What does Human-AI Mutual Feeding and the Falsification Ratchet conclude, in numbers, about “The Jagged Frontier: Why Clicking Confirm Loses Accuracy, and Why Better Individuals Make a Narrower Crowd”?

Answer: Chapter 2 treats the jagged frontier: with illustrative parameters, clicking confirm outside the frontier is right 0.604 of the time against 0.800 for answering alone, while falsifying first and then comparing leads on both sides.

Scope of the numbers: Parameters are illustrative and all calculations are generated by the book's programs. Chapter 2 treats the jagged frontier: with illustrative parameters, clicking confirm outside the frontier is right 0.604 of the time against 0.800 for answering alone, while falsifying first and then comparing leads on both sides.

中文版

Data table

The numbers below are printed verbatim by the chapter's Python program and can be reproduced by running it. printed values that match the answer: 0.604, 0.800

Accuracy by policy
regionaloneconfirmfalsify
inside0.8000.9120.954
outside0.8000.6040.863
confirm minus alone, outside: -0.196
Overall accuracy by frontier share s
saloneconfirmfalsify
0.20.8000.6660.881
0.40.8000.7270.899
0.60.8000.7890.918
0.80.8000.8500.936
1.00.8000.9120.954
Confirm beats alone only when s > 0.636
Show the full program
# Chapter 2. p21: "just click confirm" across a jagged frontier.
# A share s of tasks lies inside the model's frontier. Illustrative
# accuracies: model a_in = 0.90 inside, a_out = 0.55 outside; human
# alone h = 0.80 everywhere. Three policies:
#   alone     : h
#   confirm   : take the model's answer; a wrong answer is caught with
#               probability d = 0.15 and then solved alone:
#               a + (1 - a) d h
#   falsify   : human first writes an independent answer with a stated
#               refutation condition; agreement is accepted, a
#               disagreement is investigated and resolved correctly
#               with probability r = 0.90 when one side is right:
#               a h + r [a (1 - h) + (1 - a) h]
# a_out was chosen so that confirm falls about 19 points below alone
# outside the frontier, the size of the drop Dell'Acqua et al. report
# for consultants working outside the frontier with AI.
A_IN, A_OUT, H, D, R = 0.90, 0.55, 0.80, 0.15, 0.90


def confirm(a):
    return a + (1.0 - a) * D * H


def falsify(a):
    return a * H + R * (a * (1.0 - H) + (1.0 - a) * H)


def mix(f, s):
    return s * f(A_IN) + (1.0 - s) * f(A_OUT)


if __name__ == "__main__":
    print("Accuracy by policy")
    print("  region     alone  confirm  falsify")
    for name, a in (("inside", A_IN), ("outside", A_OUT)):
        print("  {:<8}  {:>6.3f}  {:>7.3f}  {:>7.3f}".format(
            name, H, confirm(a), falsify(a)))
    gap = confirm(A_OUT) - H
    print("  confirm minus alone, outside: {:+.3f}".format(gap))
    print("Overall accuracy by frontier share s")
    print("  s      alone  confirm  falsify")
    for s in (0.2, 0.4, 0.6, 0.8, 1.0):
        print("  {:.1f}  {:>7.3f}  {:>7.3f}  {:>7.3f}".format(
            s, H, mix(confirm, s), mix(falsify, s)))
    lo = (H - confirm(A_OUT)) / (confirm(A_IN) - confirm(A_OUT))
    print("Confirm beats alone only when s > {:.3f}".format(lo))

Single source and citation

Fan, Y. (2026). Co-Evolutionary Human-AI Writing Spirals, Falsification-Gated Ratchets and Symbiotic Capability Compounding (Version 1.0.0), Chapter 2: The Jagged Frontier: Why Clicking Confirm Loses Accuracy, and Why Better Individuals Make a Narrower Crowd. 5D Library. DOI pending

DOI pending · monograph page goes live after publication · 5DL-B-374 · v1.0.0 · CC-BY-NC-ND-4.0

BibTeX
@book{fan2026374,
  author    = {Fan, Yafu},
  title     = {Co-Evolutionary Human-AI Writing Spirals, Falsification-Gated Ratchets and Symbiotic Capability Compounding},
  year      = {2026},
  publisher = {5D Library},
  edition   = {1.0.0},
  note      = {DOI pending}
}

Other chapter answers in this book