Introduction 

The European Disability Forum (“the Forum or,” “the organisation”) is committed to using artificial intelligence (AI) in a way that is fair, ethical, safe and respectful of everyone’s rights. As AI becomes more common in our daily work, it’s important to have clear rules on how to use it properly. This policy gives guidance to all Forum’s staff on how to use AI tools responsibly protecting people’s data, avoiding misuse, and helping the organisation make good decisions. 

The Forum supports the safe, responsible and ethical use of AI. AI is used only as a tool to assist human expertise, never to replace it. We recognise the opportunities and risks of AI and will continuously monitor developments to ensure our practices remain accurate, ethical and accountable. 

Purpose of the Policy  

The Forum’s staff is already using a range of AI tools to support work, like writing, translation, and brainstorming. An internal survey showed that while staff see many benefits from AI, they also have serious concerns. These include: 

• Data protection and privacy 

• Ethical risks and bias 

• Lack of control and knowledge about how AI works 

• Trustworthiness and accuracy of AI-generated content 

This Policy exists to respond to these concerns and to support staff in using AI safely and ethically. It aims to: 

• Protect personal data and confidential information 

• Prevent the misuse of AI 

• Ensure AI use reflects The Forum’s values and mission 

• Promote fairness, inclusion, transparency and accessibility 

• Support and train staff to make informed decisions about AI tools 

Applicability 

This Policy applies to all staff and representatives across all teams and roles, when they use AI for their work at the Forum, and it covers any use of AI systems in a professional setting. It covers any use of AI tools in a professional context, whether for communication, content creation, data handling, or decision-making.  

It reflects the Forum’s commitment to transparency, accountability, and fairness in everything we do. The policy also includes our service providers. The Forum staff responsible for working with the service providers will be responsible to ensure this is included in calls for tenders and contracts. 

Legal framework 

The AI Act is a regulation in the European Union that aims to create a framework of rules and standards for the creation, implementation and application of artificial intelligence systems within EU borders. Its main objectives are to protect fundamental human rights, ensure the safety of the public, and promote trust and innovation in AI technologies.  

This policy should also be considered together with our data protection policy and aligned with the General Data Protection Regulation (GDPR). Further, this policy should be considered in alignment with the UN Convention on the Rights of Persons with Disabilities (CRPD) in terms of non-discrimination and inclusion of persons with disabilities. The AI Act sets out rules for companies that make and use AI systems. The companies that build these systems are called providers. Companies like The Forum, that choose to use AI in their work, are known as deployers under this Act. 

Risk categories 

The AI Act divides AI systems into four risk categories based on how likely they are to affect people’s rights: 

• Unacceptable risk (forbidden AI): AI systems deemed too dangerous or incompatible with European values and therefore banned entirely. The bans went into force in February 2025 

• High risk: AI systems that are permitted but strictly regulated because they can significantly influence a person’s life chances (e.g., decisions on employment, access to loans, or educational admission). The rules on high-risk AI will go into force in August 2027. 

• General Purpose AI (Limited or Medium risk): AI systems used for a wide range of tasks and present fewer risks but still subject to transparency requirements. The rules on general purpose AI will go into force in August 2025. 

• Low or no risk: AI systems with minimal impact on people’s rights that do not require special regulation. 

Unacceptable risk – Forbidden AI 

The AI Act bans AI systems that pose an unacceptable risk to human dignity, democracy, and fundamental rights, including people with disabilities.  

These are systems that can manipulate human behaviour, exploit people’s vulnerabilities, or cause physical or psychological harm. Illustrative examples of such systems include: 

• AI systems that recognise emotions in the workplace or schools. (Example: imentive AI) 

• AI systems that collect or expand facial recognition databases from the internet or surveillance cameras. (Example: Clearview AI)¹ 

• AI systems that predict criminal behaviour based solely on personality traits or personal characteristic (example: Correctional Offender Management Profiling for Alternative Sanctions [COMPAS]) 

• The use of live facial recognition in public places in public places. Exceptions apply for serious crimes such as terrorism. (Example: Hungary’s use of facial recognition to track people who attend the Pride-parade) 

• AI systems that exploit individuals with disabilities for profit or control. 

Implications for the Forum 

The Forum will not use unacceptable risk AI systems, for example we will not use emotion recognition for hiring or in general HR management. 

High risk AI

The European Commission has presented draft guidelines on high-risk AI. 

For the Forum, the potential use of high-risk AI will be in the domain of employment. 

AI systems used in employment are high-risk because they can significantly impact job opportunities, income, and labour rights, potentially perpetuating biases and invading privacy.² Below some examples of High-Risk AI applications. 

