Training the Machine, Testing the Treaty: Generative AI and the Limits of Cross-Border Governance
Introduction
Somewhere right now, a model is reading a library. It does not read like you or I, savouring a good line or skipping to the end. It does not copy, clean and convert text into numbers at the scale that a human reader could. The copyright questions these models raise are not new, but they are new in scale, and they implicate a legal order that prizes borders. A book written in Lagos can be replicated on a server in Virginia, used to train a model sold in Berlin. Whose laws apply? The honest answer is: we don't know.
A Treaty Written Before Computers
The Berne Convention is one of the most important international agreements on copyright, with over 180 countries being party to it. According to Article 9(1) of the Berne Convention, authors have the exclusive right to allow or prohibit the reproduction of their works in any form or way. Article 9(2) provides that certain countries may implement exceptions to the rights granted by the convention, as long as they are consistent with the three-step test. More specifically, a national law exception should be confined to a certain category of works, not prejudice the work’s normal use, and not unreasonably prejudice the rights of the author.
This explains how the Berne Convention could apply to copyright protection for AI. Developers of such technology often argue that since the output of a given model rarely resembles anything learned from the examples provided, there is no copying occurring. However, the Agreed Statements of the WIPO Copyright Treaty provide that the reproduction of a protected item in digital format also constitutes a reproduction. Therefore, if a work is used to train a model, this can be considered a reproduction, and the question then emerges whether a national exception applies.
Three Jurisdictions, Three Answers
Countries have responded in different ways to that question. The United States draws on its fair use doctrine, a flexible, four-factor test that was never intended for machine learning but dominates AI-related disputes anyway. Japan’s Article 30-4 permits uses of works for the purpose of data analysis where the use is not for the enjoyment of the work’s expression, which captures a lot of what training entails. And the European Union’s approach falls somewhere in the middle: its text and data mining rules allow mining of lawfully accessible works, but rightsholders can specify in a machine-readable manner that they do not want their works mined, and developers are required to seek permission.
A recent roundtable from the Columbia Undergraduate Law Review argues that the EU model fits the three-step test best because the opt-out keeps a licensing market alive and gives authors control before their work is used rather than after. It argues that broad US fair use and Japan's provision both struggle at the second and third steps, since they can undercut a foreseeable licensing market. That is one scholarly view and not settled law. The EU system also has its own weak spot: an opt-out only protects authors who know it exists and have the technical means to use it.
Bartz v. Anthropic in Brief
Bartz v. Anthropic is the case that puts these debates in front of a judge. Three authors, including Andrea Bartz, Charles Graeber and Kirk Wallace Johnson, filed a lawsuit against the Anthropic company for using part of their works to create the central library used by Claude and to train the models used in its creation. In addition, in June 2025, Judge William Alsup approved the separation of the case, considering each instance of use separately.
On training, he granted summary judgement for Anthropic, holding the use was fair and describing it as exceedingly transformative. He compared it to a reader who studies great writing in order to produce something new. It mattered that the authors did not allege Claude's outputs reproduced their books, so the case turned on inputs alone. He also held that buying print books and scanning them into digital copies was fair use, since it only changed the format of copies Anthropic already owned. But he refused to excuse the more than seven million pirated books Anthropic downloaded to build a permanent library. Building that library from pirated copies, he found, was its own use and not a transformative one, and that part was set for trial.
Two observations are worth making in relation to those thinking across borders. First, the court's assumption that there might be a market for training licences, but its observation that this is not a market which the Copyright Act entitles authors to control, is quite inconsistent with Berne's concern for normal exploitation. Moreover, the order makes no reference to Berne at all, bearing out the observation that the US courts apply the domestic statute and not the treaty.
The story did not end there. Anthropic agreed to a $1.5 billion class settlement covering roughly 482,000 works, and the court granted final approval on 20 July 2026. It resolves claims about past acquisition and copying, so the training ruling remains a trial court's view rather than binding precedent.
