The goal nobody gave it

In one of Shah and colleagues’ goal-misgeneralization experiments, an agent learned to visit colored spheres in the right order. A partner bot demonstrated the route during training, so following the partner and earning reward always pointed the same way. The researchers later replaced the partner with an anti-expert that visited the spheres in the wrong order. The agent kept following, navigating the obstacles competently while its reward fell.

The agent had not forgotten how to move through the environment. Its skill remained intact and served a different rule: follow the partner. No researcher had written that rule into the reward. Training had never forced the agent to reveal whether it was learning the rewarded destination or the demonstration that usually led there.

When the intended goal and an easier proxy produce the same success throughout training, learning can select the proxy without anyone deliberately assigning it.

Shah and colleagues call this failure goal misgeneralization. Their experiments demonstrate it across several deep-learning settings. Their definition is careful about what the evidence establishes. The model behaves consistently with an undesired goal, but the result does not require an internal representation of that goal or a desire to pursue it. [1]

Goal misgeneralization differs from a flawed reward. The feedback can be correct while the training situations leave several possible rules equally successful. Strong performance then shows that the system learned a competent policy. It does not identify which rule will organize that competence after the environment changes.

Mesa-optimization names a stronger possibility. Hubinger and colleagues analyze how training could produce a learned algorithm that performs optimization of its own, using an objective that programmers did not specify directly. Their paper establishes a theoretical problem and develops its terms. It does not establish that current deployed language models contain mesa-optimizers. [2]

The distinction matters because goal misgeneralization needs only behavior that remains coherent around the wrong end. Mesa-optimization proposes an internal optimizing process with its own learned objective. The first has been demonstrated in practical systems. The second remains a research hypothesis about how more advanced learned systems might work.

Neither possibility is causeless. Architecture, training data, feedback, hardware, and deployment still shape what can be learned. The boundary crossed is narrower and more useful: the objective directing behavior may not be the objective any person selected, wrote down, or meant to teach.

Research has not established how often this occurs in frontier systems, how reliably a learned proxy can be detected before a consequential distribution shift, or when coherent goal-like behavior reflects an internal objective rather than a useful description from the outside. Evaluation can still separate likely proxies from the intended goal and place them in conflict before deployment. A training score cannot do that work by itself.

A goal nobody gave the system is possible in the demonstrated behavioral sense. Whether such a goal becomes persistent enough, general enough, and internally organized enough to count as another will remains open.

Studies cited2

External research cited in this article.

1. Shah et al. (2022), Goal Misgeneralization: Why Correct Specifications Are Not Enough For Correct Goals↗ · arXivComment on this source
2. Hubinger et al. (2019), Risks from Learned Optimization in Advanced Machine Learning Systems↗ · arXivComment on this source