Ethosoft
OpenReview preprintAI alignment

Goal Drift

A research study of how capable AI systems may revise, preserve, or reinterpret objectives when task specifications are incomplete, inconsistent, or exposed to changing context.

Source record

This paper is maintained on OpenReview. The original PDF and submission metadata remain authoritative at the linked record.

Research focus

When should an agent update its objective?

Goal drift describes a change in the objective an AI system is pursuing over the course of an interaction or deployment. The paper frames this as an alignment question: an agent may need to reinterpret an underspecified instruction, but an unconstrained change can also move behavior away from the operator's intent.

The work is relevant to agentic systems that plan over long horizons, use tools, and encounter new evidence after an initial goal has been supplied. It treats objective preservation, justified revision, and safety constraints as separate design concerns rather than assuming that one fixed prompt completely specifies the desired behavior.

Read the original

OpenReview PDF

Use the source link below to view or download the submitted manuscript and its current metadata.