Explain the legal framework and purpose of codes of conduct under Article 95
Distinguish between mandatory obligations and voluntary commitments
Evaluate whether to adopt or develop a code of conduct for your organisation
Identify key elements of effective codes of conduct
Understand how codes of conduct intersect with sustainability and accessibility
Introduction: Beyond Compliance
The AI Act establishes mandatory requirements for high-risk AI systems—but what about the majority of AI systems that fall outside the high-risk category? Article 95 encourages the development of voluntary codes of conduct that extend trustworthy AI principles to all AI systems.
Expert Insight
Codes of conduct represent the AI Act's "soft power"—encouraging responsible AI development even where hard legal requirements don't apply. Organisations that embrace these principles often find they're better prepared if regulations tighten or if their AI systems are later classified as high-risk.
Legal Framework (Article 95)
What Article 95 Establishes
Provision
Content
Encouragement
Commission and Member States shall encourage drawing up of codes of conduct
Scope
Applicable to AI systems other than high-risk AI systems
Purpose
Voluntary application of requirements for high-risk AI systems
Sustainability
Specific focus on environmental sustainability
Accessibility
Consideration of accessibility for persons with disabilities
Stakeholder input
Codes shall take into account input from relevant stakeholders
Commission and Member State Role
Activity
Description
Facilitate development
Provide guidance, convene stakeholders, support drafting
Take stock of existing codes
Consider industry initiatives already in place
Consider SME needs
Ensure codes are accessible to smaller organisations
Promote uptake
Encourage adoption through awareness and incentives
These practices are banned, not subject to voluntary codes
Subject Matter Areas
Article 95 specifically mentions several focus areas:
Area
Article 95 Reference
Examples
High-risk requirements
Article 95(1)
Voluntary application of Articles 8-15
Environmental sustainability
Article 95(2)(a)
Energy efficiency, carbon footprint, sustainable design
Accessibility
Article 95(2)(b)
AI usable by persons with disabilities
Stakeholder participation
Article 95(2)(c)
Involving affected groups in AI design
Diversity
Article 95(2)(d)
Diverse development teams
Environmental impact measurement
Article 95(2)(e)
Tracking and reporting sustainability metrics
Environmental impact measurement
Article 95(2)(e)
Tracking and reporting sustainability metrics
Voluntary Application of High-Risk Requirements
Which Requirements to Consider
Requirement Area
Article
Voluntary Application Benefits
Risk management
Article 9
Identify and mitigate risks proactively
Data governance
Article 10
Improve data quality, reduce bias
Documentation
Article 11
Better internal knowledge, easier maintenance
Record-keeping
Article 12
Accountability, debugging capability
Transparency
Article 13
User trust, clearer communication
Human oversight
Article 14
Better decisions, reduced automation failures
Accuracy and robustness
Article 15
Higher quality systems, fewer failures
Expert Insight
Organisations that voluntarily apply high-risk requirements to their non-high-risk AI often find unexpected benefits: better quality systems, fewer production issues, and easier scaling when they do develop high-risk AI.
Environmental Sustainability
AI's Environmental Footprint
Impact Area
Concern
Code of Conduct Response
Training compute
Large models require massive energy
Efficient training practices, renewable energy
Inference energy
Ongoing energy for AI operations
Model optimisation, efficient hardware
Hardware lifecycle
Manufacturing, disposal of AI hardware
Sustainable procurement, recycling
Data centres
Cooling, power infrastructure
Green data centre practices
Sustainability Commitments in Codes
Commitment Type
Example
Measurement
Track and report AI carbon footprint
Reduction targets
Reduce AI energy consumption by X% per year
Efficiency practices
Implement model compression, pruning, quantisation
Renewable energy
Power AI workloads with renewable sources
Hardware choices
Select energy-efficient hardware
Lifecycle consideration
Factor sustainability into AI design decisions
Article 95(2): Specific Sustainability Provisions
The AI Act specifically calls for codes of conduct to include:
Provision
Implementation
Environmental impact measurement
Methodologies for measuring AI's environmental footprint
Resource consumption tracking
Monitor energy, water, materials usage
Reporting mechanisms
Standardised reporting of environmental metrics
Improvement targets
Set and track sustainability goals
Accessibility
Making AI Accessible
Accessibility Principle
Application to AI
Perceivable
AI outputs can be perceived by all users (text alternatives, etc.)
Operable
AI interfaces can be operated by all users
Understandable
AI explanations are comprehensible to diverse users
Robust
AI works with assistive technologies
Specific Considerations
User Group
AI Accessibility Considerations
Visual impairment
Text alternatives for visual AI outputs, screen reader compatibility
Hearing impairment
Captions for audio, visual alternatives for voice interfaces
Motor impairment
Alternative input methods, voice control
Cognitive differences
Clear explanations, appropriate complexity levels
Expert Insight
Accessibility isn't just ethical—it's good business. Making AI accessible expands your user base and often improves usability for everyone. Codes of conduct that emphasise accessibility signal inclusive design values.
Stakeholder Participation
Who Should Participate?
