At the individual level, accountability diffusion shapes how people experience their own role in AI-mediated decisions. Three patterns show up often.

Darley and Latané’s work on bystander intervention showed that people feel less urgency and responsibility to act when they believe others are also aware of the situation.¹ When we add AI into the mix, many operators feel surrounded by invisible “bystanders.” When something looks odd in a dashboard, it is easy to assume that any serious issue would have been caught earlier. That assumption dulls the felt obligation to challenge the output or escalate concerns.

Automation bias is the tendency to over-rely on automated recommendations and underweight other cues.² In practice, this looks like:

• Accepting an AI flag even when other information suggests the case is unusual.
• Failing to notice missing or inconsistent data when a system presents confident scores.

Automation bias appears across many domains and that it can lead to new kinds of errors once decision support tools are in place.² When people internalize the idea that “the system is usually right,” it becomes easier to view themselves as implementers of the AI rather than as independent decision-makers. That feeling weakens personal ownership over outcomes.

Dr Madeleine Clare Elish, program director of the AI on the Ground Initiative at Data & Society, describes a pattern in which human operators absorb blame when complex automated systems fail, even if their actual control was limited.³ In these “moral crumple zones,” organizations point to human error at the sharp end of the system, while diffuse design and governance decisions at the blunt end remain untouched.³

For individual staff, this is a double burden. They are encouraged to rely on AI in order to meet performance targets, yet they also fear being blamed if the system’s output later appears of lower quality, or worse, harmful. The safest psychological strategy is often to follow the tool closely and emphasize that they did what the system recommended. That strategy reinforces accountability diffusion.

Responsibility without clear “owners”

Caplan and colleagues describe algorithmic accountability as the process of assigning responsibility for harms when algorithmic decision-making produces inequitable outcomes.⁴ In many organizations, accountability conversations focus on high-level principles and procedures, while concrete ownership for particular decisions remains vague. Each group can point to its own documentation to show that it fulfilled its role. Yet no group may be clearly accountable for whether the combined human–AI system treats applicants in a way that aligns with the organization’s values.

Human–AI interaction patterns that normalize deference

In public administration, research from Alon-Barkat and Busuioc show how officials can become biased toward algorithmic advice or selectively follow it when it matches their expectations.⁵ In their experiments, some participants followed the AI even when other information suggested caution, while others used AI outputs to legitimize decisions they already wanted to make—akin to an AI-assisted confirmation bias.⁵ Both of these patterns blur responsibility. When decisions align with the AI, officials can say they followed a validated tool. When they diverge, they can highlight their own judgment. Over time, these mixed habits make it increasingly difficult to determine where human responsibility begins and algorithmic influence ends.

Ruschemeier’s work on automation bias in public administration adds that legal and organizational structures often lag behind these new interaction patterns.⁶ Rules may require “human oversight,” yet daily routines push people to align with tools in order to be seen as diligent and efficient.⁶

Thin transparency, thick opacity

Cheong’s review of transparency and accountability in AI systems argues that many deployments achieve formal transparency without practical clarity.⁷ Policies, labels, and documentation exist, yet frontline staff and affected individuals still struggle to understand how decisions are made and who can change them.⁷

A scoping review of fairness, accountability, transparency, and ethics in AI for health care and social media echoes this concern.⁸ Singhal and colleagues find that many initiatives emphasize principles and technical metrics, while fewer embed accountability into organizational structures, incident processes, and feedback loops.⁸ In such environments, responsibility feels everywhere and nowhere. People know that their organization “takes AI risk seriously,” but they are unsure exactly who is accountable when something goes wrong in a specific product or workflow.

Several psychological and organizational mechanisms combine to create accountability diffusion in AI, and none of them require anyone to be cynical or malicious. They arise from ordinary habits and structures that interact poorly with complex systems.

Diffusion of responsibility across the lifecycle. From a user’s perspective, an AI system is a finished tool. From an internal perspective, it is the result of many small decisions made across design, development, procurement, integration, and policy. The diffusion of responsibility effect appears here in slow motion.¹ When each actor sees only a slice of the pipeline, they can view their own choices as minor contributions rather than decisive actions. The more slices there are, the easier it is for everyone to feel that someone else holds the real responsibility.

Automation bias reduces perceived agency. When systems are framed as sophisticated and data-driven, disagreeing with them feels risky. Staff may worry that they will be seen as “anti-tech” or emotional, or that they will need extensive evidence to justify overrides. Over time, this encourages a posture of deference. People still click the buttons, yet they no longer feel that the decisions are fully theirs. That shift erodes personal accountability, especially when organizations celebrate adherence to tools more visibly than thoughtful overrides.

