AI Cognitive Sovereignty

The Cognitive Sovereignty stack: six tiles top-to-bottom — Model, Data, Training signal, Values, Override, Exit. Bottom two (Override, Exit) ember-tinted as the early-warning layers. Right-edge gutter: a three-state column (Held, Contested, Ceded) per layer. The Six-Layer Audit fills in this column per layer.

 

Model. Which underlying model the proxy runs on, who provides it, what happens when the model is deprecated, repriced or updated. Default failure: the model changes overnight, proxy behaviour shifts, and a working artefact is replaced by something that looks similar and decides differently, without the leader noticing until a decision lands wrongly.

​

Data. The corpus the proxy has ingested: the leader’s emails, drafts, decisions, judgements, memos. Where it sits, who has copies, what happens to it under each plausible exit scenario. Default failure: the corpus quietly becomes training signal for someone else’s product, and the leader’s voice is now in a model the leader does not own.

​

Training signal. Where the leader’s corrections, refinements and red-line edits go: into a private weight only the leader’s proxy uses, or into a shared model from which other users benefit. Default failure: the leader trains a competitor’s model for free, every correction sharpening someone else’s product, and never sees the line item.

​

Values. What the proxy refuses to do, softens, escalates, or will not say. If the leader has not set these, someone else has: the vendor’s default safety stack, a content policy written for a different jurisdiction, an alignment team the leader has never met. The proxy then represents the leader on questions where the leader’s actual position would have been different.

​

Override. What stopping the proxy looks like in operating reality, not in user-agreement language. Whether the leader can pull a draft mid-output, contradict an output already sent, retire a class of response. Default failure: an override that exists in the user agreement and not in operating practice, so the leader believes they have control they do not have.

​

Exit. Whether the leader can leave the platform, the firm or the jurisdiction with the corpus, the curriculum and the working proxy intact, and resume work elsewhere within days. Default failure: the proxy that represents the leader becomes a hostage to the venue that hosts it, and the cost of leaving rises faster than the leader’s leverage.

Override and exit are the early-warning layers. If the leader has lost these, the others are usually already lost.

​

Genesis

Reads against Andrew Likierman’s Judgement at Work (Profile Books, 2025), which frames judgement as a learnable process: if judgement is a process, a well-architected proxy can run it, and the sovereignty question is who controls that proxy. Stacks on Don Tapscott and Joseph Bradley’s You to the Power of Two (2026) on Identic AI, and on the data-sovereignty tradition: Daniel Solove (2008) on contextual integrity, Shoshana Zuboff (2019) on surveillance capitalism, Lawrence Lessig (1999) on code-as-law.

​

Why it matters now

Each layer has its own pull toward abdication: opacity on model and data; invisibility on training signal; alignment fatigue on values; the cost of testing on override; contract asymmetry on exit. The first sovereignty incident usually arrives a year after the assumption of sovereignty was made, and by then four or five layers have already moved.

​

The moves

Set five values lines for the proxy as operating refusals; five concrete refusals beat a paragraph of principles.

Test the override this week: try to stop one output after it has been sent, or pull a draft mid-stream; the test surfaces what the override actually does in operating practice. Write the exit memo: one page on what the leader would do with the corpus, the curriculum and the proxy if they left the platform tomorrow, with names and timelines.

A leader who can name the five refusals, has tested the override, and has the exit memo on file holds sovereignty on the layers most often ceded by default.

​

How ChangeSchool applies it with executives

We run senior cohorts through the Six-Layer Audit: each leader walks the six layers in front of a peer, names where their sovereignty sits on each (held, contested, ceded), and commits to one move on values, override and exit before the next session.

​

The discipline

A monthly sovereignty review, thirty minutes walking the six layers with one line per layer on what changed.

An annual sovereignty audit, half a day with agreements, settings and operating tests in hand, catching drifts the monthly review misses.

A standing exit option, a live, tested ability to take the corpus and curriculum elsewhere; it does not need to be exercised, only to exist. The leader who can credibly leave is the leader whose sovereignty on the other five layers holds.

 

 

 

‘If the proxy represents you, someone owns the proxy; the discipline of Cognitive Sovereignty is whether that someone is you.’

 

 

Viren Lall, Managing Director,

ChangeSchool LDN (2026).

virenlall.com/ai-cognitive-sovereignty

​

​

​