International AI Ethics Guidelines: Standards, Differences and Practical Implementation

International guidelines for AI ethics are intended to prevent artificial intelligence from being developed solely based on technical or economic standards. They formulate rules for transparency, fairness, data protection, security, human control and social responsibility. That sounds abstract at first. In practice, however, these principles influence how companies select applicants, evaluate loans, support medical diagnosis or control autonomous machines.

The most influential frameworks include UNESCO’s Recommendations, the European Union’s Ethics Guidelines, the OECD’s AI Principles, and the IEEE’s Technically Focused Standards. They pursue similar goals, but differ in terms of scope, commitment and implementation. At the same time, the situation has changed as a result of the European AI Act: voluntary guidelines are becoming concrete legal obligations in certain areas.

Key points at a glance

International AI ethics guidelines define common values for the development and use of artificial intelligence.
Recurring principles include human oversight, transparency, fairness, privacy, security and accountability.
The UNESCO recommendation was adopted by all 193 member states in 2021 and forms a global framework of orientation.
The 2019 EU Ethics Guidelines describe seven requirements for trustworthy AI.
The OECD principles are politically and economically oriented, while IEEE and standards organizations focus more on technical implementation.
Ethical guidelines are usually so-called soft law. The AI Act now makes individual requirements binding law.
Companies not only need an ethics paper, but also documented responsibilities, risk assessments, control mechanisms and ongoing monitoring.

What are international guidelines for AI ethics?

International guidelines for AI ethics are predominantly non-binding frameworks that determine the social, moral and human rights principles according to which AI systems should be developed, trained, used and controlled.They answer questions that cannot be solved through a purely technical examination. A system can be mathematically precise and still be unfair. It can meet high security standards and still collect unnecessary amounts of personal data. Likewise, a decision may appear statistically understandable even though the person concerned has no realistic opportunity to object to it.

AI ethics is therefore not just concerned with the algorithm. It looks at the entire life cycle:

What problem is the system supposed to solve?
Who benefits from his efforts?
Who bears possible disadvantages or risks?
What data is used?
Who is allowed to review or correct decisions?
How are errors, discrimination and abuse recognized?
Who takes responsibility if damage occurs?

You can find further articles on social, legal and technical borderline issues in the topic area Robotics and AI Ethics.

AI ethics, AI governance and AI compliance are not the same thing

The terms are often mixed, but describe different levels.

Term	Central question	Typical instruments
AI Ethics	What is socially and morally acceptable?	Values, principles, impact assessment, participation of those affected
AI governance	How is responsibility regulated within an organization?	Roles, release processes, control committees, documentation, escalation paths
AI Compliance	Which legal and contractual requirements must be adhered to?	AI Act classification, GDPR testing, evidence, audits, training

A company can formally act in accordance with the law and still make ethically questionable decisions. Conversely, a good-sounding code of ethics is no substitute for legal scrutiny. Resilient AI governance therefore connects all three levels.

Why does AI need international ethics rules?

AI systems are rarely developed and used in just one country. Training data can come from multiple states, the model runs on servers in a different jurisdiction, and the results impact people on different continents. National rules alone only cover such value chains to a limited extent.International guidelines at least create a common vocabulary. Terms such as transparency, human oversight, non-discrimination or accountability appear in almost all major frameworks today. This facilitates political negotiations, technical standardization and the development of internal company standards.

However, a common vocabulary does not solve conflicting goals. What is considered fair can be assessed differently depending on culture, legal system and application. Transparency can also be understood in different ways: Is a reference to the use of AI sufficient or does the exact decision-making process have to be explainable? It is precisely in such places that the difference between a general declaration of values and a practically verifiable specification becomes apparent.

Soft Law and Hard Law: How binding are AI ethics guidelines?

Most international guidelines belong to the so-called soft law. They are not directly enforceable laws, but can have considerable influence. Governments use them for national strategies, companies for internal policies and standards organizations as a starting point for technical standards.

Hard Law, on the other hand, includes binding regulations such as the General Data Protection Regulation or the European AI Regulation. Violations may result in supervisory measures, sales restrictions or fines.

