An Illusion of Unlearning? Assessing Machine Unlearning Through Internal Representations

The project addresses a critical 'leaky abstraction' in AI safety: the fact that Machine Unlearning (MU) often only masks outputs while leaving the underlying features intact in the weights. While the insight is highly significant for the research community, as a project it functions as a diagnostic tool rather than a defensible product. With 0 stars and 4 forks (likely from the authors/collaborators), it is in its earliest stages of dissemination. The methodology (internal representation probing) is a standard technique in mechanistic interpretability applied to a new problem (MU). Frontier labs like OpenAI and Anthropic are already heavily invested in similar internal auditing for safety and alignment; they are likely to adopt these evaluation techniques internally, reducing the need for third-party tools. The primary 'moat' here would be the specific benchmarks or datasets used for testing, but the code itself is easily reproducible by any ML researcher familiar with linear probing or activation patching. Its value lies in the 'red-teaming' insight it provides, which will likely be absorbed into broader AI safety evaluation frameworks (like those from Giskard or NIST) within 12-18 months.