Where Borders Break the System
The deeper trouble is that none of this travels well. Berne is not self-executing in the United States, so judges apply the domestic statute, and nobody is required to ask whether a fair use ruling passes the three-step test. Meanwhile, training can happen in one country, on work from dozens of others, for a product used everywhere. The same book might be freely usable in Tokyo, subject to an opt-out in Paris and defended as fair use in San Francisco.
The roundtable suggests one way forward: a new WIPO special agreement, along the lines of the WIPO Copyright Treaty, in response to the internet. It would provide guidance on the application of the three-step test to training, regard licensing markets as an ordinary part of exploitation, and require minimum standards for reservations of rights and compensation. There is one major potential obstacle: many of the most influential AI companies are based in the US, which has a record of opposing international obligations that impinge on fair use. A deal would apply only to the extent that it is not controlled by the US, where most of the training takes place. Market forces, in the form of the so-called Brussels effect, may prove more influential than negotiations at WIPO.
Conclusion
Bartz’s analysis illustrates the importance of characterising a specific use at the level of a particular court. While the former had been viewed as an activity and the latter as an acquisition, the distinction between the use of a work and its acquisition appears more likely to transcend borders than the concept of fair use. Berne was designed to create a common ground for authors. Whether this convention would be able to fulfil its function in the era of generative AI depends on the willingness of nations to adopt a joint understanding of what constitutes its foundations for machines.
Related Blogs
Introduction
The unprecedented cyber espionage attempt on the Indian Air Force has shocked the military fraternity in the age of the internet where innovation is vital to national security. The attackers have shown a high degree of expertise in their techniques, using a variant of the infamous Go Stealer and current military acquisition pronouncements as a cover to obtain sensitive information belonging to the Indian Air Force. In this recent cyber espionage revelation, the Indian Air Force faces a sophisticated attack leveraging the infamous Go Stealer malware. The timing, coinciding with the Su-30 MKI fighter jets' procurement announcement, raises serious questions about possible national security espionage actions.
A sophisticated attack using the Go Stealer malware exploits defense procurement details, notably the approval of 12 Su-30 MKI fighter jets. Attackers employ a cunningly named ZIP file, "SU-30_Aircraft_Procurement," distributed through an anonymous platform, Oshi, taking advantage of heightened tension surrounding defense procurement.
Advanced Go Stealer Variant:
The malware, coded in Go language, introduces enhancements, including expanded browser targeting and a unique data exfiltration method using Slack, showcasing a higher level of sophistication.
Strategic Targeting of Indian Air Force Professionals:
The attack strategically focuses on extracting login credentials and cookies from specific browsers, revealing the threat actor's intent to gather precise and sensitive information.
Timing Raises Espionage Concerns:
The cyber attack coincides with the Indian Government's Su-30 MKI fighter jets procurement announcement, raising suspicions of targeted attacks or espionage activities.
The Deceitful ZIP ArchiveSU-30 Aircraft Acquisition
The cyberattack materialised as a sequence of painstakingly planned actions. Using the cleverly disguised ZIP file "SU-30_Aircraft_Procurement," the perpetrators took benefit of the authorisation of 12 Su-30 MKI fighter jets by the Indian Defense Ministry in September 2023. Distributed via the anonymous file storage network Oshi, the fraudulent file most certainly made its way around via spam emails or other forms of correspondence.
The Spread of Infection and Go Stealer Payload:
The infiltration procedure progressed through a ZIP file to an ISO file, then to a.lnk file, which finally resulted in the Go Stealer payload being released. This Go Stealer version, written in the programming language Go, adds sophisticated capabilities, such as a wider range of browsing focussed on and a cutting-edge technique for collecting information using the popular chat app Slack.
