Sector Briefing 02 · Industry & Enterprise · 5-minute briefing
Fiduciary Liability in the Age of Inference.
You protect your documents. But who protects your drafts? When your lawyers and strategists use AI to 'think,' they are turning your trade secrets into training data.
This briefing makes one argument: the four risks below do not depend on the technology's weaknesses. The exposure runs through ordinary, successful use: the better the assistant, the deeper it sits in your reasoning, and the more of your strategy exists as someone else's data. No breach is required.
Is your strategy session unintentionally leaking confidential information?
The Risks
Risk 01 Holds at any capability
The Fiduciary Gap
Attorney-client privilege protects the final email. It does not protect the 45 minutes the lawyer spent typing queries into a chatbot to draft that email. That thinking process, the questions, the hypotheses, the discarded arguments, is now external property.
Why a better system makes it worse: the more capable the assistant, the deeper it sits in the reasoning chain, and the more work product it holds.
Dive Deeper
The Fiduciary Problem with AI-Assisted Analysis · The CPIA Framework
Risk 02 Holds at any capability
Inferential Leaks
Even if you never type the strategy, behavior reveals it. A system can infer that something is happening just by watching the kinds of questions your executives ask, and when they ask them. The profile assembles itself from patterns nobody considered confidential.
Why a better system makes it worse: inference quality scales with capability. The leak needs no document.
Dive Deeper
Risk 03 Holds at any capability
Judgment Displacement
A recommendation accepted without evaluation is a decision transferred. The transfer is invisible in the record, which shows a human clicking approve. The organization keeps the accountability and loses the judgment.
Why a better system makes it worse: the more reliable the output, the lower the override rate, and the harder it is to say who decided.
Dive Deeper
Risk 04 Holds at any capability
Workforce Dependency
Sustained offloading degrades the unassisted capability of the people doing the work. The organizational version is a workforce that cannot evaluate the outputs it approves, with approval records showing humans in the loop throughout.
Why a better system makes it worse: agreement is what optimized systems get better at.
Dive Deeper
The Era of Cognitive Capture · The Five Mechanisms of Cognitive Capture
49%
How much more often AI systems affirmed users’ positions than human advisors did, across eleven state-of-the-art models. Cheng et al. (2026), Science.
30 of 30
Neurotechnology companies reviewed that retained broad rights over the neural data they collected. Genser, Damianos & Yuste (2024), Neurorights Foundation.
The Leak Path
How work product leaves the building without a breach.
01 The Question
What your people type, ask, revise, and delete
02 The Observation
What the system watches around the typing
Captured03 The Inference
The model of your strategy that patterns assemble
Captured04 External Property
Where that model now lives: infrastructure you do not control
The path runs through ordinary, successful use. Nothing is stolen; it is handed over as behavior.
Required Protocols
01 / Contract
Zero-Training Contracts
Legal mandates that your data, and the metadata of your usage, cannot be used to train base models.
02 / Standard
Ephemeral Processing
Data deleted when the session ends. No long-term memory for corporate AI tools.
03 / Architecture
Isolated Infrastructure
Walled-garden AI: the model comes to your data; your data never leaves your perimeter.
Where This Sector Sits in the CPIA
Every domain applies to the enterprise. Domains 01, 02 and 04 carry the fiduciary weight, and the questions below convert directly into vendor-contract language.
What behavioral signals does the system collect beyond what your people type?
What can be inferred from usage patterns alone, and where do those inferences live?
Does the contract exclude training on your data and on the metadata of your usage?
Is retention ephemeral by default, and can deletion be verified?
Is acceptance-without-modification tracked? Does the organization know its own override rate?
Can unassisted performance still be demonstrated where accountability requires it?
The Instrument · The CPIA
The Cognitive Privacy Impact Assessment converts these risks into procurement criteria: six domains, vendor questions, and the Ephemeral Processing Standard, built to sit alongside the DPIA rather than replace it.
The CPIA Framework · The Fiduciary Problem with AI-Assisted Analysis · The Six Domains of Cognitive Capture · All Research
Cite this briefing
Cook, T. (2026). Fiduciary liability in the age of inference: AI risk to the enterprise. The Cognitive Privacy Project. cognitiveprivacyproject.org/enterprise
For advisory inquiries or research collaboration: timothy@cognitiveprivacyproject.org