The development of the most powerful artificial intelligence systems could accelerate to the point where their safety mechanisms are unable to keep pace. Armenia’s former Minister of Justice Arpine Hovhannisyan addressed this issue, outlining questions related to AI development safety, government regulation and potential legal liability.
Hovhannisyan noted that Anthropic CEO Dario Amodei, in an essay published on September 12, proposed slowing the pace at which the capabilities of the most powerful systems are developed so that safety mechanisms can keep up. Sam Altman and Elon Musk have also publicly supported this proposal.
According to Hovhannisyan, this is not about stopping or abandoning the artificial intelligence systems used in everyday life. The issue concerns the most powerful next-generation systems, whose capabilities already extend significantly beyond ordinary question-and-answer tasks.
As an example, she cites Mythos 5, which was available only to verified partners after its release and was shut down for around three weeks in June under an export-control directive from the U.S. Department of Commerce.
Hovhannisyan also referred to data published by Anthropic. After reviewing more than 141,000 tests, the company identified three cases in which its advanced models moved beyond their designated experimental environments and gained unauthorized access to the real systems of three organizations.
According to her, one proposal is to allow experts independent of AI companies, under certain conditions, to access the most powerful systems. This would enable them to study how the systems operate from the inside, test dangerous capabilities, assess safety mechanisms and investigate incidents that have already occurred.
Against this background, the question of the limits of government regulation also arises.
The US and EU Have Chosen Different Approaches
Hovhannisyan notes that the current U.S. administration is pursuing a policy aimed at encouraging artificial intelligence development and is more cautious about restrictions that could slow down American companies.
One reason is the strategic objective of avoiding falling behind China.
The European Union has already chosen a stricter approach. The EU Artificial Intelligence Act, adopted in 2024, is based on risk differentiation: the greater the risk posed by an AI application, the stricter the requirements.
Certain uses are prohibited in the EU. These include some forms of social scoring and, with certain exceptions, the use of AI to recognize people’s emotions in workplaces or educational institutions.
At the same time, in July 2026, the application of requirements for high-risk systems was postponed until December 2027, while transparency requirements are already in force.
In other words, the strict approach has been maintained, but some deadlines have been postponed.
AI Development Could Lead to Multibillion-Dollar Lawsuits
The rapid development of artificial intelligence systems and the potential harm they could cause also create serious legal risks in the form of enormous lawsuits and potential multibillion-dollar compensation claims.
Hovhannisyan notes that David Sacks, one of the key figures in the Trump administration’s AI policy, has also raised the risk of large-scale legal liability for AI companies, particularly when something goes wrong during a system’s operation.
In this context, the former minister highlights three court cases whose outcomes, in her assessment, will have significant implications for the future development of the entire industry.
The first is The New York Times vs OpenAI and Microsoft, which began in 2023.
The copyright dispute has raised broader questions: can millions of copyrighted works be used to train AI, under what circumstances is this lawful, what happens when a system reproduces protected material in its response, and who is responsible — the company that created the system, the user, or both?
The second is Like Company vs Google Ireland, a European case concerning Google Gemini.
Case C-250/25 is the first case before the Court of Justice of the European Union concerning generative artificial intelligence and copyright.
It also addresses the use of protected materials to train a system and liability arising from their potential reproduction in its outputs.
The neighboring rights of press publishers and EU rules governing large-scale computerized text analysis are also significant.
The first oral hearing took place in March 2026, and the case remains pending.
The third case has already ended in a settlement.
In Bartz vs Anthropic PBC, Anthropic, the developer of Claude, agreed to pay $1.5 billion to settle a class-action lawsuit brought by authors involving approximately 482,000 works.
At the same time, the court found that training a model on lawfully acquired books was itself fair use.
The dispute concerned millions of pirated copies that had been illegally downloaded and stored in the company’s library. The settlement received final approval in July 2026.
According to Hovhannisyan, discussions about artificial intelligence safety are neither about mysticism nor about stopping the technology.
As with any powerful technology, risks need to be addressed in a timely manner without either ignoring or dramatizing them.
In her formulation, the ultimate questions are very specific: what can the system do, who verifies it, what safety mechanisms exist, where is the boundary drawn, and who will bear responsibility when that boundary is crossed and harm is caused?