Who is actually in control, and who is taking the blame?
Curated by Business Science Daily — peer-reviewed sources, human-verified.
Learn more
About Our Curation Process
Business Science Daily curates academic research in business and management. Each featured study is selected from reputable, peer-reviewed journals, institutional repositories, or working papers (e.g., Elsevier, Sage, NBER, SSRN).
Articles are carefully summarized to ensure clarity and accuracy, with direct citations or links to original sources. Our process emphasizes transparency, academic integrity, and accessibility for a broader audience.
Learn more in our Editorial Standards & AI Policy.
Deploying AI systems increases speed, but accountability lags. The solution so far is to keep a human “in the loop”.
AI systems learn and change on their own, becoming so opaque that even their creators cannot fully understand them. This leaves a dangerous gap: accountability is assumed, but control is missing.
A new study in the Academy of Management Review argues that this gap is the real problem — not the technology itself. The solution is not more rules or warnings, but better alignment between who can influence the system, and who is responsible for what it does.
That means rethinking how power and responsibility are shared across the entire AI lifecycle. The study offers a roadmap based on six interconnected propositions.
Explore the tabs below to jump into the topics that matter most to you. Or keep reading further down the page.
AI systems learn from data produced before and during use — the first technology fundamentally reshaped by its own deployment. This creates a black-box problem and reduces control even for developers. Meanwhile, interdependencies between development and use blur the lines of accountability, creating a “many hands” problem.
Control is about influence, transparency, and predictability. Accountability is the obligation to explain and justify conduct to a forum that can pass judgment.
The paper argues: risks are mitigated if actors are held accountable only for what they can control. Misalignment — either form — increases the likelihood of undesired outcomes.
Users are in control if the system is explainable and leaves final decisions to them. Developers are accountable for providing that control.
Users have no control; they rely on the system. Developers have only partial control via intense testing and operating envelopes.
AI development and use span multiple actors: data providers, model developers, testers, user interface designers, end-users, and senior management. The entanglement of tasks makes it hard to locate accountability unambiguously.
Centralized governance (rules set by a focal firm) works for simple platforms. Decentralized governance — where all stakeholders interact based on jointly agreed rules — is better suited for complex, dynamic environments.
- Free-rider problems: overcome when actors realize that pure self-interest leads to suboptimal results for all.
- Exploitation: reduced when less powerful actors (e.g., individual users) are empowered.
- Example: Healthcare AI — physicians, data scientists, and hospital managers must jointly negotiate control and accountability.
Stakeholders have differing objectives and power. Distributive negotiation (fixed-pie) leads everyone to maximize control and minimize accountability. Integrative negotiation focuses on joint gains — giving up some control to others helps everyone meet their accountabilities.
Focuses on scientific knowledge, objectivity, quantification. Technology has deterministic effects on work.
Highlights entanglement of technology and social reality, emergent practices, value-oriented reasoning.
- Example: Physicians challenging AI diagnostic tools — they engaged only after questioning the logic based on their own expertise.
- Power management: Decentralized governance and a shared norm of accountability help reduce power differentials.
- Red teaming and sandboxes: Organizational mechanisms for continuous learning and feedback.
The theory speaks to three levels of analysis:
- Micro insight: Shifting blame from users to developers is not enough — developers also need control.
- Meso insight: AI is a testbed for distributed control and accountability across organizations.
- Macro insight: Shared governance among stakeholders can be effective, especially with accountability as a common norm.
Future research should examine the dynamics of control-accountability alignment across levels, with attention to different kinds of accountability and regulatory regimes.
- Three kinds of accountability: Visibility (explaining actions), Responsibility (duties and obligations), Liability (legal answerability). They may not always be assigned to the same actor.
- Transparency paradox: More information can make judgment harder.
- Regulators’ role: Different regulatory regimes (standards, binding law, norms) shape stakeholder interactions.
What makes the study compelling is not just its diagnostic power, but its prescriptive ambition. It doesn’t simply warn that AI is opaque and ungovernable. But it actually offers a roadmap for making it governable.
The path forward lies not in fighting AI’s autonomy, but in building social systems that can accommodate it.
Decentralized governance and integrative negotiation are not abstract ideals; they are practical responses to the very real problem of distributed accountability.
At the micro-level, the study reframes the working conditions of both AI users and developers. Research in this area has historically focused on how non-transparent algorithms stress and demotivate frontline workers. But what about the developers who are forced to build systems they cannot fully understand? Grote, G., Parker, S. K., & Crowston ask.
At the meso-level, the “many hands” problem — in which responsibility is so diffuse that no one is truly accountable — is addressed. The paper outlines how organizations can manage AI supply chains not just for data flow, but for accountability flow.
At the macro-level, researchers also ask, Can decentralized governance really work in an industry dominated by a handful of powerful platforms? The authors are optimistic.
They point to the recent commitment by sixteen Big Tech companies to publish safety frameworks for measuring AI risks as evidence that shared norms of accountability are possible.