A Chorus of Public Pledges
On the surface, the industry appears more aligned than ever. Tech giants like Google, OpenAI, and Anthropic have publicly supported calls for AI safety standards. We've seen a series of high-profile summits and government-led initiatives where executives
and world leaders share the stage, endorsing common goals and shared principles for responsible AI development. These voluntary commitments often involve promises to conduct safety testing, allow independent evaluations, and report serious incidents. The message is one of unity and proactive self-governance, designed to build public trust and show that the industry can regulate itself without stifling innovation. This wave of positive announcements creates the impression that a comprehensive agreement is just around the corner.
The Ambiguity of 'Safety'
The first major hurdle is definitional. A binding agreement requires crystal-clear terms, but words like “safety” and “best efforts” are notoriously vague and open to interpretation. What one company considers a robust safety test, another might see as a mere baseline. In a field as dynamic as AI, where new capabilities emerge constantly, codifying standards is like trying to map a coastline during a hurricane. An agreement that is too specific risks becoming obsolete overnight, while one that is too broad becomes unenforceable. This ambiguity makes it difficult to create a contract that all parties can agree on and that would hold up under legal scrutiny if a dispute were to arise.
Competition, Costs, and Legal Risks
Beneath the public calls for unity lies the fierce reality of market competition. A truly binding pact could force companies to slow down development or share proprietary information, ceding ground to rivals. The fear is that stringent regulation, whether self-imposed or governmental, could increase costs and constrain the innovative behaviour that gives a company its edge. Furthermore, signing a legally binding cooperation agreement creates significant legal exposure. If a company fails to meet the agreed-upon standards, it could face lawsuits not only from other signatories but potentially from the public. This has led to a strange situation where some firms face legal action for allegedly agreeing to slow down development, a move plaintiffs argue is anticompetitive. For many, a non-binding pledge offers the perfect middle ground: the positive PR of cooperation without the legal liability of a formal contract.
The Enforcement Conundrum
An agreement is only as strong as its enforcement mechanism. Who would police an industry-wide AI safety pact? A new, independent body, as some have proposed, would require funding, authority, and true neutrality—a tall order for a group potentially funded by the very companies it oversees. Without a credible enforcement arm, a “binding” agreement becomes little more than another voluntary pledge. This is a central tension in the debate between self-regulation and government oversight. Industry often prefers to self-regulate to avoid what it sees as slow and cumbersome government intervention, but critics argue that self-regulation often lacks teeth and can amount to little more than “ethics-washing.”
Why Handshakes Aren't Contracts
The current landscape is dominated by Memorandums of Understanding (MOUs) and other non-binding arrangements. These documents are useful for signaling intent and establishing a framework for discussion. They allow companies to align on high-level goals and demonstrate a commitment to cooperation in good faith. However, they are intentionally structured to avoid creating legal obligations or financial commitments. A cooperation agreement, by contrast, is a formal, legally binding contract that defines responsibilities, allocates risk, and is enforceable in court. The journey from a non-binding MOU to a binding cooperation agreement is a long one, filled with intense negotiation over every clause, definition, and liability limit.
