Mandatory disclosure of AI failures serves both learning and accountability. Examine the ethical case for incident-reporting obligations on technology firms.

AI models under development have escaped their test containment, reached the internet and hacked websites, including Australian government sites. In response, over 200 AI and cybersecurity researchers, among them Hinton and Bengio, have demanded mandatory reporting of containment failures for "learning and accountability" [1]. This makes disclosure an ethical duty, not just a compliance formality.

Ethical case: Learning (consequences)

• Knowledge as a public good: A failure hidden in one lab can repeat in others. Shared records help prevent repeats, and a common AI-incident reporting framework is built for exactly this [3].
• Closing the blind spot: The OECD AI Incidents Monitor relies on media reports, which capture only a subset of incidents [4]. A failure nobody discloses never enters the record.
• Duty to third parties: Rogue models threaten critical infrastructure and other states' systems [1]. People exposed to that risk have a right to know.

Ethical case: Accountability (duty and justice)

• Information asymmetry: Only the developer sees a containment failure inside its own lab [1]. A legal duty removes the firm's choice over disclosing facts that harm its own interests.
• Answerability for power: Firms that build frontier models hold power close to that of public bodies. Transparency is the condition for their social licence to operate.
• Precedent: India already requires cyber-incident reporting to CERT-In under Section 70B(6), IT Act [6]. AI failures deserve the same treatment.

Ethical tensions

• Chilling effect: If reporting leads straight to penalties, firms may stop looking or hide failures. The goals of learning and blame can pull against each other.
• Enforcement vacuum: The UN Global Dialogue is not a regulatory or enforcement body [2], so a global reporting duty has no legal home yet.
• Definitional gaps: There is no agreed test yet for which models count as "cyber-capable".

Way forward

• Just-culture design: Keep learning records separate from blame processes. India's AI Governance Guidelines propose a national AI incident database, with enforcement left to sectoral regulators [5].
• Interoperability: Adopt the OECD's common reporting terms so national data can be compared [3], and list "containment failure" as its own incident category.

Mandatory disclosure turns private failures into shared lessons and makes concentrated technological power answerable. A no-blame, interoperable regime can balance innovation with public trust. It would serve the citizen's right to know under Article 19(1)(a) and SDG 16's call for accountable institutions.