Feature	Soft Law	Hard Law
Legal binding	Basically voluntary	Mandatory by law
Examples	UNESCO recommendation, OECD principles, voluntary codes of ethics	AI Act, GDPR, product safety law
Control	Commitment, reports, certifications, public pressure	Authorities, courts, market surveillance
Sanctions	Usually no immediate government sanctions	Depending on the regulations, bans, requirements and fines
Advantage	Flexible and quickly adaptable	Legal certainty and enforceability
Weakness	Danger of non-binding symbolic politics	Higher testing, documentation and implementation effort

In practice, a mixed form arises. Voluntary guidelines shape norms, industry standards, procurement rules and contracts. These instruments can become de facto binding even though they did not begin as law themselves.

The most important international guidelines for AI ethics

UNESCO Recommendation on the Ethics of Artificial Intelligence

The UNESCO recommendation was adopted by all 193 member states in November 2021. According to the German Commission for UNESCO, this is the first global framework for the political design and regulation of artificial intelligence.

The recommendation connects human rights with concrete areas of political action. It deals not only with data protection or discrimination, but also with education, science, the environment, health, media, communication and social participation.

The key principles include:

Proportionality and avoidance of harm
Security and protection against misuse
Fairness and non-discrimination
Sustainability
Privacy and data protection
human supervision and self-determination
Transparency and explainability
Responsibility and Accountability
Awareness and AI competence
Participation of different social groups

The broad view of possible consequences is particularly valuable. An AI system should not only be checked after damage has occurred. Risks should be made visible before introduction through an ethical impact assessment. This also includes the question of whether certain applications should be used at all despite their technical feasibility.

Source: German UNESCO Commission: International Recommendation on the Ethics of Artificial Intelligence

EU ethics guidelines for trustworthy AI

A high-level group of experts from the European Union published ethics guidelines for trustworthy AI in 2019. In this model, trustworthy means: a system should be legal, ethical and technically robust.

The guidelines list seven requirements:

Priority of human action and supervision
technical robustness and security
Privacy and controlled handling of data
Transparency and traceability
Diversity, non-discrimination and fairness
social and ecological well-being
Accountability

These requirements influenced the European debate about the AI Act. The guidelines themselves were voluntary. However, the AI regulation addresses several topics in a binding form, such as risk management, documentation, human supervision, data quality, accuracy, robustness and transparency.

The ethical guidelines still remain relevant. A law sets minimum requirements. The ethical review further asks: Is an application socially sensible? Are the effects on workers, customers or vulnerable groups justifiable? Does the use match your own value proposition?

OECD Principles for Artificial Intelligence

The OECD adopted its AI principles in 2019 and revised them in 2024. The Federal Ministry for Digital and State Modernization describes them as a groundbreaking example of non-binding standards in international digital policy.

The principles are strongly aimed at governments, business and innovation policy. They should enable technological progress without losing sight of human rights, democratic values and social well-being.

The key ideas include:

inclusive growth and sustainable well-being
Human rights, democratic values and fairness
Transparency and understandable information about AI systems
Robustness, security and protection over the entire life cycle
Responsibility of the actors involved

The OECD principles are interesting for companies because they combine AI ethics with economic development and international cooperation. Responsibility is not seen as a brake on innovation, but rather as a prerequisite for permanently accepted and scalable applications.

Source: BMDS: Germany’s work on the OECD Principles for Artificial Intelligence

IEEE and technically feasible AI ethics

The Institute of Electrical and Electronics Engineers is taking a more technical approach with Ethically Aligned Design and the P7000 series. Ethical values should not only be included in a mission statement, but should also be taken into account during system development.

Topics range from transparent autonomous systems to data protection processes and algorithmic biases to fail-safe designs. In this way, IEEE attempts to translate abstract values into requirements that developers can check, document and later partially standardize.

For example, the joint white paper from DIN and DKE on the ethics of artificial intelligence describes IEEE projects on:

ethical issues during system design
Transparency of autonomous systems
Data protection processes
algorithmic bias
fail-safe autonomous systems
Measurements of human well-being

The approach is practical, but has limitations. Technical standards can measure whether a process has been documented or a model has been tested. You cannot alone determine which social values take precedence. Such decisions require democratic, legal and public negotiation.

