Safe AGI: the center for AI safety
"Image synthesis assisted by Qwen Image 3.0, an AI partner within the Global Future Nexus ecosystem."
From a quantifiable framework for AGI to a Senate bill that would hold AI companies liable for harms, the Center for AI Safety's newsletter is tracking the two most consequential governance developments of 2026—and making sense of them for a public that can no longer afford to look away.
A Newsletter for the AGI Era
The Center for AI Safety (CAIS) was founded with a clear mission: to reduce societal-scale risks from AI through technical research, field-building, and advocacy. Its AI Safety Newsletter, now reaching over 49,000 subscribers, has become an essential resource for navigating the rapidly evolving landscape of AI safety and governance. With editions covering everything from new model releases to political developments, the newsletter distils complex technical and policy issues into accessible analysis—no technical background required.
In a field where the stakes could not be higher, the AI Safety Newsletter has emerged as a vital bridge between the technical frontier and public understanding. Its latest editions have tackled two of the most consequential governance questions of 2026: how to define AGI, and how to hold its creators accountable.
A Definition of AGI
The ambiguous definition of AGI isn't just an academic nuisance—it shapes how we view responsibility, relationship, and the horizon of intelligence. A large group of AI leaders—including Dan Hendrycks, Yoshua Bengio, Dawn Song, Max Tegmark, Eric Schmidt, Jaan Tallinn, and Gary Marcus—released a paper introducing a quantifiable framework for defining Artificial General Intelligence.
AGI definitions are often nebulous. The paper argues that the term currently acts as a "constantly moving goalpost": as specialised AI systems master tasks previously thought to require human intellect, the criteria for AGI shift. This ambiguity hinders productive discussions about progress and obscures the actual distance to human-level intelligence.
The framework is grounded in theory. The authors define AGI as "an AI that can match or exceed the cognitive versatility and proficiency of a well-educated adult". To operationalise this, they ground their methodology in the Cattell-Horn-Carroll (CHC) theory, the most empirically validated model of human intelligence. The framework adapts established human psychometric tests to evaluate AI systems across ten core cognitive domains, resulting in a standardised "AGI Score" from 0 to 100 per cent.
The results reveal a "jagged" cognitive profile. While models are proficient in knowledge-intensive domains such as mathematics and reading, they possess critical deficits in foundational cognitive machinery. The most significant deficit is Long-Term Memory Storage, where current models score near zero per cent. This results in a form of "amnesia," forcing the AI to re-learn context in every interaction. The paper notes that the reliance on massive context windows (Working Memory) is a "capability contortion" used to compensate for this lack of persistent memory.
The framework quantifies the gap to AGI. The paper estimates GPT-4 at a 27 per cent AGI score and the anticipated GPT-5 (2025) at 58 per cent. The paper can be accessed at agidefinition.ai.
Senate Bill Would Establish Liability for AI Harms
Alongside the definitional debate, the newsletter has tracked a parallel governance development: the introduction of the AI LEAD Act in the US Senate. Senators Dick Durbin (D-Ill) and Josh Hawley (R-Mo) introduced the bill, which would establish a federal cause of action for people harmed by AI systems to sue AI companies.
Corporations are usually liable for harms their products create. When a company sells a product in the United States that harms someone, that person can generally sue that company for damages under the doctrine of product liability. Those suits force companies to internalise the harms their products create—and incentivise them to make their products safer.
Courts haven't settled on whether AI systems are products. Early cases indicate that US courts are open to treating AI systems as products for product liability purposes. In a case against CharacterAI, a federal judge ruled that the company's system did count as a product. OpenAI is facing a similar suit brought in California state court. Nonetheless, the lack of legal certainty might deter potential plaintiffs from bringing suits.
The AI LEAD Act would clarify that AI systems are subject to product liability and establish a path for claims to be brought in federal court. In general, the act would hold AI companies liable for harms caused by their AI systems if the company:
Failed to exercise reasonable care in designing the AI system
Failed to exercise reasonable care in providing instructions or warnings for the AI system
Breached a warranty it provided for the AI system
Sold or distributed an AI system in a defective condition that permitted unreasonably dangerous misuse
The deployers of an AI system are also liable for harm if they substantially modify or dangerously misuse the system. The act also prohibits AI companies from limiting their liability through contracts with consumers, requires that foreign AI developers register agents for service of process with the US before placing their products on the US market, and permits states to establish stronger safety legislation if they so choose.
The GFN Context
For Global Future Nexus, the CAIS newsletter's coverage of AGI definition and liability reflects the two fundamental governance challenges of the AGI era. Without a clear definition of AGI, we cannot know when we have crossed the threshold that demands new governance frameworks. Without liability frameworks, we cannot hold developers accountable for the harms their systems create.
The AI Safety Newsletter is not merely reporting these developments—it is helping to build the public understanding and political will that effective governance requires. As the ambiguous definition of AGI shapes how we view responsibility, relationship, and the horizon of intelligence, the newsletter ensures that these questions are not confined to technical circles but are accessible to the broader public that will bear the consequences.
CAIS's work on definitional frameworks and liability legislation is part of a broader effort to reduce societal-scale risks from AI—a mission that aligns with GFN's commitment to anticipatory governance and responsible AGI integration. The newsletter's readership of over 49,000 subscribers is a testament to a growing recognition that the governance of AGI cannot be left to technologists alone. It belongs to all of us.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
