The approval takes three seconds
A manager receives a recommendation generated from a system they did not build, trained on data they have never seen, configured by a team they do not know, purchased through a process they did not lead.
The recommendation is concise.
The confidence score is high.
The queue is long.
The interface asks for approval.
Three seconds later, the decision moves forward under the manager's name.
Formally, a human decided.
Morally and practically, the picture is less clear.
Did the manager understand the basis of the recommendation?
Could they have delayed it?
Would disagreement have required time the workflow does not provide?
Did they have authority to challenge the objective the system was optimizing?
Who is responsible if the outcome causes harm?
Responsibility under acceleration begins where action becomes faster than the human and institutional processes through which answerability normally forms.
Definition
Responsibility is not one thing.
It can refer to causal contribution, moral blame, legal liability, professional duty, public accountability, active responsibility for preventing harm, and the obligation to repair after harm occurs.
AI ethics has made this plurality visible through debates about responsibility gaps.
Filippo Santoni de Sio and Giulio Mecacci argue that “the responsibility gap” is better understood as several different problems, including gaps in culpability, moral and public accountability, and active responsibility. These gaps can arise from technical opacity, distributed agency, organizational structure, law, incentives, and social conditions, not only from machine learning.
Responsibility under acceleration is the problem of preserving these forms of answerability when technological systems increase the speed, scale, opacity, and distribution of action.
The phrase is useful because time changes responsibility.
The faster a system can act, the easier it becomes for review to become ceremonial.
Accountability is not the same as blame
Organizations often notice responsibility only after failure.
Who approved this?
Who signed the document?
Who was on duty?
Blame matters in some cases, but a responsibility architecture must exist before harm.
Who is supposed to notice risk?
Who has authority to stop deployment?
Who is responsible for model monitoring?
Who must investigate contradictory evidence?
Who represents affected people?
Who can change the objective?
Who owns repair?
Active responsibility is forward-looking.
It asks how institutions organize care before consequence arrives.
Acceleration makes this more important because a system can scale an error before a traditional review process catches up.
The responsibility gap is often organizational
AI can create genuine technical difficulties.
Machine-learning systems can be opaque.
Behavior may change after deployment.
Data pipelines may be difficult to trace.
But responsibility can also disappear through ordinary bureaucracy.
The vendor says the client chose the use case.
The client says the vendor built the model.
The operator says they followed procedure.
The manager says the system was only advisory.
The procurement team says it assessed compliance.
The legal team says the risk fell within policy.
Every actor has a partial explanation.
No one holds the whole consequence.
This problem existed before AI.
AI makes the chain longer, faster, and easier to describe as technical.
Meaningful human control
The concept of meaningful human control was developed partly in debates about autonomous systems.
Santoni de Sio and Jeroen van den Hoven argue that human control should be connected to relevant moral reasons and to the ability of responsible human agents to track and respond to them.
This is more demanding than putting a human in a loop.
A person cannot exercise meaningful control if they lack information.
They cannot exercise it if they have no time.
They cannot exercise it if the system is too complex to challenge and no support exists.
They cannot exercise it if institutional incentives punish refusal.
They cannot exercise it if the action has already become effectively irreversible by the time they are asked to review it.
Responsibility requires practical control.
Acceleration compresses moral attention
The moral problem of speed is not only that people make faster decisions.
It is that fewer things can become salient before action.
Consequences remain abstract.
Affected people are represented as rows.
Uncertainty is summarized into a score.
The next case arrives before the previous one has been metabolized.
The system can make a morally complex action feel operationally ordinary.
This is where responsibility intersects with judgment compression.
A person may know the principles and still have no usable interval in which to apply them.
The Institute therefore treats deliberative time as part of responsibility architecture.
Some actions should be fast.
Others need protected latency because the cost of premature closure is too high.
Responsibility cannot be outsourced to explanation
Explainable AI is important.
But explanation alone does not solve responsibility.
A person may receive an excellent explanation and still lack authority to act differently.
A model may be transparent while the organizational goal remains unjust.
An institution may document every step while affected people have no meaningful appeal.
Responsibility depends on power as well as information.
Who can say no?
Who can revise the model?
Who can change the workflow?
Who can compensate the harmed person?
Who can refuse the optimization target?
This is why governance cannot be reduced to interpretability.
The human burden problem
There is also a danger in solving responsibility gaps by assigning more responsibility to frontline humans.
A clinician is told they are ultimately responsible for an AI-assisted decision.
A content moderator is told they must catch the system's failures.
A public servant must review automated recommendations.
But the workload, training, authority, and system design remain unchanged.
The institution has not preserved human control.
It has transferred liability downward.
Responsibility without capacity is not meaningful oversight.
It is exposure.
A responsible institution must match answerability with resources, authority, information, and time.
Collective responsibility
Complex systems require shared responsibility.
Shared does not mean vague.
Different actors can hold different obligations.
Designers can be responsible for foreseeable system properties.
Deployers can be responsible for context and use.
Leaders can be responsible for incentives and escalation paths.
Operators can be responsible for professional judgment within realistic limits.
Regulators can be responsible for enforceable boundaries.
Institutions can be responsible for monitoring, repair, and redress.
This layered model is more realistic than searching for one person who “owns” an AI system.
It is also harder.
Good governance makes the layers legible.
Responsibility and agency
Responsibility presupposes some form of agency.
If a person cannot understand, influence, or refuse an action, the moral meaning of holding them responsible changes.
This is why the Institute links responsibility under acceleration to agency under technological acceleration.
The relationship is reciprocal.
Agency without responsibility becomes mere choice power.
Responsibility without agency becomes burden.
