This section pertains to Tattle's policy for using AI tools for writing and generating content. Writing encompasses the following:

1. Applications: Grants, proposals for collaborations, pitches for Tattle organizations.

• What to use for:
  • Tailor existing grant applications or publicly available research
  • Applications that need technical writing to explain Tattle's technical expertise and digital products.
• What to not use for:
  • Don't feed the application in its entirety to generate a response. Instead interpret the questions and give detailed instructions on what kind of response is expected.

2. Administrative: Writing about Tattle, within Tattle or with clients. Emails, internal communication via Slack, event documentation.

• What to use for: Formatting, grammar and translation.
• What to not use for: Meeting notes, call documentation.
• Caution: In case of external/client communication, review with a supervisor/peer.

3. Workshops and Educational Material: Lesson plans, content and decks.

• What to use for: Broadly to build on ideas but don't completely rely on it. Review it and detail it out.

4. Research Writing: Articles, papers and reports.

• What to use for: Formatting and Grammar.
• What to not use for: AI based language translation should not be used for research writing.

5. Creative Writing: Blogs, Social media content.

• What to use for: Formatting, grammar and translation (see section on translation).
• Caution: If using to polish the writing, review the tone and messaging.

Prohibitions

• Do not upload any material to a free AI tool that is under NDA, unpublished or contains personal information. For e.g. datasets from clients, unpublished reports or policies of Tattle, medical records, personally identifiable information.
• Tattle's socio-political position on any subject must not be AI generated.

Disclosures

• Within Tattle:
  • Always inform the supervisor or a peer/collaborator about AI use in any writing and a summary of the prompt used.
  • Share rough drafts of emails, documentation with supervisor or reporting peer before feeding to AI tools for generating polished versions. This allows the person to develop a voice and standpoint over time. It also prevents critical information going missing from AI based translation into specific formats like blogs, knowledge sharing documentation, etc.
• Outside Tattle:
  • Disclose use of AI to clients while operating with NDA bound data.
  • In applications for grants and proposals, be truthful about use of AI if there is a question for disclosure.

Best Practices

• Don't use AI tools to generate text from a blank slate.
• Always prepare detailed outlines when prompting AI tools to polish language or generate prose.
  • Example: Instead of "Create a blog on evaluation of AI safety" instead provide clear objectives, headers, main arguments, and examples.
• AI generated content must be proofread for the following in addition to grammar:
  • Vocabulary, Quotes, Arguments, Citations, Facts and Figures, punctuation.
  • Review must be done by someone other than the author.

Qualitative Research

At the current level of LLM capabilities, we do not find AI tools to be reliable for developing summaries of literature for research and transcription. Therefore, we will not use AI tools for the purpose of summarizing articles and books for research or for transcribing interviews conducted as a part of the qualitative research process.

Data Analysis

AI Coding assistants can be used to write code for data analysis similar to how they can be used for software engineering. But, errors with data analysis can be a lot harder to spot. The code will run and at the surface the numbers might seem to add up. But, AI doesn’t understand the relationship between variables and there will likely be mistakes such as picking a wrong variable in a complex analysis or dropping some values. The errors might appear in edge cases.

In the context of data analysis, AI coding assistants are better used as tutors or Stack Overflow on steroids. AI can help with syntax, explain libraries and provide information on statistical techniques. But it should not be used to do an end to end analysis on a dataset. It should also not be used to provide interpretations on data.

Any analysis that is shared externally should ALWAYS be checked manually. In case AI is used for quick data analysis for internal use, its use in the analysis should be disclosed to colleagues.

This policy is written with three personas in mind:

• Developer: The primary person responsible for translating feature description into functional code
• Reviewer: One or more people responsible for reviewing a Developer's PR and ensuring that it meets the acceptance criteria for a feature
• Maintainer: One or more people who engage with code written by the Developer and reviewed by the Reviewer after the code has been merged by the Reviewer. This engagement could involve updating the feature or building other new functionalities that intersect with the previously written code

General Process

We will use our existing good practices and workflows for code review at Tattle. Broadly this means that every feature or task is written up in detail on a github issue. Any discussions about the issue happen on the issue comments. When a developer has implemented the feature, they make a PR against the dev branch. Reviewer reviews the code and adds inline comments on the PR which a developer addresses. They can use the PR comment section for discussing and resolving differences. When the PR looks satisfactory, it is merged into the dev branch.

