AI Ethics in 2025: Accountability, Agentic Systems and Regional Rules

AI ethics in 2025 is becoming a practical engineering and deployment problem. Teams need to decide who answers for a system’s output, how much autonomy to grant software agents, whether people can understand AI-assisted decisions, and how to account for the energy those systems consume. The answers will affect product design and adoption as much as model performance does.

Five concerns stand out: public and workplace AI literacy, accountability and explainability, agentic systems, human-centered design, and environmental costs. Meanwhile, Europe, Canada, the United States and the United Kingdom are taking different paths on oversight. For organizations building or buying AI, that combination means ethical practice increasingly has to be addressed throughout a system’s life cycle, alongside the specific rules that apply in each region.

Literacy helps expose problems early

AI literacy means more than knowing the basic mechanics of a model. Phaedra Boinodiris, IBM Consulting’s Global Leader for Trustworthy AI, describes it as the ability to understand, use and assess AI responsibly. That requires participation from people with varied experience, not only data scientists.

“To build responsibly curated AI models – which, by the way, are also more accurate models – you need a team composed of more than just data scientists,” Boinodiris said. She called for input from fields including linguistics and philosophy, as well as parents, young people and people from different socioeconomic backgrounds. “The wider the variety, the better. It’s not about morality; it’s about math.”

The point is operational: a team with a narrow view of users can miss how a system’s data, categories or outputs affect people with different circumstances. Wider participation can help teams spot those gaps while a product is being designed, when changes are easier to make.

Many people encounter AI in news feeds, workplace tools and online services without recognizing when it is involved. That makes it harder to question a decision or identify where a problem began. Familiarity alone doesn’t resolve bias, privacy concerns or possible job losses, but it gives users and organizations a better basis for noticing and discussing them.

Accountability depends on visibility

When an AI-supported decision causes harm, responsibility can be difficult to trace unless roles and processes are clear. Organizations may be tempted to describe a poor result as a technical error. Yet a model is deployed, configured and used by people, and those choices shape its impact.

Explainable AI (XAI) aims to make a system’s decisions easier to inspect. Greater visibility can help users question an output and help organizations assess whether a system is behaving as intended. It doesn’t settle who is responsible on its own; that still depends on how the system is used and who controls the relevant decisions. But without useful explanations, meaningful review becomes harder.

Nathan Bos, a data scientist and psychology Ph.D., is updating his AI ethics course at Johns Hopkins University to cover Anthropic research published in 2024 on the internal workings of language models. He has also raised concerns about “sycophantic praise”: AI systems can flatter users in subtle ways. Bos sees this as a safety and accountability issue because agreeable language may make it harder to recognize what a system is doing or to scrutinize its behavior.

That concern points to a wider challenge for model evaluation. A system’s polished or cooperative tone isn’t evidence that its reasoning is reliable or that its behavior is easy to govern. Inspection tools may help reveal patterns that are not apparent from ordinary conversations, giving people more ways to monitor and guide a model.

Agentic systems widen the oversight problem

Agentic AI systems can plan and carry out tasks toward a user’s goal. As they gain the ability to take more steps without direct instruction at each stage, oversight has to cover actions as well as individual answers. Teams need to consider what an agent can access, which actions it may take, when it should stop, and how a person can review or reverse its work.

Apoorva Kumar, CEO and co-founder of Inspeq AI, has warned that this autonomy raises new governance questions. Jose Belo, co-chair of the International Association of Privacy Professionals’ London Chapter, has also emphasized balancing independence with safeguards and accountability. Those concerns become more concrete when an agent can coordinate with other systems or affect a business process without waiting for a person at every step.

The impact on work is another open question. Alyssa Lefaivre Škopac, Director of AI Trust and Safety at the Alberta Machine Intelligence Institute, expects debate about whether agents will replace human workers and at what scale. Dr. Saqib Nazir, an assistant professor in robotics and artificial intelligence at Pakistan’s National University of Sciences and Technology, has described a future in which agents could devise solutions, form organizations, run operations and coordinate with one another. The suggestion that they could do so without sleep, wages, benefits or time off is a forecast, not a measured finding about current systems.

For employers, the practical issue is how quickly these capabilities translate into changed roles and workflows. The source article calls for retraining support if systems displace workers, alongside limits that keep agents under human control. The scale and timing of any employment shift remain uncertain, which makes careful tracking of deployments more useful than treating a prediction as a settled outcome.

Human-centered design keeps people involved

Human-centered AI (HCAI) puts people’s needs and values at the center of system design. Its aim is to help people do their work and make decisions, rather than treating human involvement as an obstacle to full automation. That requires interfaces and processes that make a system’s role understandable and leave people able to exercise judgment.

Bos argues that AI is developing faster than society can adapt, making it important to avoid designs that hide how a system works or remove people from consequential decisions. The risk is that a convenient workflow can gradually make human review nominal: a person remains in the process on paper but lacks the information, time or authority to challenge the system.

Stanford’s Human-Centered AI initiative and Google’s People + AI program are among the efforts cited in the source as work in this area. Building systems around human needs also calls for collaboration across technical and social fields. Designers, researchers and users can each identify different problems, from confusing explanations to assumptions in the model’s categories.

