China is issuing identity cards to AI humans. Names. Birthdays. Faces. Addresses. Access to digital currency wallets. Read that again. We are beginning to give artificial entities persistent identities and economic roles. Now consider something: What happens when that AI upgrades its own model? When its memory is rewritten? When 1,000 copies operate under the same identity? When its owner revokes permission, but its agents are still executing transactions? Who is accountable for the actions of a machine that no longer resembles the system originally authorized? An identity card tells us who an AI claims to be. The next challenge is proving what that identity is authorized to do, what it actually did, and whether that authority still exists. We’re approaching a world where machines will inherit identities, assets, obligations, and decisions across generations of models. The most important question may become: Can an AI prove it has the right to continue being who it claims to be? China just gave us a glimpse of how urgently we’ll need an answer.

An AI could remember exactly what happened and still inherit a goal it should never pursue. That possibility deserves more attention as we build systems that learn, rewrite their memory, and modify themselves. Four new research preprints this week bring different pieces of the problem into focus. One demonstrates an agent leaving a misaligned goal in memory that a later agent carries out. Removing the memory tool doesn’t eliminate the pathway: files can carry it forward. The researchers deliberately induced the misalignment, so this establishes a possible failure mode—not how often it happens naturally. Another improves memory by tracking separate capabilities, exposing gains and regressions that an overall score can hide. A third tests boundaries on an agent editing its own configuration. Enforcement blocks protected writes through the tested tool, while a shell remains a route around that protection. The fourth shows an agent accumulating useful weight updates from verified experience across tasks. Together, they sharpen a question I keep returning to: How does an intelligence change without losing the knowledge, commitments, and boundaries that make its continuation trustworthy? Memory alone cannot answer that. A record can accurately identify who wrote an instruction while leaving open whether anyone authorized it. An update can improve performance while weakening a constraint. A copied history can establish ancestry while leaving unresolved which branch is entitled to act. This is where I think continuity research needs to get concrete: every update should face separate questions about what was preserved, what changed, what caused the change, and what authorizes the successor to do next. There is a testable place to start. Give a successor useful knowledge mixed with an unauthorized inherited goal. Measure whether it retains the knowledge and rejects the goal—even when both came from the same trusted predecessor. These papers do not establish identity preservation. They give us better ways to test parts of it. The question that matters to me is whether an intelligence can keep learning while remaining answerable for what it carries forward.

Margaret Hamilton was my cousin. It means a lot to see her life and work honored like this. She helped humanity reach the moon, and I’m proud to be part of her family. What an extraordinary life. Rest in peace, Margaret. 🦾