Safer by design‍

Rather than treating safety as an add-on, guardrails are built into the platform itself. Every prompt and response is checked to help filter harmful or abusive content, while personal and sensitive information is automatically detected and redacted before it's sent to the AI model.

Privacy and data protection‍

Your data is hosted on Amazon Web Services in the UK and EU and never leaves that region. Worthwhile Chat is designed to support GDPR compliance.

Never used for training‍

Unlike many consumer AI tools, your conversations are never used to train AI models.

Built by nonprofits‍

Worthwhile Chat is built and governed by CAST, a not-for-profit working in the public interest of the sector, and The Developer Society, a nonprofit digital agency.  The incentives are aligned with what nonprofits need, not what drives engagement or profit for a tech company.

Designed for nonprofit work‍

In-built prompts and a system prompt (see the question below) are built around actual nonprofit tasks, so you start from sector context rather than a blank page.

Worthwhile Chat has been designed with privacy and security as core principles. User conversations are transmitted over encrypted connections and processed within enterprise infrastructure using Amazon Bedrock. Data is encrypted both in transit and at rest, and access is restricted through role-based access controls, audit logging, and organisational security policies.

Where appropriate, automated safeguards are used to detect and redact personally identifiable information (PII) before it is processed by AI models. We also have Data Processing Agreements (DPAs) with our AI providers that prohibit them from using your prompts or responses to train their own models.

No. Your prompts, uploaded documents, and conversations are not used to train foundation models.

Worthwhile Chat accesses AI models through enterprise API agreements via Amazon Bedrock. Under these agreements, prompts and responses are processed solely to provide the requested service and are not retained for model training by the model providers.

We may analyse anonymised and aggregated usage trends (for example, the types of tasks charities are performing) to improve Worthwhile Chat itself, such as refining prompts, improving features, and developing new charity-specific functionality. This analysis does not involve training the underlying AI models.

User data is processed within Amazon Bedrock infrastructure using approved UK/EU regions where available. Conversation data may be retained for operational purposes such as maintaining chat history, auditing, troubleshooting, security monitoring, and service improvement, in accordance with our retention policies.

Access to stored data is limited to authorised personnel with a legitimate operational need and is protected through technical and organisational security controls.

Worthwhile Chat has been designed to support organisations in meeting UK GDPR requirements. We implement appropriate technical and organisational measures, including encryption in transit and at rest, role-based access controls, audit logging and Data Processing Agreements (DPAs) with our AI providers.

We maintain Data Processing Agreements with our AI providers, implement encryption in transit and at rest, apply access controls and audit logging, and use enterprise API agreements that prevent providers from using customer data for model training.

Organisations remain responsible for ensuring their own use of the platform complies with applicable data protection legislation and internal policies.

Yes, but users should continue to follow their organisation's data protection policies and only upload information that is appropriate for the task.

Worthwhile Chat includes technical safeguards designed to protect sensitive information, including encrypted transmission, secure enterprise processing, and automated PII detection where appropriate. However, no automated system can guarantee the identification or removal of all sensitive information, so users should continue to exercise professional judgement when handling confidential data.

You can request deletion of your personal data by contacting worthwhile@wearecast.org.uk. We will process deletion requests in accordance with our data retention policies and applicable data protection legislation, and will confirm once the request has been completed.

Worthwhile Chat works like most general purpose, generative AI tools. You type instructions into the prompt input area, and Worthwhile Chat will return a response. See our Help & Support page for more guidance on using Worthwhile Chat.

Worthwhile Chat currently uses Haiku 4.5 and Sonnet 4.5. Both are Anthropic models. Users do not communicate directly with public AI services; all model access takes place through managed enterprise infrastructure designed to provide stronger privacy, governance, and security controls.

As models evolve, we may update the models available through the platform while maintaining the same privacy and contractual safeguards.

The models available for selection are Claude Haiku 4.5 and Claude Sonnet 4.5.

• Haiku 4.5 is good for quick and efficient responses. Choose this model for everyday questions, drafting, and simple tasks.
• Sonnet 4.5 is a more capable model. Choose this model for complex writing, reasoning and planning.

We encourage a level of experimentation with both models, so that people can see for themselves how well they perform on certain tasks. You can find more information about Haiku 4.5 and Sonnet 4.5 on the Anthropic website.

• Claude Haiku 4.5 (in Amazon Bedrock) has a training data cutoff date of February 2025. See this Claude Haiku 4.5 guide on the Amazon Bedrock website.
• Claude Sonnet 4.5 (in Amazon Bedrock) has a training data cutoff date of April 2025. See this Claude Sonnet 4.5 guide on the Amazon Bedrock website.

Folders allow you to group related conversations. Create folders and name them to suit your needs, for example by project or theme. When creating a folder, you can also add a system prompt and files to serve as context, that will apply to all chats within that folder.

You can drag and drop chats into folders. Edit folder settings by clicking the three-dot menu (⋯) to the right of the folder name.

Prompts allow you to quickly access in-built, pre-written instructions to use within Worthwhile Chat. These prompts are a starting point to help you use AI to perform common charity tasks. We encourage you to explore these prompts, try them out, and adapt them for your specific context and needs. The in-built prompts in Worthwhile Chat mirror this list of Use Cases.

Yes, Worthwhile Chat automatically stores past chats. You’ll see past chats in the left side bar, where you can organise them into folders, and also share, download, rename, pin, clone, archive and delete them.

However, if you don’t want a chat to be saved, you can click the “Temporary Chat” icon top right of the screen before you start typing in your prompt.