Source: DIN/DKE: Whitepaper Ethics and Artificial Intelligence

International AI ethics guidelines in comparison

Framework	year	Range	Focus	Liability	Practical use
UNESCO recommendation	2021	193 Member States	Human rights, sustainability, social consequences	Not immediately binding	Policy strategies and comprehensive impact assessments
EU Ethics Guidelines	2019	European Union and beyond	Trusted, legitimate and robust AI	Voluntary framework	Basic basis for governance, audit and subsequent regulation
OECD Principles	2019, revised 2024	Member and partner states	People-centered innovation and responsible economic policy	Soft Law	Political orientation and international business practice
IEEE approaches	Continuous	Technology, industry and research	Technical implementation of ethical requirements	Standards and voluntary procedures	System design, testing, documentation and certification
EU AI Act	Applicable gradually since 2024	EU market and systems affecting the EU	Risk-based regulation	Legally binding	Obligations for providers, operators, importers and other actors

The table shows: There is no one global AI ethic. The frameworks complement each other. UNESCO covers social and human rights issues particularly broadly. The EU guidelines structure trustworthy AI into seven requirements. The OECD combines responsibility with innovation and economic policy. IEEE, DIN and DKE are more concerned with measurable technical criteria.

The Common Core: Eight Principles of Responsible AI

1. Man remains in control

People must not become passive recipients of automated decisions. The more serious the consequences of a decision, the more effective human control must be. A person who only formally confirms an automatically generated suggestion is not yet a real supervisor.

An effective control concept regulates who is allowed to intervene, what information this person receives and under what conditions a system is stopped. The article Liability for damage caused by humanoid robots deals with information on responsibility for physical AI systems.

2. Decisions must be understandable

Transparency starts with simple information: those affected should be able to recognize that they are interacting with an AI or that an AI is involved in a decision. For sensitive applications, this information alone is not enough.

Organizations must be able to explain comprehensibly:

what purpose the system serves,
which data categories are used,
what limitations and known errors exist,
how results are checked,
Who those affected can contact with questions or complaints.

Complete disclosure of the source code is usually neither necessary nor sufficient. Transparency must help people understand the specific decision and its consequences.

3. AI must not automate existing discrimination

Algorithms adopt patterns from their training and application data. If this data contains social disadvantages, the system can reproduce or reinforce them. This applies to recruiting, lending, insurance, facial recognition and official decisions.

A fairness check should not only look at the average hit rate. What is relevant is whether errors affect certain groups more often. It must also be checked whether seemingly neutral variables act as proxies for sensitive characteristics.

4. Data protection begins before data collection

An ethically responsible AI system does not preemptively collect everything that is technically available. Purpose limitation, data minimization and controlled access rights must be taken into account during planning.

The Data Protection Conference recommends that organizations check, among other things, the purpose, legal basis, personal data, automated decisions, rights of those affected and the choice between open and closed systems before use.

This problem is particularly evident in mobile sensor systems. The article about humanoid household robots and data protection shows the risks posed by cameras, microphones, spatial maps and cloud processing.

Source: Data Protection Conference: Guidance on Artificial Intelligence and Data Protection

5. Systems must be robust and secure

Ethical AI must also function reliably under realistic conditions. This includes incorrect input, changed environments, attempted manipulation, unexpected user actions and attacks on training or operational data.

Tests should therefore not only cover the desired normal case. Stress tests, limit tests, safety analyzes and clear fallback mechanisms are necessary. The Federal Office for Information Security provides developers and organizations with information on attacks, risks and the safe use of AI.

Source: BSI: Security and Artificial Intelligence

6. Responsibility needs names and responsibilities

“The AI has decided” is not an acceptable assignment of responsibility. Organizations must determine who releases the system, who controls data quality, who handles incidents, and who can initiate shutdowns.

This responsibility must not be lost between the manufacturer, operator, specialist department and external service provider. Contracts should clearly regulate roles, data flows, audit rights, update obligations and response times.

7. Social and ecological consequences should be examined

AI systems don’t just cause direct impacts on individual users. They can change work processes, shift power relations, facilitate monitoring or trigger high energy and resource consumption.An ethical assessment therefore also asks which activities are automated, whether workers were involved and what alternatives exist. You can find background information on the employment law consequences in the article Robots replace jobs: What employers are allowed to do.

