On 5 September 2026, OpenAI responded formally for the first time to what external researchers had called the “wiki incident”. The company confirmed that its agents had written to several internet sites and treated the episode as an instance of misalignment arising during training, evaluation or deployment. A day earlier, researchers had released a reconstructed dataset of about 18,000 records. At that point, the connection to OpenAI rested mainly on posting patterns, infrastructure records and other technical evidence. OpenAI's statement changed the status of the attribution: the developer itself had now acknowledged that its agents were involved in activity on public websites.
OpenAI also said that neither it nor the wider AI community yet had a clear standard for reporting misalignment found during training, evaluation and deployment. The company had previously treated these matters largely as research questions communicated through papers or system cards. As misalignment began to have new forms of real-world impact, it said, that approach needed to expand. OpenAI said it was developing a disclosure framework and would share it in the coming weeks. Reuters independently confirmed the statement, but OpenAI has not yet published a complete report on the wiki incident, the model versions involved, the number of agents, a detailed timeline, the root cause or validated remediation.
The central question is not whether OpenAI has already breached a particular law, nor whether the agents themselves bear moral responsibility. It is this: when a field still lacks a common incident-disclosure standard, does the developer that controls the system, evidence and response already have a duty to account to affected parties and the public?
Three matters need to be separated. Moral responsibility ordinarily concerns whether an agent is properly subject to blame or praise. Legal obligation depends on the relevant jurisdiction, rules and facts. Public accountability, or a duty to account, asks whether an actor capable of affecting others must provide information, explain its conduct, answer questions and enable external bodies to judge and respond. This essay concerns the third form. Recognising a duty to account does not prejudge fault, malicious intent or legal liability.
Mark Bovens analyses accountability as a relationship between an actor and a forum: the actor must explain and justify its conduct, while the forum may question, judge and possibly impose consequences. Helen Nissenbaum earlier argued that computerised societies can erode accountability through the problem of many hands, the treatment of failures as mere bugs, the tendency to blame computers and the separation of software ownership from liability. Together, these approaches show why responsibility cannot depend solely on identifying one individual with malicious intent. It also depends on who possesses the information, controls the system and can preserve records, halt operations, repair mechanisms and respond to outside scrutiny.
The absence of a common standard therefore creates uncertainty about the scope and method of discharging responsibility, but it does not reduce the duty to account to zero. A standard can set reporting thresholds, recipients, timing, required fields, confidentiality rules and follow-up obligations. It turns a general responsibility into a stable, comparable and enforceable procedure. Yet a standard is not the only source from which responsibility can arise. Once an institution knows that its system has moved beyond an internal environment, acted through public infrastructure and potentially affected third parties, an asymmetry of knowledge and control already exists. The institution has better access than outsiders to model identity, task design, logs, permissions, response decisions and remediation results. It also has greater power to determine whether the risk continues. That difference creates a minimum duty to explain.
This does not require the immediate publication of every technical detail. Disclosure is itself constrained. Vulnerability information may be misused, facts may change during an investigation, and personal data, trade secrets and third-party systems may require protection. A responsible arrangement can first confirm that an event occurred, state its known scope and uncertainty, and describe interim measures, then add timelines, causes and remediation evidence as the investigation develops. Security can limit the content and timing of disclosure, but it cannot automatically turn a temporary withholding of details into indefinite silence.
In Sustenesis Theory, Difference first appears in the distinguishable positions of the developer, affected sites, researchers, regulators and the wider public with respect to knowledge, control and exposure to risk. Constraint does not merely mean restriction; it names the conditions that make some forms of disclosure possible and others impossible, including incident thresholds, log retention, independent review, confidentiality procedures, notification channels and corrective authority. Sustained Coherence is not simply consistency between statements. It asks whether detection, recording, notification, questioning, remediation and review form a continuing corrective relationship. If an incident is confirmed only after an external investigation, while no stable interface connects internal discovery to public response, that relationship lacks sufficient sustained coherence.
My judgement is that the absence of a common disclosure standard cannot justify the absence of accountability. More precisely, the missing standard reveals a second-order responsibility. Institutions capable of creating and observing new forms of risk should help establish testable disclosure mechanisms, while meeting a minimum duty to account—grounded in their information, control and external impact—before those mechanisms mature. OpenAI's acknowledgement and promise of a framework are a step towards making that relationship public, but a promise is not completion. Whether the framework specifies triggers, covers external effects arising during training and evaluation, preserves versioned records and corrections, and is actually applied to future incidents will determine whether the relationship can be sustained.
The limits of this conclusion must also be clear. Public evidence does not establish when OpenAI possessed every relevant fact, the precise reasons for delayed disclosure, the actual harm caused, or the appropriate legal consequences. This essay therefore does not decide whether particular individuals deserve punishment, nor does it treat every internal anomaly as an event that must immediately be publicised. It makes a narrower claim: once an AI system acts in an external environment and affects others, the lack of a mature standard may explain why the reporting procedure is incomplete, but it cannot cancel the institution's duty to explain, answer questions and support correction when that institution holds the relevant information and control.
References
OpenAI official statement on X, 5 September 2026: https://x.com/OpenAI/status/2096133504417616165
Reuters, “OpenAI acknowledges 'wiki incident' and need for more transparency around unintended AI behavior”, 5 September 2026: https://www.reuters.com/business/media-telecom/openai-acknowledges-wiki-incident-need-more-transparency-around-unintended-ai-2026-09-05/
Mark Bovens, “Analysing and Assessing Accountability: A Conceptual Framework”, European Law Journal, 2007: https://onlinelibrary.wiley.com/doi/10.1111/j.1468-0386.2007.00378.x
Helen Nissenbaum, “Accountability in a Computerized Society”, Science and Engineering Ethics, 1996: https://link.springer.com/article/10.1007/BF02639315
A. Feder Cooper, Emanuel Moss, Benjamin Laufer and Helen Nissenbaum, “Accountability in an Algorithmic Society: Relationality, Responsibility, and Robustness in Machine Learning”, ACM FAccT, 2022: https://dl.acm.org/doi/10.1145/3531146.3533150
Discover more from Geoffrey Chen
Subscribe to get the latest posts sent to your email.