A mature human-AI system needs both.
The European governance context
European AI governance places significant emphasis on human oversight, accountability, fundamental rights, and risk management.
The EU AI Act creates formal obligations that vary by system category and role. Earlier European trustworthy-AI guidance similarly emphasized human agency and oversight.
These frameworks matter because they translate ethical concerns into institutional duties.
But compliance should not be confused with complete responsibility.
A system can satisfy formal requirements while still producing local practices in which people feel unable to challenge it.
Law establishes floors and procedures.
Responsibility remains a lived organizational capacity.
Time is part of moral architecture
Responsibility is usually described in terms of who knew, who chose, who controlled, and who can be held answerable afterward. Under acceleration, another variable becomes visible: how much time existed for those capacities to function before action became difficult to reverse.
A five-minute review and a five-day review are not the same governance arrangement simply because the same person signs both decisions. A human reviewer who receives hundreds of machine recommendations per hour may possess formal authority while lacking practical opportunity to exercise it. An appeal that arrives after an automated consequence has already affected housing, income, reputation, health, or liberty may technically exist while failing to preserve meaningful agency.
This is why latency is not always waste. Sometimes latency is where responsibility lives.
The required interval will differ by stakes. A spelling correction needs almost none. A medical escalation, employment decision, welfare determination, disciplinary action, or high-impact public-sector classification may require time to inspect uncertainty, hear contextual information, consult another person, or stop the process entirely.
The important question is not whether human review is present in a flowchart. It is whether the temporal design makes responsible interruption possible before consequence hardens.
Reversibility changes the burden of speed
Acceleration is less dangerous when mistakes are cheap to detect, easy to reverse, and unlikely to compound. It is more dangerous when action creates lock-in.
A generated draft can be discarded. A recommendation stored in a personnel file may follow someone for years. A ranking can alter who receives further scrutiny. A risk label can change how every later observer reads the same person. An automated denial can create deadlines, debt, or administrative burdens that remain even after the original decision is corrected.
Responsibility under acceleration therefore includes reversibility: the ability to pause, challenge, restore, and repair.
This also changes what accountability means after error. A mature system does not ask only who should be blamed. It asks how quickly the error can be surfaced, whether affected people can contest it without specialist knowledge, whether records can be corrected downstream, and whether the organization learns enough to change the conditions that produced the error.
Repair is not a softer alternative to accountability. It is one of the ways accountability becomes materially real.
How should responsibility under acceleration be evaluated?
There is no single scale.
Evaluation can map decision rights, causal chains, data provenance, review time, override authority, escalation paths, auditability, contestability, appeal, incident response, and repair.
Researchers can study whether humans understand the systems they oversee, whether they detect errors, whether they feel able to dissent, and whether institutional incentives align with formal responsibility.
Incident analysis can reveal where responsibility became diffuse.
Ethnographic work can reveal where formal control differs from practical control.
The key is to measure responsibility as architecture, not only attitude.
What the evidence does not show
The literature does not show that AI creates wholly unprecedented responsibility problems.
Organizations have long diffused responsibility through bureaucracy and complex systems.
It does not show that human control is always preferable to automation.
Humans can be biased, fatigued, corrupt, inattentive, or less accurate than automated systems.
It does not show that explainability alone solves responsibility.
It does not show that legal liability and moral responsibility always align.
And philosophical arguments about responsibility gaps remain contested. Some scholars argue that the problem is better described as a control gap or as a failure to assign indirect human responsibility correctly.
That disagreement belongs in the page.
Responsibility also needs an exit route
A person cannot meaningfully carry responsibility for a system they are forbidden to leave.
This is easy to miss in organizations where automated workflows become default infrastructure. Once the model is embedded in procurement, staffing, targets, and reporting, an employee may technically retain discretion while every practical path points toward compliance. Refusal becomes possible only at personal cost.
Meaningful responsibility therefore requires an exit route: a way to escalate, suspend, revert, or route around the system when the conditions for responsible use are not present. The route has to exist before the exceptional case arrives.
Without it, the institution may call the human accountable while designing away the human's ability to act differently.
The pause that has to be real
The manager's hand is still above the approval button.
Responsibility becomes real only if the pause can change the outcome.
If the person can inspect the evidence.
If they can ask why.
If they can refuse.
If the institution will support that refusal when it is justified.
If the chain behind the recommendation is visible enough for repair.
If those affected have somewhere to appeal.
In an accelerated system, responsibility cannot depend on good intentions surviving bad architecture.
It has to be designed into the conditions of action.
Research references
- Santoni de Sio, F., & Mecacci, G. (2021). Four Responsibility Gaps with Artificial Intelligence: Why they Matter and How to Address them. Philosophy & Technology, 34, 1057–1084. https://doi.org/10.1007/s13347-021-00450-x
- Santoni de Sio, F., & van den Hoven, J. (2018). Meaningful Human Control over Autonomous Systems: A Philosophical Account. Frontiers in Robotics and AI, 5, 15. https://doi.org/10.3389/frobt.2018.00015
- Veluwenkamp, H., & Hindriks, F. (2024). Artificial agents: responsibility & control gaps. Inquiry. https://doi.org/10.1080/0020174X.2024.2410995
- Parasuraman, R., Sheridan, T. B., & Wickens, C. D. (2000). A model for types and levels of human interaction with automation. https://doi.org/10.1109/3468.844354
- European Commission. Ethics Guidelines for Trustworthy AI. https://knowledge4policy.ec.europa.eu/publication/ethics-guidelines-trustworthy-ai_en
- Regulation (EU) 2024/1689. https://ai-act-service-desk.ec.europa.eu/en/ai-act/article-1