AI for Hiring: 

• Targeted job advertisements 

• Selecting candidates 

AI for Workforce Management: 

• Decisions on promotions and layoffs 

• Task assignments 

• Monitoring employee performance and behaviour 

Implications for the Forum 

EDF will not use AI for selecting candidates for interview recruitment or Human Resource management. 

General purpose AI

Some general-purpose artificial intelligence (GPAI) systems pose ‘systemic risks’ because they can cause significant harm to public health, safety, human rights or society. Rules on general purpose AI will be in force in August 2025. 

Examples of General Purpose AI- using Google Gemini, Co-pilot, ChatGPT and similar AI systems.  

Implications for the Forum 

• The Forum will label text, video, audio, and images which is fully AI generated. If a content is partially AI-generated, the staff member and AI working group will assess if the content needs labelling (for example, more than 50 percent of the content is AI generated). 

A staff member who used AI to generate or edit text is responsible to ensure its accuracy, respect of data protection, non-discrimination, that it does not violate the intellectual property rights of others label it.   

• The Forum staff and representatives take full responsibility for content they create or disseminate.  

• The Forum has regular staff training in place to improve skills in the tools that we use and ensure compliance with this policy.  

Core principles  

This chapter sets out the core principles and responsibilities that everyone must follow when using AI tools. When using AI at the Forum, the following key principles apply: 

Fairness 

AI must not discriminate against any individual or group. All use of AI must promote equality, inclusion, human dignity, and respect for human rights.  

Transparency 

We must be open about when AI tools are used and must seek to understand how these tools function and generate outputs. 

Accountability 

We remain fully responsible for any outcomes produced with the assistance of AI. AI does not replace human judgment or decision-making. 

Privacy 

AI must be used in ways that protect personal, sensitive, and confidential information, in compliance with data protection laws and EDF’s privacy commitments. 

Safety and security 

AI must not be used in ways that could cause harm to individuals, groups, or the organisation. 

Accessibility 

AI tools should be inclusive by design and use, supporting the needs of persons with disabilities and meeting accessibility standards. 

Advancing human rights and inclusion 

AI use at EDF should contribute positively to social justice, accessibility, and the empowerment of people most exposed to discrimination or exclusion. EDF will seek to apply AI in ways that reduce inequality, amplify underrepresented voices, remove barriers to participation, and advance inclusion.  

Examples of AI tools developed and used in line with the Forum’s core principles: the Clear Writing Agent.  

The Forum Clear writing agent is an AI agent hosted on Microsoft Co-pilot. It was internally developed and updated by the Forum communications team.  

The tool is accessible to all Forum staff. The settings are based on the Forum clear writing style guide to support staff development in clear writing.  

It reviews and suggests improvement in 4 areas: clarity, accessibility, grammar, inclusive language, and alternatives. Additionally, the Forum clear writing agent can run accessibility checks, give feedback, and explain areas of improvement.  

The tool cannot generate or download files (Word, PDF, or Excel, etc.), connect to Microsoft apps like Word, Outlook, or Teams (need a Copilot Pro account to add plugin), access internal files or external content unless pasted or linked, or edit documents in real time.  

More information on the tool is available in the Forum’s internal training folder. 

Ethical issues 

Environmental impact  

In September 2024, the United Nations Environment Programme (UNEP) has published an issue note on “The Environmental Impact of the Full AI Lifecycle Needs” 

The report categorises the environmental impact of AI across its software and hardware life cycle into three levels: direct, indirect and higher-order effects. And among the environmental impacts mentioned: 

• Resource intensive production of AI hardware: for example, making a 2kg computer uses about 800kg of raw materials, including rare earth elements often mined in harmful ways. 

• Energy consumption: AI infrastructures need a lot of energy, mostly from fossil fuels. One ChatGPT request uses 10 times more electricity than a Google Search. ³ 

• E-waste: data centres generate electronic waste containing toxic substances like mercury and lead. 

• Water usage: AI infrastructures use vast amounts of water for cooling and construction. A research conducted by experts at the University of California, Riverside titled “Making AI Less “Thirsty”: Uncovering and Addressing the Secret Water Footprint of AI Models”, estimates that using ChatGPT for 10 to 50 queries consumes approximately 500 milliliters of water, primarily due to the cooling requirements of data centers that host AI models.  

EDF is aware of the environmental cost of AI and limit our use of AI, in line with our Green EDF policy. We will also encourage employees, members, partners, and key stakeholders to do the same.  

Accuracy  

AI systems can produce useful information, but they also produce errors known as hallucinations. These errors are not anomalies or rare failures. They result directly from the mathematics that underpins large language models. These systems generate text by calculating the most likely next word based on probability. They do not check facts or reason about truth. Because of this design, hallucinations are unavoidable. 

