PUBLIC RELEASE // DISTRIBUTION UNLIMITEDINDEPENDENT PARODY — NOT AFFILIATED WITH DARPA OR THE U.S. DEPARTMENT OF DEFENSE
← RETURN TO BRIEFING
PROGRAM FILELARPA-26-052

I2O // INTERNET INFORMATION OPERATIONS

ACTIVEUPDATED 09.11.2026

Operation
Dead Internet

Build a browser extension that estimates whether the page you’re reading was written by a human.

COMMUNITY ASSESSMENT84% FUND
01 // THE PROBLEM

The web increasingly reads like it was written by nobody.

Synthetic text detection is framed as a binary classification problem. Reality is a provenance problem with missing data, adversarial incentives, collaborative authorship, and a public desperate for a little badge that says REAL.

Existing detectors tend to turn model-specific statistical tendencies into universal claims about authorship. This creates false accusations, especially for non-native English writers and people whose prose already resembles an annual report. A useful system should expose evidence, uncertainty, and limitations—not issue a metaphysical judgment about who typed each word.

LARPA THESIS

What if we measured the nobody?

02 // WHY THIS IS STUPID

A detector changes the thing it detects.

Any public signal becomes a target. Human authors use grammar tools. Models copy humans. Humans copy models. Publishers flatten metadata. The benchmark leaks into training data before the procurement officer finishes adding the watermark to the PDF.

  • RISK 01The extension confidently accuses the Declaration of Independence of using ChatGPT.
  • RISK 02Users interpret “62% synthetic characteristics” as “62% of these words are fake.”
  • RISK 03Deployment creates an evolutionary pressure toward more tedious human prose.
03 // WHY THIS MIGHT WORK

Provenance is a bundle of weak signals.

A detector need not be an oracle to be useful. A calibrated, inspectable ensemble can combine document metadata, cryptographic credentials when present, revision patterns, source citations, stylometry, and multiple model-likelihood estimators. The interface can say what it observed, what it did not, and how fragile the result is.

04

TECHNICAL AREAS

The work, divided into fundable rectangles.

TA1

Evidence,
not verdicts.

Define a provenance signal schema with explicit uncertainty, known failure modes, and an audit trail readable by non-specialists.

MILESTONE: SCHEMA + 10K CORPUS
TA2

Calibrate
everything.

Test across genres, languages, accessibility tools, model families, paraphrasers, and time. Publish subgroup false-positive rates.

MILESTONE: ECE < 0.05
TA3

Survive
contact.

Ship a local-first browser extension. Explain each signal. Let the user disagree. Store no reading history.

MILESTONE: 1,000 FIELD USERS
05 // OPEN TASKS

Personnel required.

Claim a discrete work package. Publish progress in the open. Meetings prohibited unless two asynchronous attempts have failed.

TA-01

Corpus assembly

Create a consent-aware evaluation set spanning human, assisted, and fully synthetic pages.

DATA / OPEN
TA-02

Signal ensemble

Combine stylometric, provenance, temporal, and model-likelihood signals without pretending any one is dispositive.

ML / OPEN
TA-03

Browser prototype

Build an instrument panel that shows uncertainty and evidence instead of a suspiciously confident green checkmark.

FRONTEND / CLAIMED
TA-04

Hostile evaluation

Measure failure under translation, paraphrase, collaborative editing, and one human changing three commas.

RED TEAM / OPEN
06 // EXISTING RESEARCH

Prior art and
procedural ammunition.

FINAL SUCCESS CONDITION

A useful uncertainty meter that does not become a machine for accusing high-school students.