High reliability organizations

Dietterich (2018) is an interesting article about how to make robust AI. High risk situations require the combined AI and human system to operate as a high reliability organization (HRO). Only such an organization can have sufficiently strong safety and reliability properties to ensure that powerful AI systems will not amplify human mistakes.

Reliability and high-reliability organizations

The concept of high reliability organization (HRO) comes from Weick, Sutcliffe, and Obstfeld (1999). Examples of HROs include nuclear power plants, aircraft carriers, air traffic control systems, and space shuttles. They share several characteristics: an unforgiving environment, vast potential for error, and dramatic scales in the case of a failure.

The paper identifies five processes common to HROs, that they group into the concept of mindfulness (a kind of “enriched awareness”). Mindfulness is about allocating and conserving attention of the group. It includes both being consciously aware of the situation and acting on this understanding.

This mindfulness leads to the capacity to discover and manage unexpected events, which in turn leads to reliability.

Characteristics of a high reliability organization

An HRO is an organization with the following five attributes.

Preoccupation with failure

Failures in HROs are extremely rare. To make it easier to learn from them, the organization has to broaden the data set by expanding the definition of failure and studying all types of anomalies and near misses. Additionally, the analysis is much richer, and always considers the reliability of the entire system, even for localized failures.

HROS also study the absence of failure: why it didn’t fail, and the possibility that no flaws were identified because there wasn’t enough attention to potential flaws.

To further increase the number of data point to study, HROs encourage reporting all mistakes and anomalies by anyone. Contrary to most organizations, members are rewarded for reporting potential failures, even if their analysis is wrong or if they are responsible for them. This creates an atmosphere of “psychological safety” essential for transparency and honesty in anomaly reporting.

Reluctance to simplify interpretations

HROs avoid having a single interpretation for a given event. They encourage generating multiple, complex, contradicting interpretations for every phenomenon. These varied interpretations enlarge the number of concurrent precautions. Redundancy is implemented not only via duplication, but via skepticism of existing systems.

People are encouraged to have different views, different backgrounds, and are re-trained often. To resolve the contradictions and the oppositions of views, interpersonal and human skills are highly valued, possibly more than technical skills.

Sensitivity to operations

HROs rely a lot on “situational awareness”. They are ensuring that no emergent phenomena emerge in the system: all outputs should always be explained by the known inputs. Otherwise, there might be other forces at work that need to be identified and dealt with. A small group of people may be dedicated to this awareness at all times.

Commitments to resilience

HROs train people to be experts at combining all processes and events to improve their reactions and their improvisation skills. Everyone should be an expert at anticipating potential adverse events, and managing surprise. When events get outside normal operational boundaries, organizations members self-organize into small dedicated teams to improvise solutions to novel problems.

Underspecification of structures

There is no fixed reporting path, anyone can raise an alarm and halt operations. Everyone can take decisions related to their technical expertise. Information is spread directly through the organization, so that people with the right expertise are warned first. Power is delegated to operation personal, but management is completely available at all times.

HROs vs non-HROs

Non-HROs increasingly exhibit some properties of HROs. This may be due to the fact that highly competitive environments with short cycles create unforgiving conditions (high performance standards, low tolerance for errors). However, most everyday organizations do not put failure at the heart of their thinking.

Failures in non-HROs come from the same sources: cultural assumptions on the effectiveness or accuracy of previous precautions measures.

Preoccupation with failure also reveal the couplings and the complex interactions in the manipulated systems. This in turn leads to uncoupling and less emergent behaviour over time. People understand better long-term, complex interactions.

Reliability vs performance, and the importance of learning

An interesting discussion is around the (alleged) trade-off between reliability and performance. It is assumed that HROs put the focus on reliability at the cost of throughput. As a consequence, it may not make sense for ordinary organizations to put as much emphasis on safety and reliability, as the cost to the business may be prohibitive.

However, investments in safety can also be viewed as investments in learning. HROs view safety and reliability as a process of search and learning (constant search for anomalies, learning the interactions between the parts of a complex system, ensuring we can link outputs to known inputs). As such, investments in safety encourage collective knowledge production and dissemination.

Mindfulness also stimulates intrinsic motivation and perceptions of efficacy and control, which increase individual performance. (People who strongly believe they are in control of their own output are more motivated and more efficient.)

HROs may encourage mindfulness based on operational necessity in front of the catastrophic consequences of any failure, but non-HROs can adopt the same practice to boost efficiency and learning to gain competitive advantage.

Additional lessons that can be learned from HROs (implicit in the previous discussion):

  1. The expectation of surprise is an organizational resource because it promotes real-time attentiveness and discovery.
  2. Anomalous events should be treated as outcomes rather than accidents, to encourage search for sources and causes.
  3. Errors should be made as conspicuous as possible to undermine self-deception and concealment.
  4. Reliability requires diversity, duplication, overlap, and a varied response repertoire, whereas efficiency requires homogeneity, specialization, non-redundancy, and standardization.
  5. Interpersonal skills are just as important in HROs as are technical skills.


Dietterich, Thomas G. 2018. “Robust Artificial Intelligence and Robust Human Organizations.” CoRR. http://arxiv.org/abs/1811.10840.
Weick, Karl E., Kathleen M. Sutcliffe, and David Obstfeld. 1999. “Organizing for High Reliability: Processes of Collective Mindfulness.” In Research in Organizational Behavior, edited by R. S. Sutton and B. M. Staw, 21:81–123. Research in Organizational Behavior, Vol. 21. Stanford: Elsevier Science/JAI Press. https://archive.org/details/organizing-for-high-reliability.