AI evidence extraction

Detection, Disclosure, and Verification: A Repeated Game of Generative AI and Scientific Credibility

Xufeng Zhang · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

0/2

Quality (LMQS)

Evidence

Citations

0.00

FWCI

This is an AI analysis. Read the peer-reviewed original at the publisher: https://doi.org/10.22541/au.177499312.21860755/v1

Methodology & findings

Study design

Formal game-theoretic analysis using infinite-horizon repeated game between a generator and detector.

Main result

The paper's central finding is that "stronger detection need not induce more truth-directed effort. Under natural assumptions, it induces more evasion effort instead." The authors prove formally that "in a hidden-use equilibrium" stylometric screening increases evasion effort but leaves verification effort unchanged, as demonstrated in Proposition 1 where "∂v H ∂d = 0, ∂e H ∂d = R η > 0, ∂a H ∂d = - R ν < 0." The key implication is that "Screening intensity d increases evasion effort one-for-one in marginal terms, and reduces equilibrium AI reliance, but it does not increase verification effort. Hidden-use governance therefore shifts generator effort toward laundering, not checking."

Research paradigm

formal theoretical analysis with game-theoretic modeling

Author conclusions

The authors conclude that "institutions should regulate under-verified AI use, not AI use per se." They state: "the governance problem is not simply one of identifying AI text. It is one of designing an incentive-compatible interface between AI-enabled generation and human/institutional responsibility." Furthermore, "As generation costs fall, governance should move away from origin myths and toward accountable process design. The future of scientific credibility will depend less on proving that a text was human all along than on ensuring that someone human is visibly responsible for why it should be trusted." Regarding disclosure equilibrium, they conclude that "truthful disclosure is sustainable when the one-period gain from concealment is smaller than the expected immediate penalty plus the discounted loss from falling into a harsher enforcement regime."

Limitations

The authors note that "This separability is appropriate when AI is deployed for drafting or synthesis, so that AI reliance and verification are technologically independent inputs
If AI tools also directly assist checking—for example, through automated code verification or citation validation—then λ in ρ(a, v) would be partially offset by AI-augmented verification, potentially making v H responsive to a and hence indirectly to d
Extending the model to allow such complementarity is left for future work." Additionally, the model assumes "interior parameter values so that ρ(a, v) ∈ (0, 1) on the equilibrium path" and the linear detection function is "chosen for tractability
the qualitative results only require monotonicity in the same directions."

Open questions raised

The authors identify the need to extend the model to account for AI tools that directly assist checking through automated code verification or citation validation. They state this extension is "left for future work." They also acknowledge that "the object of governance is not simply claim validity or authorship purity; it is the joint allocation of provenance, responsibility, and verification," suggesting further work is needed on this integrated governance framework.
The authors identify that extending the model to allow for AI tools that directly assist checking (automated code verification, citation validation) is left for future work. They also note that while watermarking and provenance credentials show promise, the relationship between technological governance and institutional incentive design requires further investigation.
The authors identify the need to extend the model to settings where AI tools directly assist checking (e.g., automated code verification or citation validation). They also note the need for future work on how verification and AI assistance may have complementary technological properties.

Extracted from: pdfAgreement 68%

Explore related topics

Hallucination Control Research Integrity Conceptual