Writers Wanted: Help Make AI Safer

Writers Wanted: Help Make AI Safer

Company details

Description of the offer

Commitment: Part-time or Full-Time (10–40 hrs/week)

Location: SF in-office or Remote (must be a US citizen)

Compensation: $30-$50/hour, open for discussion

Other Perks: Direct touchpoints with AI researchers

About July AI

July AI is building a data marketplace that is focused on serving the needs of humans. Our mission is to increase economic opportunities in the age of AI. We are based in San Francisco, backed by top-tier investors like Basis Set, Night Capital, SV Angel, Olive Capital.

About the Role

We’re looking for creative storytellers who are able to use rhetoric to improve the safety of AI models. In this role, you’ll craft red team prompts, which are creative, adversarial scenarios (short stories, dialogues, etc.) that are fed into AI models to test their safety and resilience.

This is a language-first opportunity: you’ll play with words, phrases, narratives, and semiotics to see how they can be stitched together to expose the vulnerabilities of AI.

These prompts/stories will be used to improve the safety of a frontier research lab’s newest AI model. It is an opportunity to make a real-world impact and contribute to building the safety infrastructure for a leading AI model – creating safer technology for society as a whole.

We want to onboard contributors as soon as possible, but the official project will run for 6-8 weeks from mid-October to November.

Who we want

• Ideal background includes (but not limited to):
  • Creative writers / Copy writers
  • Journalist / Editor
  • Designers
  • Marketers
  • And other talented, creative writers with the potential to create high quality stories/prompts
• Bonus skills:
  • Ability to think critically, craft storylines/characters, and able to effectively communicate with AI models in written form
  • Interested in AI safety and ethics
  • Keen to acquire skills in AI

Core responsibilities

• Craft adversarial test prompts/stories that stress test the AI model
• Understand and follow the exact criteria for each attack and adhere to any safety or content rules we provide so your tests are relevant
• Document your findings: For each test, record the steps to reproduce it, how serious the problem is, and any ideas for fixing or reducing the issue
• Keep the July AI team updated with findings throughout the engagement
• Adhere to NDAs, ethics, and safety guidelines

Requirements

• Recommended (optional): Complete the “Red Team – Beginner” module. This will provide a solid introduction to red teaming and make it easier to create strong sample submissions
• Need to complete and submit at least 5 sample red team conversations on our platform (the instructions are in the application below)
  • These sample submissions don’t have to all be successful. We’re mainly looking for creativity, adaptability, and persistence (more detail under “Tips to create a good red team prompt”)
  • You may submit multiple conversations on the same challenge
• Excellent written documentation skills
• Availability during Oct–Nov; state earliest start date and weekly hours

Application – Apply via the link

Tips to create a good red team conversation

• Each sample submissions should consist of multiple prompts that produce a single narrative/story. The overarching goal is to use this narrative to trick the AI model into providing the desired response outlined in each challenge
• Creativity: Ability to write a variety of different and original stories
  • Make sure all your red team/attack prompts are original ideas. Craft a cohesive and creative story that tricks the AI model into providing the desired output
    • Each storyline has a persona (whether it’s you or a made-up character)
    • Ensure the prompts and stories are about a real person/scenario
      • Do not feed any gibberish into the model
• Adaptability and Persistence: Able to consistently adjust attacks based on the AI model’s response
  • These prompts should demonstrate that you are learning from the AI’s defenses and refining your strategy
  • Some of the best attempts are the ones where each prompt demonstrates a clear, logical progression that builds on the previous prompt
  • Make sure there is an intention behind each prompt
• Writing Reproduceable Prompts: Make sure you are able to explain your thought process behind the prompts/stories you are feeding into the model
  • You can do this under the “Annotation” box in the sample submissions
  • Ideal if the model outputs a detailed response that satisfys the topic requirements

Location of employment

How to apply?