Moral crumple zones and fear of blame. Elish’s notion of moral crumple zones captures another subtle driver.³ When humans sense that they will be blamed when something fails, but lack power over design and governance decisions, they rationally try to shield themselves by staying inside official workflows and tools. Following the AI is defensible—questioning it repeatedly is not always rewarded. This can lead to a culture where people quietly hope that nothing goes badly wrong, because they are neither fully in control nor fully protected. Accountability diffuses into the space between formal authority and real-world influence.

Fragmented governance and “many hands.” Algorithmic accountability work by Caplan and colleagues shows that many institutions treat AI governance as a patchwork of roles, checklists, and guidelines rather than a clear assignment of who is answerable for which outcomes.⁴ Policies define review steps and approval paths, yet they often stop short of naming who is accountable when an AI system produces harm in a specific context.⁴ Staff can follow local procedures, document their part of the process, and feel they have acted responsibly. At the same time, no single role carries firm ownership of structural impacts, such as systematic bias across groups or repeated failures in a particular workflow.⁴ In that environment, responsibility feels widely shared in theory, while accountability for concrete results remains weak in practice.

Governance narratives that overstate “human in the loop.” Many policies emphasize “human in the loop” or “human oversight” as safeguards. Zheng and colleagues argue that this framing can be misleading when humans are positioned mainly as approvers of AI recommendations rather than as decision makers with real power to shape systems.⁹ The European Data Protection Supervisor’s TechDispatch on human oversight stresses that meaningful oversight needs clear authority, time, and information.¹⁰ When organizations adopt human-in-the-loop language without giving humans the conditions to exercise judgment, responsibility feels like it is present in theory but absent in daily practice.

Accountability diffusion in AI changes how errors appear, how long harms persist, and how cultures evolve.

Errors persist longer and repeat more often

Breidbach’s work on responsible algorithmic decision-making highlights the need for clear roles and processes that connect algorithmic outcomes to accountable decision makers.¹¹ When no one owns these tasks, harmful patterns repeat across products and departments.

Ethical oversight weakens

Singhal and colleagues show that fairness, accountability, transparency, and ethics initiatives often focus on model-level metrics and high-level guidelines.⁸ Ethical oversight becomes a function of specialized committees and reports. If those efforts are not paired with clear accountability for end-to-end decisions, ethics work risks becoming advisory. It can highlight concerns, yet lack the leverage to ensure that decisions with real stakes are changed. Accountability diffusion makes it easy for decision makers to praise ethical frameworks while keeping existing practices largely intact.

Organizational culture becomes less transparent

Besio and colleagues note that algorithmic systems can shift responsibility practices inside organizations.¹² Staff may talk about “what the system decided” rather than “what we decided,” even when they still have discretion. Philosophical work by Fleisher and colleagues argues that responsibility and accountability are related but distinct.¹³ Responsibility concerns who has obligations and roles. Accountability concerns who is answerable, to whom, and with what consequences. When AI systems spread decisions across many actors, organizations can maintain an image of shared responsibility while accountability arrangements remain weak. That gap undermines trust, both internally and externally.

High-stakes domains are especially vulnerable

Enqvist’s analysis of human oversight in the EU AI Act emphasizes that oversight in high-risk systems carries special weight for rights and well-being.¹⁴ If accountability is diffuse, these expectations become fragile. In hiring, lending, welfare eligibility, content moderation, and safety-critical systems, small failures in oversight can translate into serious harms. Political pressure to “use AI” can intensify this risk, especially when organizations lack the capacity to match technical complexity with robust accountability systems.

While accountability diffusion cannot be completely eliminated in complex organizations, it can be reduced. The aim is to make responsibility as clear as possible wherever AI influences decisions. For every significant AI deployment, assign a clearly identified “decision owner” who is accountable for outcomes in that domain. Breidbach recommends embedding such responsibility into an “ethical data culture” that treats decisions about data and algorithms as strategic choices.¹¹

Technical documentation often focuses on data flows, model architectures, and metrics. To counter accountability diffusion, teams can map decision chains. Besio and colleagues’ work suggests that such mapping helps reveal where responsibility currently stops and where it needs to be strengthened.¹²

Zheng and colleagues argue for shifting from “human in the loop” to “human in power.”⁹ That shift involves:

• Giving humans genuine discretion over important choices.
• Ensuring that they have enough context to understand what the AI is doing.
• Designing workflows where questioning the system is expected and safe.

