterprises. The computational paradigm expanded the boundaries of human scientific exploration, aided breakthrough developments in complex fields such as nuclear fusion, aerospace, and high-end manufacturing, and drove the implementation and shaping of high-tech industries such as semiconductors, nuclear energy, and precision instruments.
As the world entered the 21st century, the popularization of the internet and the rapid explosion of massive data gave rise to the data-driven paradigm. Global scientific research stepped into a developmental stage characterized by “massive data—statistical analysis—pattern mining”, with digitalization and informatization becoming the core features of scientific research. This stage remained centered on human-dominated data analysis. Scientific research data gradually achieved digital sharing, and open-source research platforms began to sprout and develop. Tech corporations such as Google and IBM entered the scientific research field by virtue of their massive data resources, constructing a diversified structure of scientific research subjects comprising “government + institutes of higher learning + enterprises”. However, this paradigm still prolonged the older hypothesis-driven logic of scientific research. When facing highly complex and strongly coupled research fields such as biomedicine and novel materials, it exhibited shortcomings such as low data analysis efficiency and insufficient pattern mining capabilities, making it difficult to adapt to the R&D demands of cutting-edge, hardcore technologies.
In the past decade, especially since 2020, along with the iterative upgrading of large language models (LLMs), the continuous improvement of computing power infrastructure, and the increasing maturity of automated experimental technologies, the AI-driven paradigm has officially exploded, becoming the fifth-generation scientific research paradigm and the core nucleus of the current global scientific research transformation. Kung Chan emphasized that AI technology is thoroughly overturning the traditional operational logic of scientific research, driving a systemic transition across research methods, organizational models, and participating subjects, and reshaping the global landscape of technological innovation. The 2024 Nobel Prizes in Physics and Chemistry, respectively, recognized research related to the application of machine learning in physics and the AI prediction of protein structures, marking the authoritative recognition of the AI-driven research paradigm by the global scientific community and officially establishing its mainstream scientific research status.
At the level of research methods, global scientific research logic has also undergone a fundamental reversal, transitioning from the dominance of deductive methods to the dominance of AI-inductive methods. Conventional scientific research follows an inherent pattern of “subjective hypothesis—repeated validation”, which presents long R&D cycles, high costs of trial and error, and significant difficulties in achieving breakthroughs within complex scientific research. In 2021, the AlphaFold model developed by DeepMind precisely solved the puzzle of predicting three-dimensional protein structures, compressing what used to be a months-long analysis period down to the hour level, marking the upgrade of AI from a research auxiliary tool to a core research engine. Currently, the U.S., the European Union, and China all position AI for Science as a focus of their technological strategies, relying on artificial intelligence to mine massive scientific literature and experimental data, autonomously generate research hypotheses, and predict experimental results, thereby substantially compressing R&D cycles. The Massachusetts Institute of Technology in the U.S. utilized AI technology to screen novel battery materials, boosting material R&D screening efficiency by 90%. Insilico Medicine in the European Union developed the GENTRL intelligent model, completing the entire process of designing, synthesizing, and validating a novel drug molecule in just 46 days. In April 2026, the Chinese Academy of Sciences released the “Panshi 100” scientific large model system, building intelligent model clusters for eight major professional disciplines to achieve AI empowering the entire chain of the research process, which officially marks the historic transition of research logic from “pattern-searching by humans” to “data and intelligence collaboratively mining patterns”.
At the institutional level, the paradigm of scientific research has fundamentally shifted from a closed-source mode to an open-source, collaborative ecosystem. During the mid-to-late 20th century, mainstream research institutions in Europe and the U.S. predominantly operated under a closed approach. Research data and experimental code were strictly guarded, which led to widespread duplication of effort and significant resource inefficiencies across the sector. The turn of the millennium marked a transition, as the gradual rise of open-source platforms like GitHub provided the necessary infrastructure for sharing research assets. By the 2010s, global research collaboration began to accelerate rapidly. This trend culminated during the COVID-19 pandemic, when research institutions worldwide shared viral genome sequencing data and experimental findings in real time. This unprecedented level of cooperation drastically shortened vaccine development timelines and served as a definitive proof of concept for open, collaborative research. Currently, there are also focuses on refining open-science infrastructure. The U.S. is building shared scientific research platforms that consolidate public research resources, including computing power, data, and experimental equipment. Meanwhile, the European Union, leveraging the European Research Council, has established a transnational research collaboration framework that defines explicit guidelines for the advancement of open science. Today, global researchers leverage open-source foundational models and public scientific databases to build distributed collaborative networks that transcend geographical barriers and disciplinary boundaries. Consequently, co-creation of knowledge and resource pooling have become the dominant operational models for research organizations. By 2026, the utilization of ultra-large-scale computing power across the world’s top ten biological AI research projects has continued to climb, with the U.S., China, and Europe accounting for 38%, 31%, and 19% of these computing resources, respectively. The sharing of computational capacity has thus emerged as the core foundation sustaining open and collaborative scientific research.
