cross-posted from: https://sh.itjust.works/post/48743094

The requirements for governance, model transparency/notification, began in August 2025. General applicability of most of the AI Act including obligations for “high-risk” AI systems begin in August 2026. Obligations for high-risk AI systems that are part of safety components in regulated products become applicable in August 2027.

By introducing a tiered, risk-based regime, the law bans certain AI practices outright, layers strict obligations on high-risk systems, and establishes unprecedented oversight for general-purpose AI and foundation models, especially those with systemic impact. Its reach is global—any company placing AI into the EU market must comply and its penalties are steep.

For businesses and their counsel, compliance isn’t optional. It’s now the price of entry into one of the world’s largest markets, and it will influence how innovation is documented, safeguarded, and patented.

It’s important to understand the AI Act’s most consequential provisions for patent strategy, particularly how regulatory documentation duties intersect with data provenance, inventorship disputes, trade secret versus patent trade-offs, claim drafting under new compliance constraints, and geo-strategic filing decisions.

The AI Act distinguishes between four key categories of AI systems:

Minimal-risk systems (AI-enabled video games, spam filters)

Limited-risk systems (chatbots interacting with users)

High-risk systems (AI in health care, education, employment, infrastructure, biometric identification)

General purpose and foundation models

While minimal and limited risk systems carry only modest obligations, the regulation of high-risk and general-purpose/foundation models forms the core of the AI Act and presents the most significant patent implications.

Before entering the market, high-risk systems must clear a conformity assessment and provide detailed technical documentation that includes:

System description and intended purpose

System architecture, algorithms, and datasets used

Risk management, testing, and validation protocols Transparency and human oversight measures

Properly managed, these records can help distinguish human inventive contributions from AI-assisted outputs, providing valuable support for defending inventorship and reinforcing patent validity in a contested landscape.

Documentation and dataset summaries may further reduce the viability of trade secret protection. Both the EU Trade Secrets Directive and US law require that information be kept confidential through reasonable protective measures.

By mandating disclosures, the AI Act can erode this confidentiality, making secrecy harder to sustain in practice. This shift alters the calculus for IP strategy. In some cases, patenting may become the safer path, as certain technologies no longer can be reliably safeguarded as trade secrets.

The EU AI Act prohibits certain uses of artificial intelligence (AI). These include AI systems that manipulate people's decisions or exploit their vulnerabilities, systems that evaluate or classify people based on their social behavior or personal traits, and systems that predict a person's risk of committing a crime. The Act also bans AI systems that scrape facial images from the internet or CCTV footage, infer emotions in the workplace or educational institutions, and categorize people based on their biometric data. However, some exceptions are made for law enforcement purposes, such as searching for missing persons or preventing terrorist attacks.

This paper provides an overview of the main sources of catastrophic AI risks, organized into four categories: Malicious Use; AI Race; Organizational Risks; and Rogue AIs. (PDF can be downloaded from the linked arxiv.org page.)

Debating the proposition "AI research and development poses an existential threat"! Witness incredible feats of mental gymnastics and denialism! Gaze in dumbstruck awe as Yann LeCun suggests there is no need to worry because if and when AI starts to look dangerous we simply won't build it! Feel your jaw hit the floor as Melanie Mitchell argues that of course ASI is not an X-risk, because if such a thing could exist, it would certainly be smart enough to know not to do something we don't want it to do! A splendid time is guaranteed for all.

Pretty solid interview with Connor Leahy about why AI Safety/Alignment is important, how transhumanists are caught in an ideological race, and some ideas for possible societal solutions/regulations.

I think this might be Leahy's best yet.

This video by Robert Miles makes a levelheaded argument for taking existential AI Safety seriously as rapidly as possible.