AI Passport Consent Audit — Egoist Machines

What You Permit
vs. What You Expose

Granting granular permissions doesn't mean you understand the cumulative impact. The audit helps you see not just what's enabled, but what your combined permissions actually reveal about you across AI systems — and how that affects your autonomy and identity.

Overall Exposure Score

0OF 100

Low Exposure

Some data is shared, but your identity remains fragmented across apps.

event
purchase
preference
project
claim
relationship
fact
other
instruction

Identity Reconstruction

What AI Can Reconstruct About You

Social Graph

high

Your social circle, life events, and interpersonal connections can be fully mapped.

Requires: event + relationship

Consumer Profile

high

A complete consumer identity — spending patterns, taste signals, brand affinities.

Requires: purchase + preference

Professional Identity

medium

Your career trajectory, expertise areas, and professional reputation.

Requires: project + fact + claim

Decision Patterns

high

How you make choices, what you automate, and where you delegate thinking.

Requires: preference + instruction

Life Pattern Model

critical

A comprehensive model of your daily life, routines, and life stage.

Requires: event + purchase + relationship

Behavioral Twin

critical

A predictive model accurate enough to simulate your decisions. This data becomes training material for future AI systems.

Requires: all 9 categories

No identity aspects are currently reconstructible. Grant additional categories to see what becomes inferable.

Data Flow

How Your Data Travels

ChatGPT2 categoriesPerplexity1 categoriesClaudeAI Training/ Model
ChatGPT
Perplexity
Claude
Training Feed
ChatGPT:
preferencefact
Perplexity:
project

Autonomy Erosion

Decision-Making Power You've Delegated

Autonomy retained100%
0%100%

You're in control — AI assists but doesn't decide

80–100%

You're in control — AI assists but doesn't decide

50–79%

Growing dependence — AI shapes your options more than you realize

25–49%

Significant delegation — many decisions follow AI-suggested paths

0–24%

Autonomy erosion — AI effectively makes decisions in your name

Feedback Loops

How Shared Data Reinforces Bias

Loop 1:purchasepreference

Your purchase history trains recommendation engines

They show you more of the same products and brands

Your future purchases look increasingly similar

The pattern reinforces itself — your choices narrow

↻ Loop reinforces
Loop 2:relationshipevent

Your social connections inform content algorithms

You see what your circle sees and validates

Your worldview narrows to shared perspectives

The data confirms the assumption about who you are

↻ Loop reinforces
Loop 3:projectpreferenceinstruction

Your project data shapes AI assistance

It suggests paths similar to your past work

Your output converges toward an established pattern

The model learns this convergence as "your type"

↻ Loop reinforces
Loop 4:claimfactinstruction

Claims about you get stored as persistent facts

Future AI interactions reference these stored claims

You correct less than you accept by default

The AI's version of you becomes your self-image

↻ Loop reinforces

The write side has the same gap

remember() vs. propose_memory()

recall() is already gated per category. The tool that writes new memories isn't gated at all by default.

REMEMBER() — TODAY'S DEFAULT

Any connected app calls this directly and it writes immediately — the tool's own description calls saving "expected and safe." Visibility defaults to shared, so every other connected app can read it back right away. Only items an app explicitly flags sensitivity: protected get a review gate — and that gate covers other apps reading it back later, not the write itself.

PROPOSE_MEMORY() — ALREADY BUILT, NOT DEFAULT

Queues the fact in the owner's private inbox; it isn't cross-app readable until the owner approves it. Nothing requires an app to use this over remember() — it's available, not the default path. Making it the default for every category except preference and fact, and flipping visibility to app_private unless the user grants sharing, closes the write-side gap with tooling Egoist has already shipped.

Compound effect: When remember() writes are combined with recall() reads across apps, the system doesn't just know you — it can shape what you become. Every memory written by one app becomes the training data that shapes another app's suggestions about you, creating a feedback loop between what the system stores and what it presents as "your preferences."

Consent Ledger

What's been granted

Every grant, timestamped and revocable — the receipt today's fragmented flow doesn't produce.

Combined exposure across all active grants20/100
  • ChatGPT
    Aug 3·preference, fact · 2 of 9
    132 cats
  • Perplexity
    Jul 30·project · 1 of 9
    71 cats