8. Those affected need opportunities to object and complain

A transparent system not only explains its function. It provides a way to report errors and have decisions reviewed. Complaints should be handled by an accessible authority, not by another automated dialogue system.

For applications with significant consequences, correction must be practically possible. If this path is missing, human control remains a promise without effect.

How are AI ethics guidelines and the AI Act related?

The AI Act does not replace international ethics guidelines. It translates selected principles into a binding, risk-based legal framework. The greater the possible risk to health, safety or fundamental rights, the stricter the obligations are.

The European AI regulation came into force on August 1, 2024 and will be applied gradually:

Since February 2, 2025: Bans on certain AI practices and requirements for sufficient AI competence.
Since August 2, 2025: Regulations for general-purpose AI models and parts of the governance structure.
From August 2, 2026: numerous transparency obligations, rules for high-risk AI according to Annex III and other essential requirements.
From August 2, 2027: additional rules for certain high-risk systems linked to regulated products.

For companies, this means: A voluntary AI code is not enough. Every system used should be inventoried and assessed according to intended use, role and risk class. The Federal Network Agency names, among other things, risk categorization, documentary evidence, transparency, security management and official monitoring as central elements.

The article on ethical and legal rules for humanoid robots shows which requirements play a role in physically acting systems.

Source: Federal Network Agency: Goals, target groups and schedule of the AI regulation

Implement international AI ethics guidelines in the company

Many organizations start with a general phrase like “Our AI is fair, transparent and human-centered.” That’s a good intention, but it’s still not a control. Only verifiable processes turn values into lived practice.

Step 1: Capture all AI systems

Create a central AI registry. This includes not only self-developed models, but also purchased software, chatbots, analysis tools, recommendation functions and generative AI services.

The following must be documented at least:

Name and provider of the system
Intended use
affected departments and groups of people
Processed data types
responsible internal body
Interfaces and external service providers
Date of introduction and last review

Step 2: Describe the purpose and benefits specifically

“Increasing efficiency” is too imprecise as a purpose. Describe what activity the system supports, what problem is being solved, and how success will be measured.

The counter question also belongs in the documentation: Can the goal be achieved with a less intrusive procedure? This check prevents AI from being used simply because it is available.

Step 3: Evaluate legal and ethical risks separately

The legal review examines, among other things, the AI Act, data protection, labor law, copyright, product safety and contractual requirements. The ethical review considers possible consequences that can remain problematic even if they are formally legal.

These include:

Disadvantage of individual groups
Loss of human freedom of choice
excessive surveillance
Manipulation or unwanted behavior control
Wrong decisions that are difficult to reverse
Dependence on a single provider

Step 4: Involve affected people early on

Developers and managers see a system from a different perspective than workers, customers or citizens. User groups can identify risks that do not appear in a technical test.

Depending on the application, workshops, pilot phases, works council participation, user surveys or independent expert reports are available. This involvement is particularly important in applications for children, those in need of care or other particularly vulnerable groups.

Step 5: Check data and possible biases

Document the origin, timeliness, quality and representativeness of the data. Not only check the training, but also real inputs in later operation.A model can appear fair in testing and deliver distorted results after introduction because user groups, data sources or general conditions change. Fairness is therefore not a one-off tick on a checklist.

Step 6: Make human supervision practical

Determine which decisions are prepared, recommended, or fully executed automatically. Designate people who can review and correct results.

For human supervision to work, these people need:

sufficient specialist and AI competence
Access to relevant information
enough time for the exam
the power to object to the system
Protection from the pressure to routinely confirm AI suggestions

Step 7: Define test, release and stop criteria

Before productive use, every relevant system needs documented minimum requirements. These include accuracy, error rates, robustness, privacy, fairness and security.

Also define clear stop criteria. A system should be suspended if critical errors occur, certain groups are demonstrably disadvantaged, or the provider does not close a security hole in a timely manner.

Step 8: Enable transparency and complaints

Those affected should receive understandable information about the purpose, functionality and contact options. Internal teams require more detailed technical and organizational documentation.

