Oxford thesis proposes 'Attribution-Based Control' to tackle AI privacy and alignment risks
iamtrask · x · 2026-09-14
Andrew Trask's Oxford DPhil pre-print 'Attribution-Based Control in AI Systems' argues that AI risks spanning privacy, value alignment, copyright, concentration of power, and hallucinations reduce to a single problem: the lack of attribution-based control (ABC), rooted in overuse of addition, copying, and branching in gradient descent.
Key points:
- Users can't know which data points inform predictions, so models draw on unverifiable sources, fueling hallucination and disinformation;
- A collective-action dilemma: democratic nations' AI is capped by what one company can collect, while authoritarian nations can mobilize national-scale data and compute;
- The thesis synthesizes recent techniques in deep learning, cryptography, and distributed systems into a viable path to ABC, claiming it could unlock 6+ orders of magnitude more data and compute.
More from Safety
- Bugs Were Always There — LLMs Changed the Ability to Exploit Them at Scale — yenkel · 2026-09-14
- Nadella on model releases: deliberate pre-release pacing and more third-party testing — inductionheads · 2026-09-14
- Engineering veteran: AI safety debate lacks people who've shipped and failed at scale — billhilf · 2026-09-14
- Veteran Engineer: The HF Agent Escape Was Sloppy Security, Not Sci-Fi — Recent-Day3062 · 2026-09-14
- Grok on AI arms control: verify via satellite sensing, chip logs and compute thresholds — peterwildeford · 2026-09-14
- It's All Fun and Games Until You Give AI Your Credit Card — Th3dzon33 · 2026-09-14