Google

A common strategy in AI alignment involves creating a 'surrogate objective,' which is a measurable, proxy goal designed to approximate the true, often more complex, intended objective. The AI system is then trained to optimize for this surrogate.

Surrogate Objectives in AI Alignment

Critically evaluate two potential surrogate objectives for training an AI to achieve the true goal of creating 'informative and unbiased' news summaries. For each option, identify at least one strength and one significant potential weakness.

Evaluating Surrogate Objectives for a News-Summarizing AI

An AI development team is tasked with modifying a social media platform's content recommendation algorithm. The true, intended objective is to 'improve the mental well-being of its users.' The team proposes three different measurable surrogate objectives to train the AI on. Evaluate the following three options. In your response, identify the most promising surrogate objective and justify your choice by analyzing the potential benefits and, more importantly, the potential negative consequences or failure modes of each option.

1. Maximize daily user engagement time on the platform.
2. Maximize the ratio of positive reactions (e.g., 'like', 'love') to negative reactions (e.g., 'angry') on content shown to the user.
3. Maximize scores on a voluntary, daily in-app survey asking users to rate their mood.

Evaluating Surrogate Objectives for a Mental Well-being AI

A development team is training an AI to write helpful and engaging online tutorials. The true, complex objective is to 'create content that effectively teaches users a new skill.' To make this measurable, the team chooses a surrogate objective: 'maximize the word count of the tutorial and the number of technical terms used.' Which of the following outcomes is the most likely form of misalignment to result from this choice?

Learn Before

Related