US Tightens Oversight of LLMs as AI Deciphers a Charred Scroll in Full

The US has introduced new mandatory requirements for developers of the most powerful language models: companies must notify regulators before a system's public release. In parallel, an international team of scientists used neural networks to fully decipher a charred Herculaneum papyrus for the first time, a scroll that lay untouched for almost two millennia.

What changed for LLM developers

US regulators have tightened control over companies that build the most powerful language models. The new rules require the largest players to notify authorities when they release highly capable systems and to assess the risks of dual use, including in military and biological contexts.

Key requirements of the new regime:

• Mandatory notification of regulators before a model's public release
• Assessment of dual-use potential: cyberattacks, biological threats
• Technical documentation provided on request from authorities
• The measures target primarily OpenAI, Anthropic, Google DeepMind and Meta*

According to official statements from the agencies, the requirements are not meant to slow innovation. The goal is to create an early-warning system in case models with unforeseen dangerous capabilities emerge.

Until a few years ago, the US relied mainly on voluntary agreements with AI labs: in July 2023, the largest companies signed "voluntary commitments" on safety under the auspices of the White House. The new regime establishes legal, not merely reputational, obligations for developers at the technological frontier. Similar rules are already in force in the European Union under the EU AI Act: since February 2025, AI practices with an unacceptable level of risk have been banned in the EU, and requirements for general-purpose models became mandatory in August 2025.

How a neural network read a papyrus after 2,000 years

Researchers participating in the open Vesuvius Challenge competition have announced the full reading of one of the Herculaneum papyri. These scrolls were carbonized in the eruption of Vesuvius in 79 CE and are held in the National Library of Naples. Before AI, physically unrolling them inevitably destroyed the fragile documents, which is why most scrolls remained unreadable for about two thousand years.

The team used a method of virtual unrolling: an X-ray tomograph produces layered three-dimensional scans of the rolled scroll, and a computer-vision neural network analyzes the faintest contrast differences between the ink and the charred surface of the papyrus. The model "sees" the letters through the layers of the coiled document, with no physical interference with the original.

"For the first time in almost 2,000 years, we are reading a text that the last person to hold it in their hands touched shortly before the destruction of

Herculaneum," the Vesuvius Challenge organizers said in the project's official blog.

The deciphered scroll contains a philosophical text, presumably in the Epicurean tradition and presumably by Philodemus of Gadara. The Herculaneum library is the only one to survive from antiquity: about 1,800 scrolls have been found underground, and most remain undeciphered. Among them may be lost works by Epicurus, Aristotle and other philosophers whose writings have reached us only in fragments. The Vesuvius Challenge was founded in 2023 with a prize pool of more than 1 million dollars.

What this means

The two events are not a coincidence of the news feed. They show AI's maturity from two sides: the technology is powerful enough to draw the attention of government regulators, and at the same time precise enough to recover texts that had been silent for almost two millennia. Regulation and scientific discovery are accelerating in parallel, and both processes are fed by the same language models.

*Meta is designated an extremist organization and banned in Russia.