Blog
Erasure, verified.
Notes on AI unlearning, right-to-erasure engineering, and evidence that survives scrutiny.
KVKK deletion requests and AI systems
A KVKK erasure request reaches the model, the retrieval layer, and the vector store, not just the database row. What Article 7 of Law 6698 requires, why deleting the record does not clear the AI, what the six-month periodic-destruction cap means for a live system, and how Turkey's 2026 enforcement is landing.
What are ghost vectors in vector databases?
A ghost vector is an embedding that survives after you delete the record it stood for. What leaves one behind, how the major vector databases defer the real cleanup, why a deleted embedding can be read back into names and faces, and what that means for a right-to-erasure request.
The right to be forgotten and AI chatbots
The right to erasure reaches AI chatbots, but honoring it is harder than deleting a row. What GDPR Article 17 requires, whether it covers what a chatbot says about you, and why a deletion you filed can still leak months later.
Does deleting data remove it from AI models?
The direct, evidence-first answer: what happens to a record after you delete it, why models and vector stores often keep the pattern anyway, and what regulators have started ordering companies to destroy instead.
What is machine unlearning?
A plain, evidence-first explainer: what machine unlearning is, how exact and approximate methods differ, why deleting data does not make a model forget, and what it takes to prove erasure held.
How much does an AI unlearning audit cost?
What teams actually pay to verify that erased data is gone from their AI surfaces, what drives the price, and how it compares to adjacent audits.