Skip to content

Commit 2aa1b00

Browse files
dcondreydcondrey
committed
add problem statement section and strengthen motivation references (fixes #1, fixes #2)
1 parent 6383841 commit 2aa1b00

File tree

1 file changed

+112
-22
lines changed

1 file changed

+112
-22
lines changed

‎draft-condrey-rats-pop-protocol.xml‎

Lines changed: 112 additions & 22 deletions
Original file line numberDiff line numberDiff line change
@@ -76,32 +76,92 @@
7676
</t>
7777
</section>
7878

79+
<section anchor="problem-statement">
80+
<name>Problem Statement</name>
81+
<t>
82+
Digital documents lack creation-process provenance. Existing
83+
cryptographic mechanisms address complementary but insufficient
84+
aspects of this gap: COSE signatures <xref target="RFC9052"/>
85+
prove key possession, trusted timestamps
86+
<xref target="RFC3161"/> prove that content existed at a given
87+
time, and media provenance standards such as C2PA
88+
<xref target="C2PA"/> track the custody and transformation of
89+
digital assets after creation. None of these mechanisms reveals
90+
how a document was produced or how it evolved during authorship.
91+
</t>
92+
<t>
93+
Non-cryptographic approaches have emerged to address this gap,
94+
but each carries fundamental limitations:
95+
</t>
96+
<dl>
97+
<dt>Surveillance-based methods:</dt>
98+
<dd>Screen recording, keystroke logging, and similar monitoring
99+
capture the authoring process but are inherently
100+
privacy-invasive, require continuous trust in a third-party
101+
archive, and produce evidence that cannot be independently
102+
verified without access to the surveillance
103+
infrastructure.</dd>
104+
<dt>Detection-based methods:</dt>
105+
<dd>Stylometric classifiers and AI-generated content detectors
106+
operate on finished output rather than the creation process.
107+
These tools are probabilistic, degrade as generative models
108+
improve, are vulnerable to adversarial evasion, and exhibit
109+
systematic bias against non-native language speakers
110+
<xref target="Liang2023"/>. Their outputs do not constitute
111+
deterministic, independently reproducible evidence.</dd>
112+
</dl>
113+
<t>
114+
The need for creation-process provenance extends across sectors
115+
and jurisdictions. AI-assisted authoring is entering clinical
116+
documentation workflows, prompting the U.S. Department of
117+
Health and Human Services to seek input on how to maintain
118+
accountability when AI tools contribute to medical records
119+
<xref target="HHS-AI-RFI"/>. The autonomous content generation
120+
capabilities of AI agents have led the National Institute of
121+
Standards and Technology to examine security implications,
122+
including the attribution of agent-produced output
123+
<xref target="NIST-AI-Agents-RFI"/>. In creative industries,
124+
the Authors Guild has launched a human authorship certification
125+
program for books, but absent cryptographic process evidence,
126+
the program relies entirely on author self-attestation
127+
<xref target="AG-Human-Authored"/>. The common requirement across
128+
these sectors is a standardized mechanism to verify how content
129+
was produced rather than inferring it after the fact. Such a
130+
mechanism must satisfy four properties:
131+
</t>
132+
<ol>
133+
<li><em>Privacy-preserving:</em> Evidence must not require
134+
disclosure of document content or invasive behavioral
135+
surveillance. Cryptographic hashes, not raw content, must form
136+
the evidentiary basis.</li>
137+
<li><em>Independently verifiable:</em> Proofs must be
138+
self-contained and deterministically checkable by any Verifier
139+
without access to external archives or proprietary
140+
systems.</li>
141+
<li><em>Tamper-evident:</em> Evidence must form a cryptographic
142+
chain resistant to retroactive modification, reordering, or
143+
selective omission.</li>
144+
<li><em>Process-documenting:</em> The mechanism must capture how
145+
a document evolved over time, not merely what it contains at a
146+
point in time.</li>
147+
</ol>
148+
</section>
149+
79150
<section anchor="use-cases">
80151
<name>Use Cases</name>
81152
<t>
82-
Stakeholders across creative industries, education, law, and
83-
commerce increasingly need to determine how content was produced,
