What is KeyGPT from The Key?
KeyGPT is a new, safe place to use two new ground-breaking features from The Key - KeyGPT and AI Wizard.
- KeyGPT, a customised version of ChatGPT, will help you brainstorm, summarise issues and explore ideas
- AI Wizard will help you with specific tasks like creating bespoke letters or creating interview questions. We will be adding more capabilities in the coming weeks
KeyGPT uses the latest generation of large language models (LLMs) as well as drawing on content from The Key. In addition to configuring our LLM, we use information about you and your school, e.g. your role and your school's location, name and phase, to give you more relevant answers.
How is this different from using OpenAI's GPT-4 or other AI models?
You should get better and more complete answers from KeyGPT when compared to using ChatGPT or other versions of LLMs.
We are using the very latest AI models (at present our feature is based on OpenAI's ChatGPT 4.0), prompting the AI to give you more targeted, high value responses in useful formats, and then giving the model additional prompts based on information that we know about you and your school.
We know that the quality of responses is greatly improved by giving better more detailed prompts to LLMs. Each time you ask a question, we enrich the question with information about your role and your school - its phase, location, context to give you the best possible response.
Who can use the KeyGPT?
KeyGPT is available to all staff in schools and trusts who are invited to join the beta site.
If you would like to add a member of staff to the service then you should add them via https://my.thekeysupport.com/manage-users/. There is no additional cost from The Key from adding additional users to your existing subscription.
Does the KEYGPT cost extra?
There is no charge for using these features and we envision the feature being included in the cost of your existing subscriptions going forward.
What sort of errors might I see in these new features?
Factuality and hallucinations: language models are known to 'hallucinate', confidently asserting that something is a fact when it is not true. We advise exercising your judgement and verifying the information you receive from KeyGPT independently, particularly in situations where you may rely on that information.
Historic data: all LLMs are trained on historic data - so, unlike content on The Key's services, content in LLMs is typically months, and sometimes years, out of date.
Bias: language models inherit biases and stereotypes from their training data. While we have made efforts to limit the impact of these within our AI-powered features, they may still emerge under various conditions.
KeyGPT doesn't know what it doesn't know: KeyGPT does not have real-time access to the internet and will not be aware of recent events, or of DfE policy changes or guidance updates. It also has a limited understanding of the data that was used to train it, how it processes user data and its own terms of service.
Gullibility: language models can occasionally be tricked into producing unsafe or inappropriate content on the premise that it is “hypothetical,” “for research”” or describing an imaginary situation. KeyGPT can exhibit similar patterns.
Limited memory: KeyGPT currently has a limited memory of the previous conversations you have had with it. This means that it may forget facts, and topics of discussion that you have shared with it in the past.
How is this different from content on The Key's services?
Articles and policies on The Key's services are based on the latest guidance from the government, sector bodies and best practice in the sector. They are carefully researched, written, edited and cross-checked by our content team to be accurate, up to date and practical.
In contrast, KeyGPT is an intelligent assistant with a conversational interface. The tool can help you brainstorm, explore ideas, summarise issues and generate documents.
Our editorial team does not check individual responses. You should not rely on KeyGPT answers to be as accurate, up to date or as balanced and nuanced as our existing leadership, governance or safeguarding content.
We do not recommend using KeyGPT on on legal and safeguarding matters.
Despite these caveats, we believe that you will find the KeyGPT feature extremely helpful across a large range of tasks - saving your time and helping you make better decisions.
How should school leaders not use the KeyGPT or any of our AI-powered features?
Not for personal information
- No personal data: do not share or seek personal data about staff, pupils, parents, carers or any individual
- Not for sensitive issues: avoid using the features for highly sensitive or confidential matters
Not as a decision-making device or for guidance on legal or safeguarding matters.
- Avoid sole reliance: do not use the AI features as the only source of information for critical decisions
- Not for legal advice or specific advice on sensitive topics like safeguarding where you can't afford to get it wrong: do not treat responses as legal counsel or definitive compliance guidance
What approach to AI is The Key taking?
Approach
Accuracy and providing the 'knowledge to act' sits at the heart of The Key's mission and culture. School leaders rightly expect that the answers and technology we provide should be safe, trustworthy, and reliable.
We have outlined the general principles of The Key Group's work around AI.
How are you working to improve the quality of KeyGPT?
Accuracy and safety
Delivering accuracy and safety is an iterative process. We have an ongoing process of review and improvement to verify that all our services are complying with our policies and principles. KeyGPT is subject to the same iterative process.
AI as a technology is in its earliest stages and far from perfect. As we at The Key work to continually improve our techniques and methodology, we'll share updates publicly.
Review and improvement
There are 4 elements that form our approach to reviewing and improving the behaviour of our AI-powered features:
- Monitoring: we automate monitoring to understand usage, the quality of responses, and where our models might be failing. These systems help to surface unsafe patterns and help to inform and prioritise the issues we need to fix
- Log review: quantitative methods are only one part of the KeyGPT structure. The Key examines report logs flagged by members, to surface issues and improve accuracy and safety
- 'Red teaming': we actively try to undermine our safety and anti-bias mechanisms doing to inform improvements and fine-tuning
- Your feedback is vital: a major way that we improve is through the feedback and suggestions that we get from users of our AI features. If you see something that you think should be fixed or improved, we encourage you to reach out to us at amy.ai@thekeysupport.com