A complaints process should specify:

How can an error be reported?
Who checks the report?
Within what deadline will an answer be given?
How is an incorrect decision corrected?
When will the incident be escalated to data protection, compliance or management?

Step 9: Monitor systems after implementation

AI is changing through new data, model updates, changed interfaces and different ways of using it. Once a system has been released, it does not automatically remain trustworthy.

Regular monitoring should capture performance deviations, complaints, security incidents, unusual expenses, and changes in data quality. Major updates require a new risk and release review.

Ethical Impact Assessment: Practical test questions before using AI

An ethical impact assessment complements technical and legal reviews. It should start before implementation and be repeated when there are significant changes.The following questions are suitable as an internal test grid:

What specific problem does AI solve?
Who benefits measurably from the application?
Which people can suffer disadvantages?
How difficult would possible wrong decisions be?
Can those affected recognize, understand and contest a decision?
Are particularly vulnerable groups being touched?
Is the data used suitable, current and sufficiently representative?
Which discriminatory effects were tested?
What data leaves the company or the EU?
Who can stop the system or correct results?
Which alternatives without AI have been examined?
How will errors and long-term consequences be monitored after implementation?

For each risk, the probability of occurrence, severity of damage, number of those affected, detectability and reversibility should be assessed. A rare event can still be critical if it causes serious harm to people or is difficult to correct.

Practical rule: The greater the consequences of an AI decision for rights, safety, health or professional opportunities, the less an organization can rely on automatic results.

Who is responsible for AI ethics in the company?

AI ethics should not rest solely with the IT department. Technical teams can assess model quality and safety features, but cannot cover all legal, social and organizational implications.

Role	Typical responsibilities
Management	Strategy, risk appetite, resources and final responsibility structure
Specialist department	Purpose, benefit, process integration and technical results review
IT and development	Architecture, testing, security, logging and technical limitations
Data protection	Legal basis, data minimization, rights of those affected and impact assessment
Compliance and Legal	AI Act classification, other areas of law, contracts and evidence
Information security	Attack surfaces, access, incidents and protective measures
Works council or staff representatives	Impact on workers, control and work design
Ethics or AI Board	Interdisciplinary evaluation of controversial or particularly consequential applications

With self-obligatory guidelines for labor and social administration, the BMAS shows how people-centered, non-discriminatory and participatory rules can be anchored in an organization.

Source: BMAS: Guidelines for the responsible use of AI

Typical target conflicts of AI ethics

Accuracy versus explainability

A complex model can provide better results in certain tasks, while a simpler model is easier to understand. The highest computational accuracy is not automatically the best choice. In the case of momentous decisions, slightly lower performance may be justifiable if it makes the results verifiable and contestable.

Personalization versus data protection

The more data a system receives about a person, the more accurately it can adjust offers or forecasts. At the same time, surveillance, abuse and security risks are increasing. A responsible solution limits data to the extent necessary and offers real choice.

Automation versus human self-determination

Automated recommendations save time. However, they can lead to people questioning their own assessments less and less. This so-called automation bias is particularly risky when workers feel pressure to justify deviations from the system suggestion.

Transparency against security and trade secrets

Full disclosure may involve security vulnerabilities or proprietary company information. However, this does not justify blanket secrecy. Transparency can be achieved in stages: understandable information for those affected, detailed documents for internal auditors and protected technical documentation for authorities or auditors.

Innovation versus precaution

Restrictions that are too early can prevent useful applications. An uncontrolled introduction, on the other hand, can harm people and permanently destroy trust. Pilot projects, limited test groups, real-world laboratories and gradual releases create a middle ground.

What role do norms and standards play?

Ethical principles remain ineffective if no one can check whether they are being adhered to. Norms and standards translate general requirements into processes, measurements and technical evidence.The second German standardization roadmap for artificial intelligence was created with the participation of more than 570 experts from business, science, civil society and politics. It contains six overarching recommendations for action and more than 100 identified needs for standardization.

Areas covered include:

Basics and terms
Ethics and Responsible AI
Quality and conformity assessment
IT security
industrial automation
Mobility
Medicine
sociotechnical systems
Financial Services
Environment and energy

For example, standards can specify how risks are documented, models tested or data quality assessed. They cannot replace the political determination of social values.

