The Two Room Thought Experiment

What happens when we try to separate expertise from access? The answer is surprisingly asymmetric: hiding a person's identity is easy. Hiding a company's? Nearly impossible.

01

The Setup

Picture two rooms, two people, no direct line of sight but can talk.

ROOM 1 Sensitive Data ROOM 2 E Expert Abstracted information Instructions

Room 1 has all the sensitive stuff: names, addresses, credit cards, confidential data. But the person in there? Not an expert.

Room 2 has the expert. They give instructions, make decisions, solve problems. But they only see what Room 1 tells them, stripped of identifying details.

So here's the question: for what kinds of tasks can this actually work?

02

Level 1: Simple Task

Ordering Food Works Perfectly

Someone needs dinner ordered. Room 1 has DoorDash open with saved payment and address. Room 2 knows how food delivery apps work.

R1"Logged in. User wants Italian, vegetarian, $30 budget."
R2"Filter Italian, vegetarian. Sort by rating."
R1"12 restaurants. Top one's 4.7 stars."
R2"That one. Look for pasta under $25."
R1"Penne $18, fettuccine $20, ravioli $22."
R2"Ravioli. Checkout with saved payment. 20% tip."
R1"Done. 35 minutes."
Result

The food gets ordered. The expert never learns who they're ordering for, where it's going, or how it's paid for. Task complete, privacy intact.

03

Level 2: Medium Complexity

Customer Support Mostly Works

A customer service expert needs to handle a complaint. Room 1 can see the customer's account, order history, personal info.

R1"Customer's upset. Laptop they ordered 3 weeks ago arrived damaged. They want a refund."
R2"When was it delivered? Any previous returns?"
R1"22 days ago, FedEx. One return 8 months back. Customer for 3 years, about $2K lifetime."
R2"Approve the refund. Send a return label. Good customer, clean history."
R1"Done."
Result

Sound business decision made. The expert saw patterns and history, not names or addresses.

For most cases, this works fine. But consider the edge: a customer who ordered a custom-engraved laptop with accessibility modifications, mentioned they're a professor at a local university, and previously bought textbooks on quantum physics? That combination of details starts to become identifying on its own. These cases are rare, though. Individual anonymization holds up pretty well.

04

Level 3: High Complexity

Investment Evaluation Breaks Down

A VC firm needs to evaluate a startup. Strict NDA. Room 1 has the pitch deck and financials. Room 2 is a senior investor, 25 years in the game.

R1"20 employees, $47M raised, $5M ARR. AI agents for customer support."
R2"CAC to LTV ratio?"
R1"$1,200 CAC, $8,500 LTV."
R2"Interesting. What's their actual differentiation? Technical architecture?"
R1"They claim a novel approach that cuts hallucinations 60% versus standard models."
R2"Is it actually novel? Who are the competitors? IP situation?"
R1"Main competitors are [3-4 well-known AI companies]. 2 patents pending."
R2"Hold on. If those are the competitors, I know exactly what space we're talking about. And there are maybe three companies that fit this profile. I need team backgrounds, go-to-market strategy, the actual tech stack..."
Where It Breaks

The abstraction just collapsed. Every piece of information the expert needs to make a real decision (the technical approach, the competitors, the market segment, the team) is effectively the identity. Even without the company name, Room 2 can figure out who they're evaluating.

05

The Complexity Threshold

A pattern becomes visible across these levels:

Simple Tasks Individual PII Organizational Identity
  • Simple tasks work perfectly. Generic instructions, no specific context needed.
  • Individual PII can usually stay hidden. Edge cases exist, but they're uncommon.
  • Organizational identity fails fast. The details needed to evaluate a company are what identify it.
The Core Asymmetry

A person's name doesn't change how you order their food. But a company's technology, its competitors, its market position? That is the evaluation. You can't separate the two.

06

Theoretical Foundations

Claude Shannon's information theory and decades of privacy research formalize what the thought experiment reveals intuitively.

Shannon's Information Entropy

Shannon formalized a way to measure information content. Entropy quantifies the average "surprise" in a random variable — roughly, how many yes/no questions you'd need to pin down an unknown value. When a combination of attributes is rare, each attribute carries high self-information, and the set becomes identifying.

H(X) = -Σ p(x) log2 p(x)
High entropy = many possible values. Low entropy = few candidates, easy to identify.

Common things have low information content. Rare things become identifying.

k-Anonymity

Privacy researchers use k-anonymity: data is k-anonymous if each record is indistinguishable from at least k-1 others. In plain terms: how many people (or companies) match this description?

k ~ 50,000+
"35-year-old software engineer in San Francisco" Tens of thousands match. Identity well-protected.
k ~ 1-3
"50-person AI startup, $47M raised, 60% hallucination reduction" Maybe two or three companies fit. Identity exposed.

For individuals, k is typically in the thousands or millions. For organizations with specific market positions, k drops to single digits almost immediately.

Differential privacy (Dwork, 2006) takes this further: rather than trying to hide identities after the fact, it mathematically bounds how much any single record can influence a query result. But even differential privacy can't help when the query itself is the identity — which is exactly the organizational case.

07

Real-World Applications

$4.45M
Avg. data breach cost
IBM, 2023
$340B+
Global BPO market
GlobalData, 2023
40%
Rank data privacy as #1 AI concern
Deloitte, 2024
€2B+
GDPR fines in 2023
Enforcement Tracker

How This Plays Out by Domain

Domain Individual PII Organizational Identity
Healthcare Patient processing works Hospital review? Can't hide it
Legal Client comms mostly work M&A due diligence fails
Finance Transaction processing works Investment eval? Impossible
HR Resume screening works Executive search breaks down
08

Implications

The Limits of Anonymization

You can strip a person's name from a document and still get useful work done. Try that with a company, and you've removed the very thing that makes evaluation possible.

What Expertise Actually Is

It's not just following procedures. It's pattern recognition across specific contexts. Simple tasks reduce to generic steps. Complex decisions require context that, for organizations, is inseparable from identity.

AI and Automation

  • Simple tasks automate easily with minimal context
  • Medium complexity requires careful trade-offs between privacy and capability
  • High complexity may resist both anonymization and full automation for the same reason

The Trust Question

For complex tasks, the solution isn't better anonymization. It's better trust frameworks: contractual, technical, and legal structures that enable necessary information sharing with appropriate protections.

Where does the threshold lie in your domain?

References

  1. Shannon, C.E. (1948). "A Mathematical Theory of Communication." Bell System Technical Journal.
  2. Sweeney, L. (2002). "k-Anonymity: A Model for Protecting Privacy." Int'l Journal of Uncertainty, Fuzziness and Knowledge-Based Systems.
  3. Narayanan, A. & Shmatikov, V. (2008). "Robust De-anonymization of Large Sparse Datasets." IEEE S&P.
  4. IBM Security. (2023). "Cost of a Data Breach Report 2023."
  5. Dwork, C. (2006). "Differential Privacy." ICALP.