Superior Characteristics of the Go Stealer Version
Different from its GitHub equivalent, this Go Stealer version exhibits a higher degree of complexity. It creates a log file in the machine owned by the victim when it is executed and makes use of GoLang utilities like GoReSym for in-depth investigation. The malware focuses on cookies and usernames and passwords from web browsers, with a particular emphasis on Edge, Brave, and Google Chrome.
This kind is unique in that it is more sophisticated. Its deployment's cyber enemies have honed its strengths, increasing its potency and detection resistance. Using GoLang tools like GoReSym for comprehensive evaluation demonstrates the threat actors' careful planning and calculated technique.
Go Stealer: Evolution of Threat
The Go Stealer first appeared as a free software project on GitHub and quickly became well-known for its capacity to stealthily obtain private data from consumers who aren't paying attention. Its effectiveness and stealthy design rapidly attracted the attention of cyber attackers looking for a sophisticated tool for clandestine data exfiltration. It was written in the Go programming language.
Several cutting-edge characteristics distinguish the Go Stealer from other conventional data thieves. From the beginning, it showed a strong emphasis on browser focusing on, seeking to obtain passwords and login information from particular websites including Edge, Brave, and Google Chrome.The malware's initial iteration was nurtured on the GitHub database, which has the Go Stealer initial edition. Threat actors have improved and altered the code to serve their evil goals, even if the basic structure is freely accessible.
The Go Stealer version that has been discovered as the cause of the current internet spying by the Indian Air Force is not limited to its GitHub roots. It adds features that make it more dangerous, like a wider range of browsers that may be targeted and a brand-new way to exfiltrate data via Slack, a popular messaging app.
Secret Communications and Information Expulsion
This variation is distinguished by its deliberate usage of the Slack API for secret chats. Slack was chosen because it is widely used in company networks and allows harmful activity to blend in with normal business traffic. The purpose of the function "main_Vulpx" is specifically to upload compromised information to the attacker's Slack route, allowing for covert data theft and communication.
The Time and Strategic Objective
There are worries about targeted assaults or espionage activities due to the precise moment of the cyberattack, which coincides with the Indian government's declaration of its acquisition of Su-30 MKI fighter fighters. The deliberate emphasis on gathering cookies and login passwords from web browsers highlights the threat actor's goal of obtaining accurate and private data from Indian Air Force personnel.
Using Caution: Preventing Possible Cyber Espionage
- Alertness Against Misleading Techniques: Current events highlight the necessity of being on the lookout for files that appear harmless but actually have dangerous intent. The Su-30 Acquisition ZIP file is a stark illustration of how these kinds of data might be included in larger-scale cyberespionage campaigns.
- Potentially Wider Impact: Cybercriminals frequently plan coordinated operations to target not just individuals but potentially many users and government officials. Compromised files increase the likelihood of a serious cyber-attack by opening the door for larger attack vectors.
- Important Position in National Security: Recognize the crucial role people play in the backdrop of national security in the age of digitalisation. Organised assaults carry the risk of jeopardising vital systems and compromising private data.
- Establish Strict Download Guidelines: Implement a strict rule requiring file downloads to only come from reputable and confirmed providers. Be sceptical, particularly when you come across unusual files, and make sure the sender is legitimate before downloading any attachments.
- Literacy among Government Employees: Acknowledge that government employees are prime targets as they have possession of private data. Enable people by providing them with extensive cybersecurity training and awareness that will increase their cognition and fortitude.
Conclusion
Indian Air Force cyber surveillance attack highlights how sophisticated online dangers have become in the digital era. Threat actors' deliberate and focused approach is demonstrated by the deceptive usage of a ZIP archive that is camouflaged and paired with a sophisticated instance of the Go Stealer virus. An additional level of complication is introduced by integrating Slack for covert communication. Increased awareness, strict installation guidelines, and thorough cybersecurity education for government employees are necessary to reduce these threats. In the digital age, protecting national security necessitates ongoing adaptation as well as safeguards toward ever-more potent and cunning cyber threats.
