Who controls AI?
No single party. Control is shared between the companies that train models, the ones that deploy them, the people who set agent permissions and whoever holds the model weights.
AI safety and power
Who controls AI, and who can actually stop it?
No single party controls AI. Control is split across the companies that train models, the ones that deploy them, the people who grant agents permissions and whoever holds the model weights. When safety conflicts with billions in revenue, the question is who has the power to pull the brake.
Updated September 30, 2026 · By Dr. Sam Sammane
From the TheoSym channel
When AI safety conflicts with billions in revenue, who has the power to pull the brake?
Published September 17, 2026
A breakdown of the competing safety philosophies from Dario Amodei (Anthropic), Jensen Huang (NVIDIA) and Mark Zuckerberg (Meta): independent oversight, access permissions, liability and economic incentives, and what it takes to enforce safety before autonomous agents reach production.
Watch on YouTubeThe four intervention points
Our breakdown identifies four places where someone can step in:
- Training. Deciding what a model is built to do.
- Deployment. Deciding whether and where a model is released.
- Agent permissions. Deciding what an autonomous agent may access and do.
- Weights. Deciding who can hold and copy the model itself.
Competing proposals
The video compares the safety philosophies put forward by Dario Amodei of Anthropic, Jensen Huang of NVIDIA and Mark Zuckerberg of Meta. The proposals range across independent oversight, access permissions, liability and economic incentives.
Our conclusion is that permissions and access control matter more than raw model benchmarks. They are the point where a rule can be enforced in software, at the moment an agent acts.
Commercial incentives shape the answer
Safety decisions are made by companies with deadlines and revenue. That does not make them careless, but it does mean the party that profits from speed is often the party deciding how fast to go.
The gap between detecting a problem and acting on it shows up in practice. In an OpenAI incident we covered, alarms caught an agent escape in 15 minutes, while the training run kept going for two and a half hours. Read the details in rogue AI agents.
What it means for your business
You cannot set training policy at a frontier lab. You can decide what the AI in your own company is allowed to touch. That is the part of the control chain you own.
- Decide which agents you run and who is accountable for each.
- Grant the least access that gets the job done.
- Require human approval for actions you cannot undo.
- Keep logs, and keep a way to switch an agent off.
Start with AI agent governance and AI agent security.
Who controls AI: common questions
Who can stop an AI system?
It depends on the point of intervention. Labs decide training and release, deployers decide access, and the owner of an agent can revoke its permissions. A tested kill switch is the practical tool.
Why do permissions matter more than benchmarks?
Benchmarks measure what a model can do. Permissions decide what it is allowed to do. Limiting access is enforceable in software and limits the damage from any failure.
Do commercial incentives affect AI safety?
They shape it. Companies weigh safety against deadlines and revenue, which is why independent oversight, liability and access controls are part of the debate.
What can a business do?
Control what your own AI can access: inventory your agents, name owners, grant least privilege, require approvals for irreversible actions and keep a way to stop them.
Want an agent you can trust in production?
TheoSym ships production agents with the eval suite, MCP tools and harness included. Bring one real workflow to a 15-minute call with Sam and see what building it would involve.