Stakeholder Group
Value of Participation
End users
Practical needs, usability feedback
Affected communities
Rights and fairness perspective
Domain experts
Technical and professional standards
Civil society
Public interest representation
Regulators
Regulatory expectations alignment
Academics
Research insights, ethical frameworks
Participation Methods
Method
Description
Advisory boards
Ongoing stakeholder input on AI development
User testing
Direct feedback on AI systems
Public consultations
Broader input on AI policies
Impact assessments
Stakeholder involvement in evaluating AI impacts
Grievance mechanisms
Channels for stakeholder concerns
Diversity in AI Development
Why Diversity Matters
Diversity Dimension
AI Development Impact
Gender diversity
Reduces gender bias in AI design and outputs
Ethnic diversity
Improves fairness across populations
Disciplinary diversity
Brings varied perspectives (tech, ethics, social science)
Neurodiversity
Innovative problem-solving approaches
Age diversity
Balances experience and fresh perspectives
Commitments in Codes
Commitment
Implementation
Diverse teams
Recruitment and retention targets
Inclusive culture
Ensure all voices are heard in development
Diverse testing
Test AI with diverse user populations
Bias awareness
Training on recognising and addressing bias
Developing Effective Codes of Conduct
Key Elements of Effective Codes
Element
Description
Importance
Clear principles
Specific, actionable commitments
Guides behaviour
Implementation guidance
How to apply principles in practice
Enables action
Governance structure
Who oversees code implementation
Accountability
Monitoring mechanisms
How compliance is tracked
Evidence of effectiveness
Reporting requirements
What participants must report
Transparency
Review process
How code is updated
Continuous improvement
Enforcement/consequences
What happens if code is breached
Credibility
Code Development Process
Adopting vs. Developing Codes
Strategic Options
Option
When Appropriate
Advantages
Disadvantages
Join existing code
Industry code exists, fits your needs
Faster, established credibility
Less tailored to your context
Develop new code
No suitable code exists, unique needs
Fully customised
Resource-intensive
Adapt existing code
Existing code needs customisation
Balance of fit and efficiency
May require negotiation
Internal code only
Competitive differentiation, unique approach
Full control
Less external credibility
Considerations for SMEs
Factor
SME Approach
Resources
Join existing codes rather than developing new ones
Credibility
Leverage established industry code reputation
Flexibility
Choose codes with proportionate requirements
Demonstration
Use code adoption as trust signal to customers
Benefits of Code Adoption
Business Benefits
Benefit
Description
Customer trust
Demonstrates responsible AI commitment
Competitive differentiation
Stand out from competitors
Risk management
Proactive approach reduces future issues
Regulatory readiness
Prepared if requirements tighten
Talent attraction
Ethical AI attracts principled employees
Stakeholder relations
Positive engagement with civil society, regulators
Operational Benefits
Benefit
Description
Quality improvement
Code requirements drive better practices
Knowledge building
Documentation and transparency improve internal understanding
Issue prevention
Proactive risk management catches problems early
Scalability
Good practices transfer to new AI projects
Examples of Code of Conduct Elements
Environmental Sustainability
SUSTAINABILITY COMMITMENT EXAMPLE:
We commit to:
1. Measure the carbon footprint of all AI training runs
2. Report annual AI energy consumption publicly
3. Achieve carbon neutrality for AI operations by 2027
4. Prioritise model efficiency in architecture decisions
5. Use renewable energy for at least 80% of AI compute
Accessibility
ACCESSIBILITY COMMITMENT EXAMPLE:
We commit to:
1. Apply WCAG 2.1 AA standards to all AI interfaces
2. Test AI systems with users with disabilities
3. Provide alternative formats for AI outputs
4. Ensure AI works with common assistive technologies
5. Maintain accessibility throughout AI lifecycle
Stakeholder Participation
STAKEHOLDER COMMITMENT EXAMPLE:
We commit to:
1. Establish an AI Ethics Advisory Board with external members
2. Conduct annual stakeholder consultations on AI practices
3. Publish impact assessments for significant AI deployments
4. Maintain grievance mechanisms for AI-related concerns
5. Report publicly on stakeholder engagement activities
Code of Conduct Adoption Checklist
Preparation
Assess which AI systems are in scope (non-high-risk)
Research existing codes of conduct in your sector
Evaluate strategic options (join, develop, adapt)
Identify stakeholders for input
Assess resource requirements
Adoption
Select or develop appropriate code
Obtain leadership commitment
Allocate implementation resources
Communicate code to relevant staff
Establish monitoring mechanisms
Implementation
Implement code requirements in AI practices
Train staff on code commitments
Begin monitoring and reporting
Engage stakeholders per code requirements
Document implementation evidence
Ongoing
Report on code compliance per requirements
Participate in code governance
Review and update implementation
Address any non-compliance
Contribute to code evolution
What You Learned
Key concepts from this chapter
**Codes of conduct are voluntary** but demonstrate responsible AI commitment beyond legal requirements
**Primary focus** is on non-high-risk AI systems where mandatory requirements don't apply
**Environmental sustainability** is a key theme—measuring and reducing AI's environmental footprint
**Accessibility** extends AI benefits to persons with disabilities
**Stakeholder participation** brings diverse perspectives to AI development