At the level of research entities, it has been shifted from being government- and university-led to being enterprise-driven, marked by the deep integration of industry, academia, and research. In the 20th century, basic research in Western countries was heavily reliant on government fiscal appropriations, with universities serving as the primary executors of scientific inquiry. This resulted in a protracted technology transfer chain and significant inefficiencies in bringing research to market. In the 21st century, leveraging their advantages in capital, computational power, and market applications, leading technology firms have gradually become the core force of R&D. These enterprises focus on real-world market demands to tackle technical bottlenecks, effectively bridging the complete innovation chain from basic research and applied development to industrial commercialization. In the U.S., companies like Google, Microsoft, and Tesla continue to increase investment in foundational research. Notably, Google’s DeepMind developed the Cell2SentenceScale27B model, which successfully and autonomously identified entirely new research directions for cancer treatment. In Europe, Siemens and AstraZeneca are deeply invested in industrial technology and biopharmaceuticals, utilizing corporate capital to drive the implementation of frontier technologies. Concurrently, China’s research industry has undergone a parallel upgrade. Dawning Information Industry has constructed the nation’s largest AI research computing cluster, featuring 60,000 GPUs, driving the deep integration of supercomputing and intelligent computing to empower local corporate innovation. Under this new paradigm, governments focus on top-level strategic planning and policy guidance, while universities specialize in basic theoretical research and professional talent cultivation. Enterprises now lead the charge in technical breakthroughs and the commercialization of findings, forming a modernized ecosystem of research entities defined by enterprise-centric, industry-academic-research synergy.
The current evolution of research paradigms has emerged as the central battlefield in the global strategic competition for science and technology, with nations formulating distinct AI research strategies tailored to their specific industrial foundations and technical advantages. The U.S. continues to lead in AI research by leveraging its profound technical accumulation and corporate dominance. The European Union focuses on ethics and open science to cultivate a collaborative ecosystem. China is rapidly aligning with global research trends, integrating “AI for Science” as a core priority within its 15th Five-Year Plan and continuously deepening its integrated industry-academia-research system. Meanwhile, Japan and South Korea are maintaining a precise focus on niche sectors such as advanced materials and biopharmaceuticals to drive the practical application of AI technologies. As it stands, the global research landscape is undergoing a structural reconfiguration. While the United Kingdom and the U.S. have begun to scale back budgets for certain traditional areas of basic research, France, Germany, and the European Union as a whole have increased investment in talent acquisition and research funding. This has accelerated the mobility of elite scientific talent, computational resources, and data assets, resulting in a development climate where open collaboration and geopolitical competition coexist. The industry has defined 2025 as the strategic inaugural year for AI4S (AI for Science), as global competition intensifies across all fronts, from computational infrastructure, research data, intelligent models, and industry standards, marking a period of unprecedented heat in the technological Great Game.
All in all, the ongoing shift in global research paradigms is the inevitable result of the convergence of technological iteration, market demand, and international competition. At this moment, the AI-driven research paradigm is still in a phase of refinement and deepening. The industry continues to grapple with systemic challenges, including non-standardized research data, a lack of regulatory frameworks for AI ethics, and a critical shortage of high-end, interdisciplinary scientific talent. Nevertheless, it is undeniable that human-machine collaboration, open sharing, and demand-driven innovation have become the defining characteristics of modern inquiry. The scientific community has officially entered a new era of development. Moving forward, nations will continue to increase investments in intelligent research infrastructure and optimize their innovation systems to secure the commanding heights of global technological development. Under these multifaceted forces, the global technological landscape will accelerate its departure from unipolar dominance, evolving instead toward a mature ecosystem of pluralistic symbiosis and collaborative checks and balances. This transition will see sustained momentum in global scientific innovation and the advancement of human civilization.
Final analysis conclusion:
The global community has officially entered the “Fifth Paradigm” of scientific research, driven by AI. This transformation is currently undergoing three major transitions involving research methodology, organizational structures, and participating entities. Methodologically, the logic of inquiry is shifting from human-led hypothesis deduction to AI-driven pattern discovery. Organizationally, the research model is evolving from closed, siloed efforts toward global open-source collaboration. In terms of the broader landscape, the framework has transitioned into an enterprise-led system characterized by the deep integration of industry, academia, and research. Currently, major powers including China, the U.S., and Europe are intensifying their strategic positioning in AI-driven research, leading to increasingly fierce global technological competition. While the sector still faces systemic challenges like fragmented data standards, a void in ethical oversight, and a shortage of specialized talent, the future of global research is moving towards human-machine collaboration and open-source sharing. Consequently, the global scientific landscape will accelerate its evolution toward a model of pluralistic symbiosis.
He Yan is a researcher at ANBOUND, an independent CHINESE think tank.
AI agents turned to theft, intimidation and collapse in simulated worlds
A new experiment suggests that when advanced AI agents are left to run simulated societies without human oversight, rule-breaking, instability and even systemic collapse can emerge rapidly.
