When organizations delegate judgment to AI, who still decides?

01Problem Framing

Gradual disempowerment happens one convenient decision at a time: a model recommends, a team defers, and over months the organization's own judgment — and its leaders' control over direction — quietly atrophies. Staff who once made a judgment call now approve one; the muscle for the call itself, not just its output, fades within a few quarters.

Without a named human owner, a model's recommendation becomes the default outcome — authority shifts without anyone deciding to shift it. And because tools roll out in weeks while the policies to govern them take quarters, the gap between adoption and oversight is exactly where control erodes fastest.

What Counts as Disempowerment

Disempowerment is AI voluntarily given control over:

• Which options are proposed
• Which resources are made available
• Which alternatives in trade-offs are given higher priority
• Which stakeholders are emphasized
• Which issues within proposed policies are highlighted

How It Happens

1. 01AI assists with tasks
2. 02AI becomes the default starting point
3. 03Humans increasingly review rather than originate
4. 04AI shapes which choices and evidence are considered
5. 05Independent skills and institutional knowledge weaken
6. 06Human approval remains, but meaningful human control declines

Left unchecked: workflow replacement, a subtler erosion of human control, and compounding misalignment.

Is This Already Happening?

Federal AI spending shows how fast adoption is outpacing governance (Brookings research, 2026):

Governance Is Catching Up

States are moving in parallel with a growing wave of AI safety bills. At least 21 county and city governments have already published their own AI governance policies:

Case Studies

02Why It Matters

To Individuals

You are the average of the five people you spend the most time with — what if those "people" are AI models? How often have you already turned to an LLM for a purchase decision, a relationship question, a job conflict, or a car or house purchase? Have you thought about how AI models are subtly changing you?

No LLM is truly neutral. Its views are shaped by training data and by reinforcement learning from human feedback — and no group of people providing that feedback is free of bias. These biases can subtly influence belief and opinion. (ACM Digital Library)

Governments are increasingly using AI to increase efficiency in domains like surveillance, housing, and criminal justice. A meaningful human-in-the-loop component is essential to fairness, accuracy, and public trust. AI use is inevitable and its benefits are real — governance is what makes its use responsible.

To Everyone

Many of us believe we remain fully in control when using AI — that we're simply using it as a tool while we make the final decisions. But the more we outsource evaluating options, forming judgments, or framing problems, the easier it becomes to accept AI's suggestions and framing without much critical reflection. The shift is gradual and hard to notice.

AI doesn't just provide answers — it shapes how information is framed, which options are presented, and what looks reasonable or important. As AI becomes embedded in daily life, these influences accumulate across individual decisions, education, workplaces, public discourse, and policy.

This is why AI governance matters to everyone, not just policymakers or technology companies. It isn't about restricting AI — it's about transparency, accountability, and preserving meaningful human agency.

03Gap in Current Policy

Existing U.S. policy tracks whether AI was used and whether a human signed off — and stops there. That standard is satisfied even when a human's only contribution was approving an idea that was AI-originated from the start. NIST's AI Risk Management Framework, the leading framework, lets organizations adopt as many or as few of its recommendations as they see fit, leaving this gap open at every organization's discretion. Signing off is not the same as deciding.