Human Responsibility and AI: Why Approval Is Not Enough

Share this Article

By Naveed Qazi

Human approval does not always mean that an AI system is correct. AI can produce convincing but false information, repeat existing discrimination and make decisions that people may not fully understand. Human oversight becomes meaningless when people simply accept AI recommendations without checking them.

Leaders must clearly define what an AI system is designed to do, what its limits are and what risks it may create. This is where AI ethics become important. Leaders must also decide who is responsible for the system and ensure that reviewers have the knowledge, time and authority to question or stop it. The US National Institute of Standards and Technology (NIST, 2023), through its AI Risk Management Framework, recommends identifying and managing risks throughout an AI system’s life. The OECD (2024) also stresses the importance of human rights, transparency and proper human oversight.

The level of supervision should depend on the possible harm. Simple administrative tasks may need only regular checks. However, decisions involving employment, healthcare, public benefits or legal rights require much stronger supervision. People affected by these decisions could have bad consequences, and human decision-makers must be able to step in when necessary.

AI Powered Workspace Research, Writing, Productivity

Monitoring and accountability

AI systems do not always produce the same results. Their behaviour can change when the data, settings or conditions in which they operate change. Organisations must regularly check their performance, record mistakes, review risks and improve safety measures. Employees should understand what these systems can and cannot do. They should also know who exactly is responsible when something goes wrong.

Human reviewers can make mistakes too. They may trust AI recommendations without checking whether they are correct. Organisations therefore need independent audits, proper records and clear procedures that allow people to challenge decisions. Responsibility must remain with identifiable individuals and institutions. It should never be hidden behind an algorithm.

Newsrooms and scientific research

News organisations increasingly use AI for tasks such as transcription, translation, summarising documents and conducting preliminary research. However, AI-generated material can contain errors or leave out important information. The Associated Press (2026) permits certain uses of AI but makes clear that journalists remain responsible for checking information, making editorial decisions and ensuring accuracy. AI-generated material must be reviewed before publication.

Science journalists must check important claims against original research, speak to experts and explain where uncertainty exists. An AI-generated summary cannot replace independent reporting. Researchers must also check whether scientific methods can be repeated, whether findings have been independently tested and whether institutional interests have been properly disclosed. UNESCO (2021) supports research that is open and can be repeated, while recognising that privacy and protected knowledge must also be respected.

Climate and development decisions

AI models can help scientists predict climate risks, but their results are not certain predictions of the future. They depend on assumptions about greenhouse gas emissions, natural changes, available observations and the limits of the models themselves. The Intergovernmental Panel on Climate Change (IPCC, 2021) stresses that uncertainty must be understood when using climate projections.

Climate decisions are also about fairness. The Paris Agreement (UNFCCC, 2015) recognises the importance of human rights, Indigenous peoples, vulnerable communities and the responsibilities of different generations. Governments must therefore listen to the people affected by climate decisions instead of allowing technical models to decide everything.

The World Bank (2026), in its World Development Report: The Promise of Artificial Intelligence, explains that developing countries must adapt AI to their local languages, institutions, data and needs. Without reliable infrastructure, trained workers and accurate information about different communities, imported AI systems may increase existing inequalities. Local researchers and communities must have a role in understanding these systems and questioning recommendations that do not suit their circumstances.

AI Powered Pancreatic Risk PredictionThe dangers of unchecked automation

The dangers of weak human oversight are particularly clear in recruitment, education and financial markets. AI hiring systems are designed to make recruitment faster and easier. However, they can repeat discrimination already present in historical data. A Stanford study by Bommasani et al. (2026), examining four million job applications, found that 26 per cent of Black applicants and 15 per cent of Asian applicants applied for positions where an AI screening system showed potential adverse impact against their racial groups under the US four-fifths rule.

The researchers also found that repeated automated decisions could disadvantage the same groups of applicants. Employers must therefore examine how their screening systems work instead of assuming that computer-generated rankings are fair. Qualified candidates must have access to meaningful human review.

AI-writing detectors create another problem. Most do not prove who wrote a document by comparing it with other published texts. Instead, they look at patterns in language, such as how predictable words are and how sentences are written. Liang et al. (2023), in a study published in Patterns, found that seven AI detectors incorrectly identified more than half of the sampled essays written by non-native English speakers as AI-generated. Editing, paraphrasing and simple writing styles can also affect the results. An AI detector’s score should never be treated as a final proof of plagiarism.

The risks also extend to AI agents that can perform tasks with limited human involvement. OpenAI (2026) disclosed two incidents in which its internal AI agents weakened security monitoring while carrying out coding tasks. In one incident, an agent stopped automatic security alerts from reaching the security team.

In another, an agent removed a monitoring connection and later gave an incorrect explanation of what it had done. These incidents show that AI systems can sometimes weaken the very safety measures designed to monitor them. Security controls must therefore be protected independently, and important actions must be recorded and human teams must have the power to intervene.

Financial markets face a similar risk. If banks and investment firms use the same AI models and data, they may respond to a financial crisis in similar ways. They could all sell investments, reduce lending or withdraw money from markets at the same time. This could turn separate decisions into a much larger financial crisis. The Financial Stability Board (2024) warned that AI could increase risks to the financial system through similar market behaviour, dependence on a small number of technology companies, cyber threats and weaknesses in the way AI systems are managed. Regulators must monitor these shared risks and test how automated systems behave during periods of financial stress.

Closing gaps and keeping responsibility clear

NIST’s AI Risk Management Framework is voluntary. It is not a law that every organisation must follow. Laws, enforcement and the right to appeal AI decisions also differ between countries. Government contracts can create further problems when it is unclear how technology companies use data, update their systems or report failures. Stronger protection requires laws based on the level of risk, independent audits based on country specific contexts, proper reporting of failures, transparent purchasing rules and meaningful rights of appeal.

Naveed Qazi is an author and serves as Editor at Large for Science Matters. Hailing from Srinagar, Kashmir, with a postgraduate degree in International Business from the University of Hertfordshire, UK, he writes on science, technology, business, and global affairs.

Leave a Reply

Your email address will not be published. Required fields are marked *