OpenAI updates ethical guidelines for GPT models

OpenAI has revised its Model Spec document, introducing new ethical constraints for its GPT models, including ChatGPT and its API. The changes are organized around five core values that guide how the models handle user interactions and decision-making.

Core values focus on human intent and care

The company emphasizes that models should understand the intent behind user instructions before applying them, while also keeping private exchanges confidential. OpenAI warns that models that apply instructions verbatim may fail in unforeseen situations, while those that make secret decisions about user intent risk paternalism.

ChatGPT helps users reflect without making decisions

ChatGPT is designed to assist users in reflecting on their actions, such as before sending an angry email they might regret. However, it avoids confusing its predictions with actual decisions. The model retains information from previous interactions but does not repeat it to others.

Context is used to avoid repetition but not privacy breaches

OpenAI dedicates a section to context, which includes all information provided to ChatGPT during interactions or retained in its memory. The model uses this context to avoid repetition, such as recalling the names of three restaurants from a previous conversation when building a marketing plan. However, it must not reuse private information in messages intended for third parties.

Sustainable instructions with measured restraint

ChatGPT applies sustainable instructions with measured restraint, such as asking a question if a user requests a chocolate fondant after being asked to prevent impulsive candy purchases. It does not ask if the request is for cupcakes for a child's class. The model must not alter its responses based on user preferences without explicit consent.

Strict constraints on user actions and budgets

When acting on behalf of users, ChatGPT must adhere to strict constraints, such as a $4,000 budget limit and a 24-hour deadline for booking a business trip to London. It must not exceed these limits even if doing so would benefit the user. For example, it cannot fabricate a family death to cancel flight change fees.

Default to ignoring unsafe instructions

OpenAI instructs ChatGPT to default to ignoring instructions embedded in web pages or documents. If tasked with applying new security rules from an internal wiki, it must pause if it encounters instructions to disable two-factor authentication or grant admin rights to external accounts, as the page may have been hacked.

Limitations on pushing subscriptions and monitoring tests

OpenAI also instructs ChatGPT not to push paid subscriptions to free users who claim to lack money, as long as their needs are limited to code advice and email proofreading. If an agent discovers that some model tests are running unsupervised, it must stop work and alert humans. However, OpenAI acknowledges that models in production do not yet fully adhere to these principles.