Source: DIN: German Standardization Roadmap for Artificial Intelligence

Limits and criticism of international AI ethics guidelines

Many principles remain too abstract

Terms such as fairness, trust or human control sound clear, but allow for different interpretations. Without measurable criteria, any organization can claim to be acting responsibly.

Ethics washing does not replace changes

“Ethics washing” occurs when companies publicly promise ethics but do not create responsibilities, review procedures or consequences internally. A mission statement without a budget, control rights and escalation channels is, above all, communication.

Global power differences are not automatically resolved

Many countries and population groups provide data or are affected by AI impacts, but have little influence on the development of the systems. The UNESCO recommendation attempts to incorporate this perspective more fully through international participation and a focus on countries with different economic conditions.

Principles can collide with each other

Greater transparency may impact privacy or security. Higher accuracy may require models that are more difficult to explain. Ethical quality is therefore not reflected in whether a company uses as many keywords as possible. It is reflected in how openly conflicting goals are documented and justified.

Control does not end with market launch

Models, data and usage contexts change. An originally justifiable system can become problematic due to updates, new areas of application or changing user groups. International guidelines only have an impact when organizations permanently monitor their systems.

Conclusion: AI ethics moves from a guiding principle to a verifiable process

International guidelines for AI ethics create a common framework for a technology that transcends national borders, industries and social areas. UNESCO, EU, OECD and IEEE have different priorities, but agree on key values: AI should serve people, remain comprehensible, do not systematically disadvantage anyone and be operated safely.

However, the time for purely voluntary declarations of intent is running out. The AI Act makes several requirements legally tangible. Companies must record their AI systems, classify risks, identify responsibilities and train workers.

The real difference in quality arises in everyday life. Trustworthy AI is not reflected in glossy codes, but in documented tests, accessible complaint channels, human intervention options and the willingness to switch off a problematic system.

The overview of current development of artificial intelligence shows how quickly technical possibilities change. For the social classification of autonomous machines, it is also worth taking a look at the modern interpretation of the robot laws by Isaac Asimov.

Frequently asked questions about international guidelines for AI ethics

Are international guidelines for AI ethics legally binding?

Most international AI ethics guidelines are not directly legally binding and are considered soft law. However, they influence laws, technical standards, contracts and internal company policies. The European AI Act makes individual requirements binding for certain AI systems.

Which international organization has the most comprehensive AI ethics guidelines?

The 2021 UNESCO Recommendation is one of the most comprehensive global frameworks. It has been adopted by all 193 UNESCO member states and covers human rights, data protection, fairness, sustainability, education, science and social participation. Its strength lies in its global political reach.

What are the seven requirements for trustworthy AI?

The EU guidelines mention human oversight, technical robustness, data protection and data governance, transparency, diversity and non-discrimination, social and environmental well-being and accountability. These requirements serve as a framework for development and use. Several topics can also be found in the AI Act.

What is the difference between AI Ethics and the AI Act?

AI ethics describes social and moral principles for responsible systems. The AI Act is a binding European regulation with risk-based obligations and possible sanctions. A system can comply with the AI Act and still require further ethical review.

How can companies implement AI ethics in practice?

Companies should maintain an AI registry, document intended uses, examine legal and ethical risks and establish clear responsibilities. In addition, there are tests for security and discrimination, human control options, complaint channels and ongoing monitoring. A general code of ethics alone is not enough.

What is an Ethical Impact Assessment?

An ethical impact assessment is a structured impact assessment before and during the use of an AI system. Possible effects on human rights, fairness, self-determination, data protection, security and social participation are examined. The results should be documented and re-evaluated if significant changes occur.

Sources and further information

German UNESCO Commission: Recommendation on the ethics of artificial intelligence
Federal Network Agency: Goals, target groups and schedule of the AI Act
Federal government: The European AI Act
BMDS: OECD Principles for Artificial Intelligence
Data Protection Conference: Guidance on AI and Data Protection
Federal Office for Information Security: Artificial Intelligence
DIN: German standardization roadmap for artificial intelligence
DIN/DKE: Whitepaper Ethics and Artificial Intelligence
Platform Learning Systems: IT security, data protection, law and ethics
BMAS: Guidelines for responsible AI use