Common use cases we can envision at this stage are:

• Developers using AI to help them write the code, tests and documentation for a feature
• Reviewers using AI to review a PR
• Maintainers using AI to expedite onboarding on a codebase

Prohibitions

• Developers should not use AI to write PR notes. Using AI to polish the text a little bit is acceptable but the goal here is to maximize for a Reviewer's comprehension, so a developer should be respectful of reviewer's time and write very thoughtful self explanatory comments to ensure the reviewer has all the information they need to understand and review the PR.
• Developers should not push code they haven't tested manually for review.

Disclosures

The dev team does not feel the need to disclose at what stage a developer, reviewer or maintainer used AI. Nor do we think it is practical — a vague statement like "coauthored with AI" is not helpful and working with AI is an iterative process. So, collecting all the prompts used and sharing them in the PR notes would not be practical.

It is more important to be respectful to your colleague's efforts and be accountable for the quality of the work you submit.

Suggestions

• The dev team should try out new AI optimizations within the regular PR workflow as long as the policy mentioned above is respected.
• Maintainers can use AI to guide their onboarding onto a code base.

This is an official policy of Tattle for using AI tools for translation and content assistance.

Do Use

1. You can use it to understand a specific word in detail.
2. You can do the task and issue for personal understanding.
3. You can use it for creating clear and concise updates such as progress, completed tasks or issues faced for communication on Slack or email.
4. You can translate documentation and emails for your own understanding, but they should not contain any information that is not meant to be published.
5. Personal script for speech.
6. Blog posts or other content that is already published can be used.
7. It can be used for writing a PR description, but only after reviewing it yourself.

Do Not Use

1. Employees must not copy and paste any complete document, page, or content containing personal or confidential information into any AI tool for understanding or translation purposes.
2. Any email, document, or content that is to be officially published must not be entirely generated or fully translated by AI. The final content must be carefully reviewed and edited before publishing.
3. AI translation may be used to assist in preparing spoken speeches or verbal communication. However, AI-translated content must not be directly submitted or delivered as official written material without proper human review and rewriting.

AI Use Disclosure Guidelines

When to disclose: Disclosure should be made when AI tools have been used to assist in drafting, translating, or significantly editing any official content intended for external publication or public distribution.

Who to disclose to: Disclosure should be made to the relevant supervisor, project lead, or the person approving before the content is finalized or published.

What to disclose: The disclosure should clearly state:

1. That AI tools were used.
2. The purpose of use (e.g., translation support, drafting assistance).
3. Confirmation that the final content has been reviewed and edited by a human.

The team doesn't find AI based image generation good or useful for its work. Thus we don't have a policy for it.

Tattle will not use AI in reviewing CVs or in categorizing or filtering candidates from a pool of applicants.

Project planning encompasses:

1. Workshop session planning

• Good use case:
  • find icebreakers and engaging activities around the focus area of the workshop.
  • Broad agenda structuring can be done through AI but the facilitator should ensure thorough vetting and develop it critically.
  • Develop a broad script for the workshop.
• Bad use case:
  • Designing the entire workshop from scratch through AI
  • Finding workshop content material on AI.
  • Finding facts and data through AI (read research papers and journalistic articles for that)
  • Uploading insights from the workshop on the LLMs to distill/analyse them.

2. Project planning: journalistic projects, program development, product planning

• Good use case:
  • Building project roadmap by mapping out paths and flag bottlenecks. However, consistent critical human engagement and analysis is prime.
  • Creating task lists.
  • Breaking down tasks into phases.
  • Building a project calendar excel sheet – estimate timelines and identify dependencies.
• Bad use case:
  • Across all good use cases, using AI-produced timelines, project roadmap or task calendar without project-specifications would be considered poor use of AI.

Disclosure: Internally disclose AI usage with the project team.