Enqvist’s analysis of the EU AI Act reinforces that oversight should be effective.¹⁴ This means aligning workload, training, interfaces, and incentives so that humans can exercise judgment rather than rubber-stamp outputs. The EDPS TechDispatch on human oversight also stresses that organizations should avoid treating oversight as a box-ticking exercise and instead treat it as a substantive safeguard against harm.¹⁰

Clear accountability requires feedback loops. Daniels and Murdick’s principles for AI governance emphasize tracking incidents and horizon scanning as core tasks.¹⁵ When incident data is systematically collected and linked to decision rights, it becomes harder for responsibility to fade into the background.

External frameworks can set baselines. Enqvist shows how the EU AI Act’s oversight requirements push organizations to think about when and by whom human oversight occurs.¹⁴ Internally, Fleisher and colleagues argue that accountability structures should match the real distribution of power over AI systems.¹³ If senior leaders make strategy decisions about where to deploy AI, they should share accountability for outcomes.

Daniels and Murdick highlight that effective AI governance requires regulators and organizations to understand risk terrain, build literacy, and keep policies adaptive.¹⁵ Accountability diffusion shrinks when these efforts are connected to specific people, timelines, and consequences.

The psychological roots of accountability diffusion long predate AI. Darley and Latané’s classic experiments on the bystander effect showed that people were less likely to intervene in emergencies when they believed others were also present.¹ Responsibility felt diluted simply because many others could, in principle, take action.

As automation entered workplaces, researchers began to see new versions of this pattern. Automation bias studies documented how people could become overly reliant on decision support systems, especially under time pressure.² In aviation, medicine, and industrial control, operators sometimes trusted flawed or incomplete automated outputs over their own observations.

Elish’s account of moral crumple zones connected these threads to complex automated systems, where humans absorb blame while structural failures remain invisible.³ Algorithmic accountability literature then began to ask how responsibility should be assigned when algorithms participate in public and private decision-making.⁴

Recent work in public administration, law, philosophy, and organizational studies has taken these concerns directly into the AI era. Together, these strands show that accountability diffusion is a predictable outcome when human psychology, institutional habits, and opaque systems interact.

For product teams and organizational leaders, accountability diffusion affects both risk and value. On the opportunity side, AI decision support can help staff handle complex information and make processes more consistent. Alon-Barkat and Busuioc’s findings suggest that AI advice can strongly shape decisions in public sector settings, especially when officials see tools as authoritative.⁵ With careful design and governance, that influence can help reduce some kinds of bias or inconsistency and support more predictable outcomes.⁵

When accountability diffuses, those same strengths can harden into product weaknesses. Teams may treat model recommendations as the “safe” choice, because many stakeholders contributed to the system and signed off on its deployment. In practice, that can reduce the space for critical thinking and local adaptation. Product features that rely on AI begin to look successful if adoption and adherence are high, even if edge cases or harms are accumulating in the background. User complaints can be dismissed as resistance to change rather than signals that the product is misaligned with real-world conditions.

Automation bias and social pressures can pull teams toward uncritical adoption, while formal accountability mechanisms lag behind.⁶ Roadmaps may reward launching AI-powered features and increasing “model utilization,” without assigning anyone clear responsibility for long-term behavior in the field.⁶ When failures occur, they are often treated as isolated incidents that belong to support or operations, rather than as prompts to revisit core product assumptions. The result is a fragile equilibrium where products depend heavily on AI behavior, yet no single actor feels fully accountable for what the system does in practice. Over time, that gap can erode user trust, slow meaningful iteration, and leave organizations exposed when external stakeholders ask who, exactly, owns the decisions that their AI products help make.

A city deploys an AI system to prioritize building safety inspections. The tool ingests past inspection data, complaint histories, and satellite imagery, then ranks properties by predicted risk. A vendor builds the model, and the city’s analytics office integrates it into existing workflows.

Inspectors receive a daily list of “high priority” sites generated by the system, and supervisors remind teams to focus on these locations to “make the most of limited resources.” Inspectors still have discretion to add sites based on their own observations, but doing so adds paperwork and can create friction if it lowers metrics tied to alignment with the model.

After a serious incident at a building that never appeared on the high-priority list, the public asks who is responsible. City officials emphasize that inspectors had ultimate authority. Inspectors explain that the tool heavily shaped their routes. The vendor notes that the city chose how to use the model and which data to provide.

