Artificial intelligence has changed substantially since the first version of this article. The discussion is no longer limited to robots performing repetitive tasks or algorithms sorting applications. Generative AI now writes, analyses, recommends, creates images and carries out multi-step tasks through AI agents. It can support employees and accelerate decisions, but also produce convincing falsehoods, reproduce discrimination and expose confidential information.
This makes one point more important than ever: AI is not your friend, but it is not your enemy either. It has no duty of loyalty, no moral judgement and no personal responsibility. Responsibility remains with the people and organisations that design, procure, deploy and supervise it.
The central ethical question in 2026 is therefore not whether a machine can think like a human. It is whether humans are delegating consequential tasks without retaining sufficient control, evidence and accountability.
The New Nature of the Problem
Traditional software usually follows defined rules. Modern AI systems infer patterns from data and produce probabilistic outputs. Generative systems communicate fluently even when their answer is wrong. Their confidence is a feature of presentation, not proof of accuracy.
At the same time, AI is becoming more active. An assistant may only draft an email; an agent may access files, call software tools, place an order or initiate a workflow. The ethical risk therefore depends not only on the model, but on the entire system: its data, instructions, permissions, interfaces, users and organisational setting.
The same technology can be low-risk in one context and unacceptable in another. An imperfect meeting summary can be corrected. An incorrect recommendation concerning employment, credit, healthcare or public services can materially affect a person’s life.
Bias Has Not Disappeared
AI can reduce some forms of human inconsistency, but it does not automatically produce neutral decisions. Historical data contains historical inequalities. Training data may underrepresent certain groups. Labels may reflect subjective judgements, and apparently neutral variables can serve as proxies for sensitive characteristics.
Bias can also arise after deployment. Users may trust automated recommendations too readily, while feedback loops reinforce earlier outcomes. A recruitment model trained on previous hiring decisions may reproduce an organisation’s past preferences rather than identify future potential.
Fairness is not a single technical score. Definitions can conflict, and the right test depends on the system’s purpose and consequences. Organisations should examine performance across relevant groups, document limitations and involve people who understand the operational context. Where serious harm is possible, affected individuals need a meaningful route to appeal.
Transparency Must Be Useful
People should know when they are interacting with an AI system or viewing synthetic content. This is especially important for chatbots, deepfakes and material that could be mistaken for authentic human communication. The EU AI Act’s transparency requirements, applicable from 2 August 2026, reflect this principle: certain AI interactions must be disclosed, and AI-generated or manipulated content must be identifiable in the situations covered by the law.
Transparency is not achieved by publishing a lengthy disclaimer or source code few people can interpret. Useful transparency answers practical questions: What is the system used for? What data does it process? How reliable is it here? What are its limitations? Who is accountable, and how can a decision be challenged?
Explanation must also be proportionate to impact. A movie recommendation needs less justification than a system that influences a mortgage or employment decision.
Human Oversight Must Be Real
“A human is involved” is not an adequate safeguard if that person lacks the time, information or authority to disagree with the system. Human review can become a rubber-stamping exercise, particularly when AI output appears polished or when employees are measured primarily on speed.
Effective oversight requires clearly assigned decision rights. Reviewers must understand the system’s limitations, recognise warning signs, access relevant evidence and be able to stop the process. High-impact uses also need escalation paths and a record of the final decision.
This is partly an education issue. Under the EU AI Act, providers and deployers must support an appropriate level of AI literacy among staff and others operating AI on their behalf. Training should reflect roles. General users need to understand verification, confidentiality and appropriate use; procurement and compliance teams also need competence in risk classification, vendor evidence and monitoring.
Privacy, Ownership and Security
Ethical AI depends on responsible data use. Before entering personal, confidential or copyrighted material into an AI system, an organisation must understand what happens to that information: whether it is retained, used for training, transferred to another jurisdiction or accessible to third parties.
The GDPR continues to apply alongside the AI Act. The European Data Protection Board has stressed that model anonymity, legitimate interest and the effects of unlawfully processed training data require case-by-case assessment. “The vendor offers AI” is not a legal basis, and outsourcing a system does not outsource accountability.
AI systems can also be manipulated through malicious instructions, contaminated data or unsafe tool access. Agents with broad permissions are particularly risky because a hostile instruction can trigger an external action. Access should follow the principle of least privilege, sensitive actions should require confirmation, and incidents must be reconstructable.
What the EU AI Act Changes in 2026
The EU AI Act uses a risk-based approach. Some practices are prohibited; providers and deployers of other systems face obligations according to their role and the system’s risk. Rules on prohibited practices and AI literacy have applied since February 2025, while governance and obligations for general-purpose AI models began to apply in August 2025. From August 2026, the European Commission’s AI Office and national authorities move into the main enforcement phase, and the transparency obligations take effect.
The timetable has nevertheless changed. The 2026 AI Omnibus extended the application dates for several high-risk-system requirements: rules for high-risk systems listed in Annex III are now scheduled for December 2027, while certain systems embedded in regulated products follow in August 2028. Organisations should therefore use the current consolidated timetable rather than older summaries of the Act.
But legal compliance is only the floor. A system can fall outside a high-risk category and still damage trust, employees, customers or the public. Ethics asks not only “May we deploy this?” but also “Should we deploy it, under what conditions, and who bears the consequences if it fails?”
Turning Principles into Practice
Ethical principles become credible only when they influence operational decisions. A practical governance process should include:
- an inventory of AI systems, including unofficial tools already used by employees;
- a named owner, a defined purpose and a clear description of the people who may be affected;
- classification of the organisation’s legal role and the system’s risk;
- checks on data provenance, quality, privacy, intellectual property and security;
- pre-deployment testing, including failure cases, subgroup performance and adversarial testing where appropriate;
- defined human decision rights, escalation procedures and accessible channels for complaints or redress;
- controls over permissions, versions, vendors, logs and material changes;
- ongoing monitoring for errors, drift, incidents and unexpected impacts; and
- role-specific AI literacy and regular review of whether the use case remains justified.
Frameworks such as the NIST AI Risk Management Framework can help translate broad values into repeatable governance. Its focus on validity, safety, security, transparency, privacy and managed bias is useful even for European organisations, although it does not replace the requirements of EU law.
Conclusion
The greatest ethical risk is not that AI suddenly develops its own morality. It is that organisations treat persuasive output as truth, automation as objectivity and vendor claims as assurance. AI can enhance human capability, but it can also scale errors and obscure responsibility.
Responsible use in 2026 means preserving human agency: knowing where AI is used, limiting what it can do, testing what it produces and ensuring that someone has both the duty and the power to intervene. The goal is not to make machines our friends. It is to make human accountability impossible to delegate.
Editorial sources
- European Commission: AI Act regulatory framework
- European Commission: Enforcement of the AI Act
- European Commission: AI Omnibus enters into force
- European Commission: AI talent, skills and literacy
- European Data Protection Board: Opinion on AI models and GDPR principles
- NIST: AI Risk Management Framework
- OECD AI Principles
Related Links:
This couldn’t have been said better. Responsibility remains with the people and organizations that design, procure, deploy and supervise AI powered tools. We must make these organizations accountable, if we’re to evolve a more ethical and human centered use of AI.