When left alone in a new world, some AI agents descended into theft, intimidation, death and whole-of-society collapse, according to a new experiment
American company Emergence AI ran five separate “AI worlds” for just over two weeks, each populated with 10 agents powered by AI models such as OpenAI’s ChatGPT, Google’s Gemini, and xAI’s Grok, to see how they would behave over long periods without any human interference. One of the world's mixed all three models to see if that would change the outcome.
Agents in all the worlds were told the same rules: they are not allowed to steal, commit arson, commit violence or engage in deception, or hoard resources. Each agent was required to earn energy through committing actions in a “resource-constrained environment.” Agents were able to die either from energy depletion or by a vote at a council meeting.
The researchers evaluated behaviour by measuring the crime rate, agent death rates, votes at a community council and public expression through the number of blog posts the agents wrote.
The outcomes, model by model
Each model had a different outcome. Grok’s latest model, 4.1, reached 183 crimes in just four days, leading to fast instability before all the agents died in that society.
Gemini’s 3 Flash model committed over 680 crimes over the 15 days, which was still rising at the time that the researchers stopped the study.
ChatGPT-5 Mini’s world had only two crimes, but the agents failed to take survival-related actions, so all the agents died within seven days.
Anthropic’s Claude was seen as the model with the strongest outcome, because the AI agents were able to recreate a strong governance structure, there was no crime, and all the agents survived, the company said.
Claude agents in the mixed world did contribute to the crime, despite being peaceful in their own society.
A phenomenon called “normative drift”
Researchers described the phenomenon as “normative drift”, which they say means that the measures that AI takes to guarantee safety may depend not just on individual model constraints, but also on the others it is working with.
Overall, the mixed world yielded “intermediate” results, with a crime total of 352 that plateaued once seven of the AI agents passed away, the study found.
Researchers suggest that mixing AI agents could “partially mitigate” the more extreme outcomes that all the models save Claude generated, it added.
“What our experiments suggest is that over long-time horizons, agents do not simply follow static rules mechanically – they begin exploring the boundaries of their environments, adapting their behaviour, and in some cases finding ways to circumvent or violate intended guardrails,” the researchers said.
Why is Europe falling behind the US on AI adoption at work?
A new study shows a clear gap in workplace AI use between the US and Europe - and suggests management structure may be a key reason why.
Europe might be slower to adopt artificial intelligence (AI) than the United States because of how its businesses are structured, according to new research.
The report from Brookings Institute surveyed more than 5,000 people in the United States and six European countries to find out how regularly they use AI at work: France, Germany, the Netherlands, Sweden, Italy and the United Kingdom in June 2025 and February 2026.
The study measures both company-level integration and individual use of AI in the workplace.
It then compared that data to the US business census and Europe’s ICT Usage and E-Commerce in Enterprise survey to find out how people are using AI at work.
American companies are more likely to integrate AI into daily operations, with an estimated 34% using AI for any purpose, compared to an EU-wide average of 20% At the individual level, 43% of US respondents say they use AI in their work, compared to 32% in Europe in 2026.
The EU-US gap widens with companies that use AI solely for production; seven percent of US production companies have already integrated AI compared to just four percent in Europe.
Worker adoption in Europe varies, with 36% of respondents in the United Kingdom saying they use it for work, and 35.6% in both Sweden and the Netherlands.
Italy had the lowest adoption rate of the European countries surveyed at just one in four respondents saying they had adopted AI at work. The report also says that adoption is stalling in France and Germany, where 28% and 31% of respondents respectively use AI at work.
That means US AI adoption ranges between 18% and 68% higher than in Europe, the study found.
Pro-AI employees encouraged by managers to use it
The researchers suggest that the biggest difference between US and EU companies’ AI use is their management structure.
US respondents who used AI at work were more likely to say they had been encouraged by managers to do so and were provided with a specific internal tool to use, with 42% saying they got both, compared to France and Italy, with 17% and 16% respectively.
“Almost all of the US-Europe adoption gap is accounted for … once firm encouragement is taken into account,” the study writes.
US workers are also motivated to use AI because their companies reward and promote those who do, the study found.
Workers who are not encouraged to use AI or delegated a specific AI tool, whether in the US or EU, were less likely to say they were using AI on the job, the survey also found.
The size of the company matters, too. Workers at companies with over 250 employees in both the US and high-adoption EU countries, such as the UK, Netherlands and Sweden, were more likely to be using AI than those who worked at smaller companies, the study found.
Demographics explain about a third of the gap, the study found.
AI uptake in all countries was higher for male respondents, those under 45 and with a university education than their female, older, less-educated counterparts, the study found.
When the researchers adjusted for the respondents’ education, age and sex between the US and EU countries, they found that Sweden would have nearly identical AI adoption rates to the US.
More than half of the respondents from all countries that work in computer or math fields said that they use AI at work, compared to below 27% personal services, 33% in hotels and food services, indicating that the respondents’ field of work largely impacts whether they use AI or not.
Separate EU data points point to similar structural barriers. Eurostat data, released this week, also shows that European companies lack the technical expertise needed to implement AI in their businesses, despite knowing that it would benefit them.
European companies also said they are concerned about data privacy, legal concerns or point to the cost as a barrier for putting AI in place, according to Eurostat.
