Mecha-nudges for Machines

Giulio Frey; Kawin Ethayarajh

arxiv: 2603.23433 · v2 · pith:LUAIHHRSnew · submitted 2026-03-24 · 💻 cs.AI

Mecha-nudges for Machines

Giulio Frey , Kawin Ethayarajh This is my paper

Pith reviewed 2026-05-19 17:51 UTC · model grok-4.3

classification 💻 cs.AI

keywords mecha-nudgingAI agentsmachine-usable informationEtsy listingsBayesian persuasionV-usable informationonline marketplacesChatGPT release

0 comments

The pith

Sellers on Etsy began including more information useful to AI agents after ChatGPT launched.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper examines how online marketplaces are adapting to the rise of AI decision-makers by introducing mecha-nudging as presentation changes that influence AI agents without harming human users. It combines Bayesian persuasion and V-usable information to measure these shifts in a common unit of bits. Analysis of over six million Etsy listings reveals a 0.143-bit increase in machine-usable information for predicting agent curation decisions after ChatGPT's release. This increase is robust across models and absent in placebo or human-usable measures, and exceeds effects from generic LLM rewriting. The work provides large-scale evidence that such systematic adjustments are already occurring unnoticed.

Core claim

After ChatGPT's release, Etsy listings contain significantly more machine-usable information for predicting agent curation decisions, increasing by 0.143 bits out of a maximum possible increase of 0.355. This shift is robust across prompts, token choices, labeling models, and fine-tuning architectures; absent in a regulated-text placebo; and far larger than the effect of generic LLM rewriting. In contrast, a human study finds little to no change in human-usable information.

What carries the argument

The combined framework of Bayesian persuasion and V-usable information that quantifies environmental changes in bits to detect mecha-nudging across contexts and models.

If this is right

AI agents making curation decisions will encounter listings optimized for their information processing.
Human users continue to see decision environments with little added usable information.
Similar mecha-nudging could emerge on other platforms where AI agents select or rank options.
Marketplaces may face pressure to distinguish AI-targeted changes from standard improvements.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Platforms could develop tools to detect and label mecha-nudges for transparency.
This pattern may extend to search results or recommendation systems beyond product listings.
Long-term, it raises questions about whether markets will evolve separate presentation layers for AI versus humans.

Load-bearing premise

The observed increase in machine-usable information after ChatGPT's release stems from deliberate adjustments to influence AI agents rather than unrelated trends in listing quality or platform changes.

What would settle it

Repeating the analysis on listings from a similar platform with no AI-agent exposure or on pre-ChatGPT data showing no comparable increase would challenge the link to mecha-nudging.

read the original abstract

AI agents are becoming active decision-makers on the Internet. As they make decisions in the same environments as humans, the environments themselves can change to influence them. We call this $\textit{mecha-nudging}$: changes to how choices are presented that systematically influence AI agents without materially degrading the decision environment for humans. To measure this phenomenon, we combine two frameworks -- Bayesian persuasion from economics and $\mathcal{V}$-usable information from computer science -- to get a common unit (bits) for quantifying how environments change across a wide range of interventions, contexts, and models. We apply this framework to over six million Etsy listings and find that, after ChatGPT's release, listings contain significantly more machine-usable information for predicting agent curation decisions, increasing by 0.143 bits out of a maximum possible increase of 0.355. This shift is robust across prompts, token choices, labeling models, and fine-tuning architectures; absent in a regulated-text placebo; and far larger than the effect of generic LLM rewriting. In contrast, a human study finds little to no change in human-usable information. Our results provide the first large-scale evidence that systematic mecha-nudging is already occurring in the wild, but going unnoticed.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper measures a 0.143-bit post-ChatGPT rise in machine-usable information on Etsy but the link to deliberate AI-targeted nudging rests on timing and indirect controls that leave room for other explanations.

read the letter

The main thing to know is that this paper finds listings on Etsy became more informative for AI curation agents after ChatGPT launched, with a 0.143-bit increase in V-usable information out of a possible 0.355 bits. They contrast this with little change for humans and show the shift is bigger than generic LLM rewriting and absent in a regulated-text placebo. That observational pattern is the concrete contribution. They combine Bayesian persuasion and V-usable information to get a shared bits scale, which is a practical move for comparing interventions across models and contexts. The scale of six million listings plus checks across prompts, tokenizers, labeling models, and fine-tuning setups gives the result some weight. The human contrast is useful too. The soft spot is the causal story. The increase could reflect broader seller efforts to improve visibility or SEO around the same time rather than specific attempts to influence AI agents. The placebo and robustness checks narrow some alternatives but do not directly test seller adoption of AI tools or isolate intent from concurrent platform or quality trends. Details on data exclusion and exact significance tests are not in the abstract, so the precision of the 0.143 figure is hard to judge without the full methods. This is for readers working on AI safety, platform regulation, or how digital markets adapt to new agents. Someone studying information measures in real environments would get value from the framework and data. It deserves peer review because the empirical scale and cross-framework application are substantial enough for referees to evaluate and tighten the attribution claims.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces 'mecha-nudging' as systematic changes to choice environments that influence AI agents more than humans. It combines Bayesian persuasion and V-usable information frameworks to quantify changes in bits, then applies the measure to over six million Etsy listings. The central empirical finding is that after ChatGPT's release, machine-usable information for predicting agent curation decisions rose by 0.143 bits (out of a reported maximum possible increase of 0.355 bits). The increase is reported as robust across prompts, tokenizers, labeling models, and fine-tuning architectures; absent in a regulated-text placebo; larger than generic LLM rewriting; and unaccompanied by comparable changes in human-usable information.