84-
not merely what it contains. Post-hoc detection techniques
85-
(stylometric classifiers, watermark detectors) are demonstrably
86-
insufficient for this purpose: they degrade as generative models
87-
improve, are defeated by paraphrasing, and exhibit systematic
88-
bias against non-native language speakers, with one study
89-
finding that detectors misclassified over 61% of essays by
90-
non-native English writers as AI-generated
91-
<xref target="Liang2023"/>. Institutions relying on these
92-
detectors face significant litigation risk, as courts have begun
153+
The provenance gap identified in
154+
<xref target="problem-statement"/> has concrete consequences
155+
across multiple domains. Where detection-based methods have been
156+
deployed as substitutes for process evidence, the results have
157+
been problematic: Liang et al. found that GPT detectors
158+
misclassified over 61% of essays by non-native English writers
159+
as AI-generated <xref target="Liang2023"/>, and courts have
160+
begun
93161
ruling that sanctions based solely on AI detection scores lack
94-
valid evidentiary basis.
95-
</t>
96-
<t>
97-
PoP addresses this gap through process attestation: capturing
98-
tamper-evident evidence of how content is created, rather than
99-
making probabilistic inferences about finished output. This
100-
approach complements existing media provenance standards such as
101-
C2PA <xref target="C2PA"/>, which tracks the custody and
102-
transformation of digital assets after creation. Where C2PA
103-
attests to what happened to content, PoP attests to how content
104-
came into existence.
162+
valid evidentiary basis. The following use cases illustrate how
163+
process attestation addresses needs that neither content
164+
detection nor custody-based provenance can satisfy.
105165
</t>
106166
<t>
107167
Importantly, PoP does not assert that human-authored content is
@@ -1657,6 +1717,7 @@ jitter-seal = HMAC-SHA-256(seal-key, seal-input)
16571717
</references>
16581718
<references>
16591719
<name>Informative References</name>
1720+
<xi:include href="https://bib.ietf.org/public/rfc/bibxml/reference.RFC.3161.xml"/>
16601721
<xi:include href="https://bib.ietf.org/public/rfc/bibxml/reference.RFC.3552.xml"/>
16611722
<xi:include href="https://bib.ietf.org/public/rfc/bibxml/reference.RFC.6973.xml"/>
16621723
<xi:include href="https://bib.ietf.org/public/rfc/bibxml/reference.RFC.9266.xml"/>
@@ -1764,6 +1825,35 @@ jitter-seal = HMAC-SHA-256(seal-key, seal-input)
17641825
</front>
17651826
<seriesInfo name="Internet-Draft" value="draft-mihalcea-seat-use-cases-01"/>
17661827
</reference>
1828+
<reference anchor="HHS-AI-RFI" target="https://www.regulations.gov/docket/HHS-ONC-2026-0001">
1829+
<front>
1830+
<title>Request for Information: Accelerating the Adoption and Use of Artificial Intelligence as Part of Clinical Care</title>
1831+
<author>
1832+
<organization>U.S. Department of Health and Human Services, Office of the National Coordinator for Health Information Technology</organization>
1833+
</author>
1834+
<date year="2025" month="December"/>
1835+
</front>
1836+
<seriesInfo name="Federal Register" value="2025-23641"/>
1837+
</reference>
1838+
<reference anchor="NIST-AI-Agents-RFI" target="https://www.regulations.gov/docket/NIST-2025-0035">
1839+
<front>
1840+
<title>Request for Information Regarding Security Considerations for Artificial Intelligence Agents</title>
1841+
<author>
1842+
<organization>National Institute of Standards and Technology</organization>
1843+
</author>
1844+
<date year="2026" month="January"/>
1845+
</front>
1846+
<seriesInfo name="Federal Register" value="2026-00206"/>
1847+
</reference>
1848+
<reference anchor="AG-Human-Authored" target="https://authorsguild.org/news/ag-launches-human-authored-certification-to-preserve-authenticity-in-literature/">
1849+
<front>
1850+
<title>Authors Guild Launches "Human Authored" Certification to Preserve Authenticity in Literature</title>
1851+
<author>
1852+
<organization>The Authors Guild</organization>
1853+
</author>
1854+
<date year="2025" month="January"/>
1855+
</front>
1856+
</reference>
17671857
</references>
17681858
</references>
17691859

0 commit comments

Comments
 (0)