On September 8, 2026, Jacob Coxon resigned from Anthropic and made his reasons public. He had worked in research at both OpenAI and Anthropic, and in his statement he warned that the industry is racing toward self-improving superintelligence without adequate safeguards in place.
Those are his words and his judgment, reported here as what he said rather than as a settled conclusion about the future. Reasonable people inside these companies disagree about timelines and about risk. What is not in dispute is that the pace of frontier development is being set by a small number of organizations, and that most of us are not in the room.
That is an uncomfortable position for a leader. It is also not the whole picture. The question I keep returning to with founders, CEOs, nonprofit directors, and mid-sized organizations is narrower and far more useful: what can we decide about AI inside our own environment?
Agency Is Local
Agency, in this context, means deciding where, when, and under what conditions AI operates in your organization. Which workflows it touches. What information it can see. Which recommendations carry weight. Who reviews them. Who can stop the system.
Human-first AI is what you do with that agency. It is the deliberate use of these choices to strengthen human capability, judgment, dignity, and responsibility instead of quietly eroding them. It is not a posture of refusal. It is a posture of authorship.
Ethical Microcosms
An ethical microcosm is a plain idea with a slightly academic name: a local organization that chooses responsible standards on its own, before, and often in the absence of, industry agreement. It does not require a treaty, a regulator, or a frontier lab to change course. It requires a leadership team willing to say what is acceptable here.
These environments matter more than they appear to. Standards tend to travel through practice long before they travel through policy. Employees carry them to their next role. Customers learn what to expect. Vendors adjust to what buyers actually require.
Four Decisions You Can Make This Quarter
Decision 01
Define AI decision rights
Write down which decisions AI may inform, which it may draft, and which must remain under accountable human authority. Consequential decisions about employment, pay, access, safety, credit, care, or eligibility belong to a named person who can explain the reasoning and reverse the outcome.
Decision 02
Control data access and permissions
Decide what the system may read, how long it retains anything, how quality is checked, and what consent was actually given. Access should be granted deliberately and narrowly rather than inherited from whatever a vendor default allows.
Decision 03
Refuse invasive surveillance and involve employees
There are honest reasons to measure work, and there is a line past which measurement becomes surveillance. Involve the people doing the work in selection, testing, and redesign. They know where the exceptions live, and their participation makes the system more accurate as well as more humane.
Decision 04
Audit for bias, drift, and unintended effects
Systems change as data and usage change. Review outputs on a schedule, look for patterns that disadvantage particular groups, watch for quiet degradation, and assign a named human owner for every system in production.
The Tensions Are Real
None of this is free. Speed and governance pull against each other, and a review step that protects people also slows a launch. Vendor platforms deliver capability quickly while making internal control harder to keep. Cost reduction can be genuinely necessary in the short term and still hollow out human capability over the longer one.
Leaders do not resolve those tensions by pretending they are absent. They resolve them by naming the tradeoff out loud, deciding on purpose, and revisiting the decision when the evidence changes.
What We Actually Control
We may have no control over the larger AI system taking shape around us. We have agency over the environments we create.
That agency is exercised in ordinary places: a permission setting, a review threshold, a conversation with the team whose work is about to change, a decision to say no to a tool that would be convenient and corrosive.
Every organization that uses AI with judgment, dignity and care becomes evidence that another way forward is possible.
If you want a structured starting point, the Human-First AI Readiness Scorecard takes four to five minutes and shows where your organization is ready and where judgment needs protection. You may also find our earlier piece on what AI actually scales a useful companion.
Sources and Interpretation
The account of Jacob Coxon's September 8, 2026 resignation and his public statement is reported as fact. The framing of agency, ethical microcosms, and the four decisions above is Sacred Algorithms' interpretation, not a claim made by any company named here.
- Jacob Coxon, public resignation statement, September 8, 2026 — primary statement (opens in a new tab)
- Anthropic risk documentation — Responsible Scaling Policy (opens in a new tab)
- OpenAI safety materials — OpenAI Safety (opens in a new tab)
Frequently Asked Questions
What does agency over AI mean inside an organization?
Agency means deciding where, when, and under what conditions AI operates in your environment: which decisions it touches, what data it can reach, who reviews it, and who can pause it.
What is an ethical microcosm?
An ethical microcosm is a local organization that deliberately chooses responsible standards for AI even when the wider industry has not agreed on them. It is a small, real environment governed with care.
Do smaller organizations really have influence over AI?
They have limited influence over the frontier race and considerable influence over their own environment. Every organization sets the terms for how AI affects its people, customers, and decisions.
What should a leader decide first?
Start with decision rights: name the decisions that must remain under accountable human authority, and name the person responsible for each AI system in use.