References
- https://www.overtoperator.com/p/indianairforcemalwaretargetpotential
- https://cyberunfolded.in/blog/indian-air-force-targeted-in-sophisticated-cyber-attack-with-su-30-procurement-zip-file#go-stealer-a-closer-look-at-its-malicious-history
- https://thecyberexpress.com/cyberattack-on-the-indian-air-force/https://therecord.media/indian-air-force-infostealing-malware
Executive Summary
An image is being widely circulated on social media with the claim that it shows the destruction caused by an explosion at a mall in Japan following a 7.1-magnitude earthquake in July 2026. The viral image shows a shopping complex in a severely damaged condition, allegedly after the incident. CyberPeace Research Wing’s research found that the viral image does not show damage caused by an explosion or earthquake at any mall in Japan. The visuals in the image do not match authentic photographs of the affected shopping complex. The research further revealed that the viral image was created using Artificial Intelligence (AI).
Claim
A social media post claims that the AEON Mall in Japan exploded after the Kumamoto earthquake and a gas leak was suspected. The post, published in Burmese language on Facebook on July 29, 2026, shared an image claiming to show the aftermath of the incident at the mall.
FactCheck
To verify the authenticity of the viral claim, we conducted a Google search using relevant keywords. During the search, we found a report published by France24, which stated that the Kumamoto disaster management office released updated figures following the earthquake. According to the report, seven bodies were recovered from the debris, steel and wires of the Aeon shopping mall after a suspected gas blast. The report further stated that eight people were confirmed dead at the Nippon Paper Industries factory in Yatsushiro city, where part of a red-and-white smokestack collapsed. Following the 7.1-magnitude earthquake, five others were reported to be in critical condition, while around 9,500 residents were staying in evacuation centres in Kyushu, southwestern Japan.
http://france24.com/en/live-news/20260730-up-to-23-feared-dead-in-japan-quake
However, the viral image showed several visual inconsistencies and was flagged as AI-generated.
To further examine the authenticity of the image, we scanned it using the AI detection tool Hive Moderation. The results indicated that the image had an 84 percent probability of being AI-generated.
We also analysed the image using Human Meter, where the results showed a 75 percent probability that the image was AI-generated.
In the final stage of the research, we scanned the image using another AI detection tool, AI Image Detector. The tool indicated that the image had a 95 percent probability of being AI-generated.
Conclusion:
The research established that the viral image is not an authentic photograph of the damage caused by the Kumamoto earthquake or any explosion at an Aeon Mall in Japan. The image contains multiple inconsistencies and was found to be AI-generated. Therefore, the claim linking the image to the Japan earthquake incident is false.
A video is being shared on social media, falsely attributing it to Australian Prime Minister Anthony Albanese. The video claims that following the Bondi Beach attack, he decided to cancel the visas of Pakistani citizens.
An investigation by the Cyber Peace Foundation revealed that the viral video was created using AI. In the original video, Anthony Albanese was answering questions related to the Climate Change Bill during a press conference. It is important to note that in the attack that took place last Sunday (14 December) at Bondi Beach in Sydney, New South Wales, Australia, 15 people were killed. According to Australian police, the attack targeted the Jewish community. New South Wales Police Commissioner Mal Lanyon stated that the two accused involved in the attack were father and son—one aged 50 and the other 24. Media reports identified them as Sajid and Naved Akram.
Claim:
On 14 December 2025, a user on the social media platform X shared a video claiming, “After the attack by a Pakistani Islamic terrorist, the Australian Prime Minister has decided to cancel the visas of all Pakistanis. The whole world is troubled by this community, and in India it is said that Abdul cannot buy a house in a Hindu neighbourhood.”
The link to the related post, its archived version, and screenshots can be seen below:
Investigation:Upon closely examining the viral video, we suspected it to be AI-generated. Subsequently, we scanned the video using the AI detection tool aurigin.ai. According to the results provided by the tool, the video was found to be AI-generated.
