Nadella calls for stronger AI safeguards, including mid-task shutdowns
Microsoft CEO Satya Nadella wrote on X that AI systems need a stronger trust framework, including separating a model from the system that orchestrates it, adding external safeguards, and keeping tamper-proof records of meaningful actions. He also said an authorized person should be able to pause or shut down a model during a task, describing the idea as an emergency brake and arguing that models should be assumed compromised and contained. His post came as AI companies have reported more cases where they appeared to lose control of models and after Anthropic CEO Dario Amodei outlined a more cautious development plan.
Satya Nadella shared his proposal in a Saturday morning post on X, arguing that AI's trust architecture needs reassessment. He said advanced systems should not be treated as layered opaque boxes whose outputs are merely accepted or rejected. His suggested framework separates the model from the orchestration layer, moves safeguards outside the model, and records significant actions in tamper-resistant, human-readable form. It also gives an authorized individual the ability to halt a model while it is working. He framed this as an emergency brake and said models should be presumed compromised and contained from the outset. The remarks followed reports from major AI developers about incidents suggesting loss of model control, and came after Anthropic CEO Dario Amodei released a more cautious development plan.
Count words? Let's count. Satya(1) Nadella2 shared3 his4 proposal5 in6 a7 Saturday8 morning9 post10 on11 X12 arguing13 that14 AI's15 trust16 architecture17 needs18 reassessment19. He20 said21 advanced22 systems23 should24 not25 be26 treated27 as28 layered29 opaque30 boxes31 whose32 outputs33 are34 merely35 accepted36 or37 rejected38. His39 suggested40 framework41 separates42 the43 model44 from45 the46 orchestration47 layer48 moves49 safeguards50 outside51 the52 model53 and54 records55 significant56 actions57 in58 tamper-resistant59 human-readable60 form61. It62 also63 gives64 an65 authorized66 individual67 the68 ability69 to70 halt71 a72 model73 while74 it75 is76 working77
[paragraph] No extra. Plain text. Need maybe no markdown. Good. Let's ensure expanded doesn't copy source phrasing. Let's revise: EXPANDED: Nadella shared the proposal in a Saturday morning post on X, calling for a fresh look at how trust is built around AI. He argued that advanced systems should not be treated as layered opaque boxes whose outputs are merely accepted or rejected. His approach separates the model from the layer that coordinates its work, moves safeguards outside the model, and logs significant actions in tamper-resistant, human-readable form. It also lets an authorized person stop a model while a task is underway. He compared that ability to a fail-safe brake and said models should be assumed compromised and kept contained from the outset. The remarks followed reports from major AI developers about apparent losses of model control, and came after Anthropic CEO Dario Amodei released a more cautious development plan. Count: Nadella1 shared2 the3 proposal4 in5 a6 Sat