As AI systems gain the ability to act, organisations need to show how people can question their decisions, prevent harm and get mistakes corrected.
A person can approve an AI-assisted decision without having the time, evidence or authority to question it.
Consider a hiring process. An AI tool helps shortlist applicants. A rejected candidate asks for a review. The recruiter opens the same score that informed the rejection and confirms the result.
There has been human involvement. Whether there has been meaningful scrutiny depends on what the recruiter could actually examine and change.
That distinction deserves more attention in AI adoption. Fairness, transparency and accountability are easy to endorse. They become harder questions when someone must decide what information a reviewer receives, how much time they have and whether they can stop a decision from taking effect.
My view is that an organisation’s commitment to responsible AI should be judged partly by what happens when someone has good reason to disagree with it.
When AI can act, oversight needs to happen earlier
In February 2026, the US National Institute of Standards and Technology announced its AI Agent Standards Initiative, addressing systems capable of autonomous action. Its priorities include research on agent security and identity.
That development gives the ethics discussion a practical urgency. A tool that drafts a response leaves an opportunity for someone to check it. Connect the tool to a customer database or an external messaging system, and the question becomes what it is authorised to do before anyone intervenes.
A later correction cannot fully undo the disclosure of private information.
For project leaders, the boundary between recommendation and action should therefore be an explicit design decision. Which actions require approval? What can the system access? Who can pause it? What record remains if an action is disputed?
These decisions belong in the project scope and operating model. Leaving them until after launch makes effective oversight harder to establish.
Accuracy starts with choosing the right target
One of the most useful studies in AI ethics predates the current generation of chatbots.
In a 2019 paper published in Science, Ziad Obermeyer and his colleagues examined a commercial algorithm used to identify patients for additional care. At the same risk score, Black patients were considerably sicker than White patients.
The algorithm predicted healthcare spending as a substitute for medical need. Unequal access to care meant that spending understated the needs of some patients.
The researchers estimated that correcting the disparity would raise Black patients’ share of those automatically identified for additional help from 17.7% to 46.5%.
The problem lay in what the system had been asked to predict. A target that looked reasonable produced an unfair basis for allocating help.
A similar question belongs in any AI procurement discussion: does the performance measure reflect the purpose of the service?
For example, a customer-service system could be rewarded for closing complaints quickly, even when the underlying issue remains unresolved. A hiring model might reproduce patterns in past recruitment without establishing whether those patterns identify people who can do the work.
A convincing demonstration of predictive accuracy will not answer those questions. The organisation must first decide what a good outcome means and whose experience it is measuring.
Human review is work that has to be resourced
Research also challenges the assumption that combining people and AI automatically produces the best result.
A 2024 meta-analysis in Nature Human Behaviour, by Michelle Vaccaro, Abdullah Almaatouq and Thomas Malone, examined 106 experiments reported across 74 papers. On average, human–AI combinations outperformed humans alone, but fell below whichever performed best independently: the human or the AI.
The underlying studies were published between January 2020 and June 2023, with substantial variation across tasks. The finding is a reason to test the combined process rather than assume that human involvement improves it.
For an employer, that means looking closely at the reviewer’s job. Can they examine the original evidence? Do they understand the system’s limitations? Can they request another opinion or override a recommendation?
Time matters too. An employee expected to process a queue quickly may have little opportunity to investigate an unusual case. Giving that person responsibility without changing their workload creates a predictable conflict.
If a process depends on human judgement, that judgement has to be budgeted for.
Interface design can help or hinder it. In a 2021 experiment involving 199 participants, Zana Buçinca and her colleagues used a meal-modification task with simulated AI. Designs that prompted more deliberate thinking reduced overreliance compared with simpler explainable-AI interfaces. However, participants gave the designs that reduced overreliance most the least favourable subjective ratings.
A smoother experience, then, is not sufficient evidence of a better decision process. Teams should examine whether users detect important errors, as well as whether they like the interface.
The appropriate amount of checking should reflect the consequences of a mistake. Reviewing a suggested meeting title and reviewing a decision affecting someone’s access to a service are different responsibilities.
Putting Qatar’s accountability principle into practice
Qatar’s published ethical AI guidance provides a useful foundation for this discussion. The development and deployment guidelines available through the government’s Sharek platform, on a page dated May 2024, include a clear principle: “Assign ultimate accountability to humans.”
“Assign ultimate accountability to humans”
The guidance also addresses communication with affected users, human intervention and the ability to appeal or override decisions with substantial impact.
For organisations adopting AI today, the practical question is how to turn that principle into project requirements, staff responsibilities and a functioning route to correction.
A supplier should be able to explain the system’s capabilities and limitations. The adopting organisation must decide where it will use the tool, what influence its output will carry and who will investigate a disputed outcome.
Those responsibilities cross technical and operational boundaries. A model developer may be able to explain a score but lack authority to change a service decision. A manager may have that authority but lack access to the evidence. Effective oversight requires the two to work together.
The NIST AI Risk Management Framework offers a useful operational reference, covering oversight roles, feedback and appeals, monitoring and arrangements to disengage systems.
I would make a demonstration of the review process part of acceptance testing. A project should show how a mistake will be investigated and corrected alongside what the AI can do.
Language can determine whether someone is heard
For services used in Qatar and across MENA, local evaluation also needs to examine how people describe a problem.
AraDiCE, published at COLING 2025 by researchers including a team at Qatar Computing Research Institute, benchmarks Arabic dialectal and cultural capabilities. Its evaluation includes Gulf, Egyptian and Levantine contexts and identifies difficulties in dialect identification, generation and translation among the models tested.
That research supports a practical requirement: test the language used by the people the service is intended to serve.
Suppose a person explains a complaint in dialectal Arabic. The system summarises it, assigns a category and passes it to a reviewer. If the summary loses an important qualification, the reviewer may respond correctly to an incomplete account.
Fluency in the final answer would conceal the original misunderstanding.
Testing should therefore follow meaning through the whole process. Reviewers should be able to consult the original message, and users should have a way to clarify information the system misunderstood.
The same attention belongs in an appeal route. Clear instructions, usable language and a reasonable way to provide evidence affect whether a person can challenge an outcome at all.
Test a disagreement before launch
Research on contestable AI treats the ability to challenge a system as a design concern throughout its life. Kars Alfrink and his colleagues explain why human review can be inadequate when reviewers lack the information or authority to change decisions. They also warn that appeals can shift the burden of finding problems onto affected people.
An organisation can begin addressing this with a controlled test case.
Introduce a known error into a test scenario and ask the designated reviewer to investigate it using the records, permissions and tools they will have in practice. Check whether new evidence can be submitted, whether the reviewer can act on it and whether a correction reaches the other systems affected.
The exercise should reveal who owns the case when responsibility passes between teams. Access to evidence must also be appropriate to the task, with controls that protect personal information.
After launch, challenges should inform monitoring. If several cases reveal the same problem, correcting them individually leaves the underlying cause in place. Someone needs authority to investigate the pattern and decide whether the data, model or business process should change.
A low number of complaints will not settle whether the service is working fairly. It may reflect a good experience, or a difficult complaints process. Organisations need to examine what users encountered and what happened when they sought help.
AI ethics covers much more than disputed decisions. But this is a useful place for leaders to make their commitments concrete.
Before approving a consequential AI-assisted process, ask the team to show how a well-founded challenge would be handled. The answer should identify the evidence, the responsible people and the action they can take.
That is something an organisation can demonstrate before asking people to trust the system.