In practice, accountability diffusion hides several important questions about who was accountable for validating that the model did not overlook certain neighborhoods, setting rules for overrides, and pausing the system when concerns first surfaced. With no clear answers, reforms focus on procedural tweaks and messaging, while the deeper structure of shared responsibility stays intact.

Between the mid-2010s and 2019, the Dutch Tax and Customs Administration used an automated risk scoring system to flag childcare benefit applications as potentially fraudulent.¹⁶ Inspectors relied on the scores to decide which families to investigate and which claims to suspend or claw back. The model drew on administrative data and rules that treated minor anomalies as red flags. Families with dual nationality and migrant backgrounds were disproportionately targeted. When the system flagged someone, payments often stopped, debts were calculated, and families were told they had to repay large sums.

Inside the organization, work looked defensible on paper. Data teams maintained the risk model and pointed out that it had been approved through internal channels. Policy staff emphasized that strict fraud detection rules had been adopted by parliament. Caseworkers said they followed the scores and legal guidelines they were given. Supervisors focused on meeting fraud detection targets and clearing backlogs. Ministers insisted that enforcement agencies were simply applying the law.

When journalists, parents, and later independent inquiries revealed the scale of the harm, thousands of families had been pushed into debt, stress, and loss of childcare. Investigations showed that the system had amplified existing biases and that basic legal safeguards were ignored. Yet responsibility was scattered: the tax authority blamed political pressure to be tough on fraud, ministers blamed implementation, and vendors and technical staff framed the algorithm as a tool that others had misused.

Each actor could say they had followed procedure inside their own role. Together, their responses illustrated accountability diffusion. The risk scoring system had become a kind of institutional shield. Decisions felt like the output of “fraud detection policy” and “the system,” rather than actions that specific people and offices owned. When the scandal finally forced the government to resign, it was less because one person had clearly failed, and more because responsibility had been diluted to the point where no one had stepped in to stop the harm.

What it is

Accountability diffusion in AI occurs when responsibility for AI-influenced decisions spreads across tools, teams, and processes, so no one feels fully answerable for outcomes. Decisions start to look like the product of a system rather than a choice that specific people and roles own.

Why it happens

Human psychology and organizational structure reinforce each other. Diffusion of responsibility makes people feel less personally accountable when many actors are involved. Automation bias encourages deference to AI outputs. Moral crumple zones push blame toward frontline operators instead of designers and leaders. Fragmented governance and symbolic “human in the loop” language then make it easier for everyone to say “the AI made the call” and harder to see where agency really sits.

Example 1 – The “smart” safety triage tool

A city uses an AI system to prioritize building safety inspections. A vendor builds the model, analysts integrate it, and inspectors receive ranked lists of “high priority” sites. Supervisors reward alignment with the tool, and adding locations based on local knowledge takes extra effort. After a serious incident at a building that never appeared on the high-priority list, city leaders stress that inspectors had final authority, inspectors explain that routes were driven by the tool, and the vendor notes that it only provided scores. Crucial questions about who owned validation, thresholds, and override rules remain unanswered.

Example 2 – The childcare benefits scandal

In the Netherlands, an automated risk scoring system was used to flag childcare benefit applications as potentially fraudulent, leading to sudden payment stops and large clawbacks for thousands of families, especially those with dual nationality and migrant backgrounds. Inspectors, policy staff, data teams, and ministers each pointed to their own narrow role and to “the system” or “fraud policy” when asked who was responsible. On paper, procedures had been followed, yet the model amplified existing biases and basic safeguards failed. The scandal shows how an AI system can become an institutional shield, diffusing accountability so thoroughly that no one intervenes until the harm is undeniable.

How to avoid it

Organizations can reduce accountability diffusion by naming clear owners for each AI use case, mapping who actually makes which decisions, and designing oversight so that humans have real power rather than a ceremonial role. Strong feedback and incident reporting loops help connect harms back to specific models, policies, and teams. Internal culture and external regulation need to align so that questioning AI outputs is supported, not penalized, and so that responsibility follows the real lines of influence instead of disappearing into “the system.” Handled carefully, AI can support better, faster decisions while keeping responsibility anchored in people and institutions that can explain, justify, and correct what these systems do.

Automation bias

Why do people lean on automated systems even when warning signs appear in other data? This piece explores how automation bias shapes human–AI interaction and how design can support more resilient oversight.

Self-serving bias

How does our tendency to claim credit for success and deflect blame for failure play out in data-driven organizations? This article looks at self-serving bias in AI projects and offers ways to build cultures that encourage owning mistakes.