AI outputs can combine accurate information with false or invented details. This behaviour reflects how the system works. It does not indicate misuse or poor configuration. Research, including a joint report by the European Broadcasting Union and the BBC, confirms that hallucinations are a structural feature of these systems. This conclusion is reinforced by recent academic research, including the paper “LLMs Will Always Hallucinate, and We Need to Live With This”, which shows that hallucinations arise from the core mathematical design of large language models and cannot be eliminated. 

Users can reduce the frequency of errors by giving careful instructions and by verifying sources. These measures help, but they cannot remove the underlying mathematical limitation. Any claim that hallucinations can be fully prevented contradicts the established functioning of large language models. That is why AI policies like this document should explain this clearly and require users to independently check all AI‑generated content. 

These risks can appear in work in several practical situations, for example: 

• Background research: AI may produce clear and confident summaries that combine accurate information with outdated or incorrect details, which can weaken the evidence base of a policy.  

• Citing sources: AI can generate or misattribute references, creating the risk of including sources that are unreliable, incorrect, or do not exist.  

• Drafting policy briefs: The text may sound persuasive and well-structured, but still contain subtle inaccuracies, unsupported claims or missing context.  

Because of these risks, your work should always involve careful factchecking, source verification and critical review of any AI-generated content before it is used. 

Bias and discrimination in AI 

AI can sometimes be unfair because of how it is designed and the data used to train it. AI systems learn by analysing large amounts of information. If that information includes past examples of bias or discrimination, the AI may learn and repeat those patterns. It can also reflect the views or assumptions of the people who created it. 

For example, if an AI system is trained mostly on data from one group of people, it may not understand or respond well to people from other groups. This problem can be worse when using AI tools in languages other than English if the system was mainly trained in English. This kind of bias could show up in your work in different ways, such as: 

• If you ask an AI image tool to show people collecting water, it might automatically create a picture of a poor village, unless you tell it otherwise. 

• If you ask for an image of a company CEO, it might show a white man in a suit. 

• If you use a tool that creates spoken audio, the voice will probably have an American accent.  

These kinds of results can reinforce stereotypes and make existing inequalities worse. 

Staff responsibilities 

To uphold these principles, we have the following responsibilities when using AI tools: 

Awareness of AI limitations 

• We understand the capabilities, biases, and limitations of any AI tool they use. 

• We are fully responsible for the quality and accuracy of our work.  

Human oversight and responsibility 

• We can use AI tools as a starting point to support work, but always review, edit, and apply our own knowledge and judgment to ensure accuracy.  

• Information generated by AI is not always accurate or correct to ensure we don’t spread misinformation inadvertently we should always do a thorough review of information generated by AI, including by asking it to provide sources. 

• We must critically review and verify AI-generated outputs. AI cannot replace human oversight; final responsibility always lies with the persons who are responsible for the content.  

Use AI ethically and transparently 

• We must use AI in ways that align with the Forum’s ethical values and must be transparent about AI use in our work.  

• Respect for copyright and intellectual property.  AI outputs must be checked to ensure compliance with copyright and intellectual property rules. Proper attribution must be given when required.  

• We expect that we thoroughly review all AI-generated content that we use in any outputs, including verifying factual statements and citations. 

• Respectful and Inclusive content creation 

• We should ensure that the content we create respects the dignity of individuals. 

Protect Data and sensitive information 

• We must avoid sharing confidential or sensitive data with AI systems, unless data privacy and security are assured. 

• Reporting and accountability: Any issues, errors, or ethical concerns related to AI use must be promptly reported to supervisors or internal AI working of the Forum.   

• Commitment to ongoing learning: We should actively stay informed about AI developments, emerging risks, ethical practices, and regulatory requirements. 

The Forum IT tools: 

When people use AI software that are not part of the Forum’s official software package. They should inform the operations manager of the Forum and their manager. This is to the Forum can ensure that the use of this is compliant. They can ask the AI working group for an assessment.  

AI Working group  

To make sure AI is used in a safe and fair across the organisation, the Forum will set up an AI working group. This group will guide staff, monitor how AI is used and support good decision-making about AI tools and practices.  

Main tasks of the AI working group 

• Develop, review and regularly update the AI policy and practical guide 

• Share updates, good practices, and lessons learned in the practical guide 

• Review, advise and approve new AI tools before they are used 

• Make sure accessibility and inclusion are considered in all AI decisions 

• Collect feedback from staff on AI use, challenges and concerns 

• Monitor how AI is used and suggest improvements when needed. 

This group helps ensure AI is used in a coordinated, transparent and responsible way, with input from across all Forum’s staff members.