Terms of Use

Last updated: 27 April 2023

Schedule

  1. Name of Service: Pair
  2. Nature of Service
    1. This Service facilitates the provision of Large Language Models (LLMs) powered applications to Singapore public sector agencies. This includes Pair Chat, a fast, powerful and enterprise-ready version of ChatGPT for Singapore public sector officers, as well as subsequent applications that GovTech may develop under the Pair name.
    2. Pair’s applications are designed to enable officers and agencies to use LLMs safely and productively in their core work. This is achieved through a number of ways, including stronger data security than COTS alternatives, enterprise controls and analytics, and guides for effective LLM usage. These features are intended to help officers leverage LLMs as assistive tools for the wide range of natural language tasks that they have been found useful for, such as summarisation, brainstorming, finessing of language, getting feedback on one’s writing, and programming.
    3. SNDGO has approved Pair’s applications for up to Restricted/Sensitive Normal information on non-Secret devices. You shall not use Pair for anything beyond what has been approved by SNDGO.
    4. GovTech is NOT responsible for ensuring the accuracy or quality of the output, or biasness of what Pair produces. Given the probabilistic nature of machine learning, the output from Pair’s underlying large language models may not correctly capture facts, events, policies or people. In fact, these large language models may “hallucinate”, presenting fictitious information in a convincing manner. As output reflects training data from the internet, Pair may also present biased or offensive viewpoints. Users should be aware of these limitations, and evaluate and ensure the accuracy of work produced by Pair before using it in their work.
    5. You are not permitted to use this Service if you are not using the Service for and on behalf of a Singapore public sector agency (“Agency”). Notwithstanding this, the Terms of Use binds your Agency and you will also be personally responsible for use of this Service. The Agency shall be responsible for all of your acts and omissions.
  3. Third party software/services
    1. Pair currently utilizes the Microsoft Azure OpenAI Service as a Limited Access Service. The applicable Third Party Terms for such components are the Microsoft Customer Agreement (including the Product Terms and any applicable addendums and annexes, as well as the Code of Conduct for Azure OpenAI Service). Some models may be in preview and the Preview Terms of Use apply in these cases. You shall not, in particular, use the Azure OpenAI Service to discover any underlying components of the models, algorithms, and systems, such as exfiltrating the weights of models and shall not use web scraping, web harvesting, or web data extraction methods to extract data from the Azure OpenAI Service or from any data or content output by the Azure OpenAI Service.
  4. Special terms
    1. Data Security
      1. No data storage or logging by third party large language model provider: Pair uses models from Azure OpenAI, a large language model provider, with modified content filtering and abuse monitoring. These models are hosted in the United States, but this arrangement for use with information up to Restricted/Sensitive Normal is possible because the request and response data sent to Azure OpenAI is not stored and unavailable for any purposes, whether for debugging, abuse monitoring, content filtering or improvement of the same, or otherwise. A description of Azure OpenAI’s data, privacy and security setup can be found at this Link .
      2. No use of government data for training of models by third parties: Any data sent from Pair Chat to Azure OpenAI is not used by Microsoft to train its own models.
      3. Built on Government Commercial Cloud: Pair’s server infrastructure utilises the Government Commercial Cloud.
    2. Large Language Models
      1. Azure OpenAI: Pair currently uses a range of models from Azure OpenAI, including text-davinci-003, gpt35turbo, and gpt4.
      2. Additional Options: Pair is working to onboard alternative large language models from a range of providers, including NVIDIA, Cohere, Google, and Anthropic. Pair will only utilise models with equivalent or superior data security arrangements as is presently available through Azure OpenAI with modified content filtering and abuse monitoring. GovTech will inform users of changes to the core large language model used in Pair’s applications, and its accompanying data security setup.
    3. Enterprise Analytics and Controls
      1. Analytics: GovTech collects usage metrics of officers’ interactions with Pair’s applications, including but not limited to number of tokens used, chat messages and sessions. In addition, while Pair Chat can operate without any data storage of message contents, GovTech may log these messages to provide agencies’ greater oversight and understanding into their officers’ use of the service. Reports relating to these analytics may be made available to agencies upon request.
      2. Access Controls: Pair Chat allows the setting of domain-level or officer-level access controls, which can be used to manage the rollout of Pair Chat in agencies and across the Government.
    4. What you are responsible for
      1. You are responsible for (in addition to compliance with the Terms of Use):
        1. Abiding by Pair’s Usage Policy
        2. Abiding by Third Party Terms
        3. Complying with PMO (SNDGO) CIRCULAR NO. 1/2023: USE OF LARGE LANGUAGE MODELS IN THE PUBLIC SECTOR (as amended from time to time) and any other applicable rules when using Pair. The AI Policy Group (SNDGG) has approved Pair for use with up to Restricted / Sensitive Normal data.
    5. Usage Policy
      1. Pair’s applications shall be used safely and responsibly. These usage policies apply to all users of Pair’s applications. If we discover that your usage does not follow these policies, we may ask you to make the necessary changes. Repeated or serious violations may result in further action, including suspending or terminating your account.
      2. These policies may change as we learn more about use and abuse of our tools.
      3. By using Pair’s applications, you acknowledge that:
        1. Large language models generate text using a statistical model. They can generate text that sounds true but is wrong about people, events, and facts.
        2. Large language models do not have access to the internet or real-time data. They have limited knowledge of facts beyond their training data.
        3. Pair’s applications are meant as assistive tools and utilise preview software components i.e. there may be issues and errors. You shall not use its output as-is without checks. Officers and their Agencies remain responsible for ensuring that their work is accurate and appropriate.
        4. Pair is only cleared for up to Restricted / Sensitive Normal information. Do not use it for any work involving higher security classifications. Be especially careful about what information you provide through your queries or the source/reference material that you provide Pair with.
      4. Pair’s applications are meant for public sector work. They shall not be used for personal matters.

Built by:

OGP