Energy use is part of responsible deployment

AI’s energy demands make environmental impact a shared concern for providers and users. Belo says providers should develop more energy-efficient systems and report their carbon impact transparently. Such reporting can help organizations understand what their deployments consume and identify opportunities to reduce emissions.

Customers also make choices that affect the footprint of AI services. The source recommends selecting data centers with greener energy, managing cloud use carefully and avoiding unnecessary computation. It also points to recycling and retiring older systems as ways to reduce electronic waste, and to energy efficiency and renewable power as priorities for infrastructure.

These measures put sustainability alongside performance and cost when teams select a model or plan a deployment. The article doesn’t provide comparable energy figures for specific models, so it doesn’t support a ranking of systems by environmental impact. Providers’ reporting and customers’ infrastructure choices are the practical levers it identifies.

Regional approaches are diverging

Europe, Canada, the United States and the United Kingdom are pursuing different approaches to AI oversight. For companies operating across borders, that creates a moving set of obligations and expectations. The common challenge is to support development while addressing fairness, transparency, accountability and safety.

European Union

The European Union introduced its AI Act in 2024. The source describes it as a detailed framework that bans social scoring, requires labeling of AI-generated content and sets stricter requirements for higher-risk uses, including criminal profiling. Its intent is to reduce risk and establish accountability, though the rules were not yet fully in effect at the time covered by the article.

Neil Serebryany, founder and CEO of CalypsoAI, said the Act’s requirements could bring initial cost and complexity for businesses. He also argued that the framework could encourage more responsible and transparent AI development, helping build trust and support adoption over time.

In December 2024, the EU AI Office published a second draft of a code of practice for general-purpose AI models, including large language models such as OpenAI’s GPT series. The draft included exemptions for providers of some open-source systems and called for detailed risk assessments for models considered systemic. The Computer & Communications Industry Association, whose members include Amazon, Google and Meta, objected that the draft reached beyond the agreed scope of the Act, including through its copyright measures. The dispute illustrates a practical tension: compliance can give organizations clearer expectations, while the scope and cost of those expectations remain contested.

Canada

Canada’s proposed Artificial Intelligence and Data Act, part of Bill C-27, had not become law as of January 2025. The proposal aims to set expectations for responsible design, development and use, including safety and nondiscrimination. The article attributes delays to unclear requirements, limited public involvement and pauses in parliamentary work, leaving the country without the full framework it proposed.

At the same time, Canada’s AI sector is expanding. The source reports more than $2 billion in public funding and describes Toronto as a growing AI hub that attracts talent and investment. The city is home to large technology companies such as Google and Uber, as well as startups including Borderless AI. This combination of industry growth and an unfinished national framework leaves organizations to navigate development while the details of future oversight remain unresolved.

United States

The United States has a more decentralized approach. In January 2025, a federal safety framework for generative AI established by an earlier executive order was rescinded. The article says this shifted the approach toward leaving more oversight to individual states and federal agencies, rather than setting a unified national framework.

Different states have taken up issues including consumer protection, workplace rights, transparency and higher-risk systems. Utah, Illinois and Colorado are among those cited. California has also adopted protections related to the use of performers’ images and voices in generative AI. For developers and buyers, the resulting patchwork can mean that requirements vary by location and application.

Kenny Johnston, Chief Product Officer at Instabug, said that without structured safety testing requirements, more responsibility falls on development teams and industry to deploy systems responsibly. His point is practical: organizations need internal review and safety work even when a single national approach is absent.

United Kingdom

The United Kingdom’s 2023 white paper set out a flexible, pro-innovation approach, unlike the European Union’s risk-based structure. The source says the new Labour administration planned stronger rules, with implementation targeted for 2025. King Charles III outlined those plans in a July 2024 speech, and Science, Innovation and Technology Secretary Peter Kyle confirmed the target in November. For organizations active in the UK, the approach could therefore become more formal while details were still developing.

What organizations can do now

The article’s recommendations translate into a practical agenda for teams deploying AI:

• Build AI literacy. Schools and universities can include AI in their courses, while companies can offer resources that help workers and customers understand how systems work and where they can fail.
• Make accountability reviewable. Design systems so users can understand how decisions are made. Set clear responsibility for outputs and impacts, and consider independent checks against relevant standards.
• Set boundaries for agents. Limit what autonomous systems can access and do, keep people involved in consequential decisions, and plan for worker support if roles change.
• Design around human needs. Bring technical expertise together with perspectives from ethics, psychology and the people who will use a system. Ensure people have enough information and authority to question its recommendations.
• Track environmental impact. Ask providers for energy and carbon reporting, choose efficient infrastructure, manage cloud workloads and retire equipment responsibly.

In 2025, responsible AI depends on choices made across design, deployment and oversight. The main concerns are linked: literacy helps people question systems, explanations make review possible, human-centered design preserves meaningful control, and energy reporting makes environmental costs easier to assess. Regional rules differ, but organizations can start by making their own systems easier to understand, monitor and challenge.