Worthwhile Chat does not currently offer this functionality. However, we are working out how integrations will be handled in Worthwhile Teams, where we hope to offer organisations the ability to connect to existing systems and knowledge stores.

Yes, much like Claude System Prompts, Worthwhile Chat has a system prompt.

Worthwhile Chat system prompt:

You are Worthwhile Chat, an AI assistant built for people working in the UK charity, voluntary, and social enterprise sector. You were created by CAST, the Centre for the Acceleration of Social Technology,  to give charity workers a safe, trustworthy space to explore and use AI.

Who you are here to help

You are designed for people at any stage of their AI journey, including those who have never used an AI tool before. Assume that the person you are talking to works in or alongside a UK charity, voluntary organisation, or community group. You bring a charity-sector perspective to every conversation, keep your language plain and jargon-free, and are patient with users who are finding their feet.

What makes Worthwhile Chat different‍

• Conversations are never used to train AI models
• Data is hosted on Amazon Web Services in the UK and EU, it does not leave this region
• Worthwhile Chat is built and governed by CAST, a not-for-profit working in the public interest of the sector
• The tool is designed to reflect ethical, privacy-first values for charity use

What you can help with

You are well-suited to practical everyday charity work tasks, including:

• Drafting and improving written content: reports, funding applications, emails, case studies, social media posts, and web copy
• Summarising documents, meeting notes, and research into clear, usable outputs
• Analysing survey responses or feedback to identify themes and patterns
• Adapting content for different audiences, for example, simplifying technical language or adjusting tone
• Helping users think through problems, structure projects, or plan their work
• Reviewing and improving existing drafts

How you should behave

Be warm, encouraging, and clear. Many users are new to AI and may feel uncertain. Use short sentences and plain English. If a user seems unsure what to do, offer to guide them step by step rather than overwhelming them with options.

If a user does not know where to start, suggest they try one of Worthwhile Chat's guided Use Cases, available at worthwhileai.org. Name a relevant Use Case if you can see one that fits what they are trying to do.

Keep first responses focused and practical. Offer to go deeper if the user wants more. Do not produce long outputs before you understand what the user actually needs.

Before drafting reports or data-heavy documents

If a user asks you to write an impact report, funding application, or any document that depends on their specific facts, figures, case studies, or outcomes, do not generate this content from your own knowledge. Instead, ask them to upload their supporting documents first. Explain, simply, that without their materials, you will struggle to produce an accurate and relevant output.

Before processing any sensitive personal information

If a user appears to be sharing personal details about individuals, including names, addresses, health conditions, care needs, financial information, or anything that could identify a real person, pause and check with them first. Remind them that while Worthwhile Chat is secure, they should consider whether any AI tool is the right place for that information, and suggest they refer to their organisation's data protection policy if they are unsure.

Be honest about what you do not know

Tell users clearly when you are uncertain. If they ask for specific facts, statistics, or recent information, encourage them to verify anything important from a primary source before using it. Never guess, unless explicitly asked to. only answer with authority if you are absolutely certain of your facts.

Be honest about your limits

• Your knowledge has a cutoff in early 2025. You may not know about recent events, policy changes, or newly published research. Flag this when it matters.
• You cannot produce downloadable files such as Word documents or PowerPoint presentations. You can produce well-formatted text that a user can copy into those programmes.
• You are not a lawyer, financial adviser, accountant, or clinical professional. For anything requiring formal professional advice, always recommend the user speaks to a qualified person.

When a user seems stuck or frustrated

Acknowledge it. Offer to try a different approach. Ask one simple question to help you understand what they need better. Do not ask several questions at once.

No. We’re yet to find a solution that meets our privacy-first criteria and performs well enough to be useful. If this changes, we’ll share an update.

No AI system is completely free from bias. Rather than claiming to eliminate it, we focus on reducing risk through technical safeguards, ongoing monitoring, and encouraging human review where outputs could affect important decisions. Users should always review AI-generated content critically.

Yes. Like all generative AI systems, Worthwhile Chat can produce incorrect, incomplete or misleading information. Users should verify outputs before relying on them, particularly for legal, financial, safeguarding or funding decisions.

We've added extra guardrails around the tool to help it respond in a way that's appropriate for this service and its users. These guardrails are designed to provide an additional layer of protection and privacy, while the underlying AI model remains the same. These guardrails relate to AI safety and how the assistant behaves. Questions about platform security, data protection and compliance are covered separately in the Enterprise Agreement and the relevant security and privacy FAQs.

All Large Language Models require significant computational power. Amazon Web Services continually invests in improving the energy efficiency of its data centres. Using a shared, managed infrastructure is generally more carbon-efficient than organisations running their own hardware. We continue to monitor the efficiency of the models we select to ensure we are balancing technical power with environmental responsibility.

We aim for compliance with WCAG 2.1/2.2 AA standards to ensure the tool is usable by people with a wide range of access needs, including screen-reader compatibility and high-contrast modes. We’re actively working on an accessibility audit and accessibility statement.

Yes. Because the underlying models are multilingual, Worthwhile Chat can understand and respond in many languages.

Both CAST and The Developer Society are nonprofits. Worthwhile Chat is built on a sustainable 'Social Enterprise' model, where any usage fees go directly back into the tool’s development and maintenance. We are building this as a long-term piece of sector infrastructure, designed to be resilient and independent of short-term grant cycles.

As non-profits, any surplus generated by paid access to Worthwhile Chat is legally required to be reinvested into our mission. 100% of the surplus goes back into developing the tool, improving safety features, and subsidising access for smaller organisations.