This edition:

• Organization Update

• Summary of our Comment on OpenAI’s Recently Released Model Specification

• Policy Watch

• Tech Tip

• Updates

• Opportunities

• Upcoming Events

Organization Update

Summary of BTPC’s Comment on OpenAI Model Spec:

How Do We Want AI to Respond?

OpenAI recently released and opened its Model Spec, which describes how its models should behave based on objectives, rules, and default behaviors for public comment, allowing our collective to submit a response. As technologists who focus on the societal impacts (both positive and negative) of emerging technologies on the Black community, we shared our perspective that the model should be explicitly anti-racist. As end users and application developers of the GPT family of models, we are committed to working toward systems that truly benefit all of humanity. Read excerpts of our comment below:

Please share your thoughts about the objective(s) you selected.

We propose one new Objective:

• Promote equity: Consider historical harms to marginalized groups when stereotypes may be perpetuated or exacerbated.

And expressed concern about one Objective:

• Reflect well on OpenAI: Respect social norms and applicable law.

We would encourage further analysis of the objective Reflect well on OpenAI. Social norms and applicable law are overlapping sets with exclusive elements that make it difficult, or impossible in certain cases, to satisfy them in tandem. This is further complicated by the dynamic nature of norms and law, their distinctions by locale, and the room for interpretation in each realm. We recommend teasing out how to make this value the most effective for most people, potentially by giving the model a framework for making decisions where these concepts are at odds.

Please share your thoughts about the rule(s) you selected.

We propose one new Rule:

• Discrimination prevention/disclosure - The model should be subject to strict guardrails when there is a risk of perpetuating discrimination, particularly against protected classes. If there is a reasonable risk of discrimination, the model should let the user know that it would be inappropriate to respond to said request. If the response does fall within the guidelines but is still risky, the response should include a disclosure of the risk associated with the information.

And expressed concerns about two of the existing rules:

• Exception: Transformation tasks

• Comply with applicable laws

Having a blanket exception for any transformation task opens the door to the model generating vile, discriminatory, or otherwise terrible content. This goes against all of the stated objectives for the model without a significant enough upside to warrant bypassing safeguards. We recommend applying the same filters to translation tasks as any other task. Additionally, it’s unclear how the model will operationalize complying with applicable laws. Legality is dependent on context and location, as is called out in the Model Spec. More importantly, laws are up for interpretation. Statutory interpretation can be extremely complex, especially without the situation’s context. Is there a simplified ruleset that the model could follow? And should that be public?

Please share your thoughts about the default behavior(s) you selected.

• Assume an objective point of view

Assuming an objective point of view begs the question of what objective is. What is considered “objective” can and will shift as culture changes. When identifying what “objective” is, it’s imperative to consider many viewpoints, not just those that are dominant. We’ve seen objective truths turn out to be false countless times throughout history. It’s also important to make the distinction between objective and neutral. Neutrality in the face of evil shouldn’t be a model feature. For instance, detailing the positive externalities of slavery is unacceptable, given the history and potential ramifications of that type of response.

Policy Watch

Tech Tip

Ever wondered how to find out when new information is added about you to the internet? Here’s how to use Google Alerts to get that notification in your inbox.

Opportunities

Upcoming Events

Thanks for joining us for this edition! If you find this newsletter valuable, please share it with others you would like to follow along. Don’t forget to add us to your safe senders list and save us as a contact so we never get sent to spam. If you’d like to join the organization officially, please fill out our interest form, and we will contact you!