Significance. If the attribution to deliberate mecha-nudging holds, the result supplies the first large-scale empirical evidence that online marketplaces are already adapting listings to AI agents. The cross-framework use of bits as a common unit and the scale of the Etsy corpus are clear strengths. The multiple robustness checks and the contrast with both a placebo and human judgments further support the contribution, provided the missing statistical and derivation details are supplied.

major comments (2)

[Abstract and empirical results] Abstract and results reporting the 0.143-bit increase: the maximum possible increase of 0.355 bits is stated without derivation details, data-exclusion rules for the six-million-listing corpus, or statistical significance tests. These elements are load-bearing for the claim of a 'significant' and 'robust' shift.
[Controls and attribution discussion] Controls and attribution discussion: timing around November 2022 plus the regulated-text placebo do not yet isolate AI-specific seller intent from concurrent non-AI trends in listing quality, SEO, or platform algorithms that could also raise V-usable information for the tested models.

minor comments (2)

[Framework section] The notation and precise definition of V-usable information should include an explicit equation or direct citation to the source framework so readers can follow the bit calculations without external lookup.
[Results tables/figures] Any table or figure presenting the 0.143-bit figure should report confidence intervals or p-values to substantiate the 'significantly more' language.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their thoughtful and constructive comments. We address each major comment below, indicating revisions made to the manuscript.

read point-by-point responses

Referee: [Abstract and empirical results] Abstract and results reporting the 0.143-bit increase: the maximum possible increase of 0.355 bits is stated without derivation details, data-exclusion rules for the six-million-listing corpus, or statistical significance tests. These elements are load-bearing for the claim of a 'significant' and 'robust' shift.

Authors: We agree these details are essential. The revised manuscript now includes a full derivation of the 0.355-bit maximum in new Appendix A, computed from the upper bound of V-usable information under the Bayesian persuasion setup. We specify the data-exclusion rules: duplicate listings, non-English text, entries with fewer than 10 words, and those missing required fields were removed to arrive at the final corpus. We have added bootstrap-based statistical tests showing the 0.143-bit increase is significant (p < 0.001) with 95% CI [0.119, 0.168]. revision: yes
Referee: [Controls and attribution discussion] Controls and attribution discussion: timing around November 2022 plus the regulated-text placebo do not yet isolate AI-specific seller intent from concurrent non-AI trends in listing quality, SEO, or platform algorithms that could also raise V-usable information for the tested models.

Authors: We acknowledge the challenge of isolating seller intent in observational data. The revised discussion section now explicitly addresses concurrent trends, noting that general SEO or quality improvements would be expected to increase human-usable information as well, which is not observed. The regulated-text placebo (no increase) and the contrast with generic LLM rewriting further support the AI-specific interpretation. While these checks strengthen attribution, we note that definitive proof of deliberate mecha-nudging would require seller-level data beyond the current study. revision: partial

Circularity Check

0 steps flagged

No circularity: empirical measurement on external data using pre-existing frameworks

full rationale

The paper's central result is an empirical difference-in-differences style measurement of V-usable information (in bits) on over six million real Etsy listings before versus after ChatGPT's release. It combines two independently established frameworks—Bayesian persuasion and V-usable information—without deriving either from the present data or from author-specific fitted parameters. The reported 0.143-bit increase (out of a 0.355-bit maximum) is computed directly from the external dataset; robustness checks across prompts, tokenizers, models, and a regulated-text placebo are statistical controls, not definitional reductions. No load-bearing step reduces by construction to a self-citation, an author-defined ansatz, or a fitted input renamed as a prediction. The derivation chain is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The abstract relies on the assumption that Bayesian persuasion and V-usable information can be combined into a shared bit-scale metric without additional free parameters beyond those internal to the cited frameworks. No new physical entities are postulated. The mecha-nudging concept is a definitional framing rather than an invented causal mechanism with independent evidence.

axioms (1)

domain assumption Bayesian persuasion and V-usable information can be combined to produce a common unit (bits) for quantifying how environments change across interventions, contexts, and models.
Invoked in the abstract to justify the measurement framework applied to the Etsy data.

invented entities (1)

mecha-nudging no independent evidence
purpose: Label for changes to choice presentation that systematically influence AI agents without materially degrading the environment for humans.
New term introduced to frame the observed phenomenon; no independent falsifiable prediction or evidence supplied beyond the Etsy measurement itself.

pith-pipeline@v0.9.0 · 5738 in / 1578 out tokens · 75313 ms · 2026-05-19T17:51:11.076908+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We combine the Bayesian persuasion framework with V-usable information... arg max IM(τ(X)→YM) s.t. IH(τ(X)→YH) ≥ IH(X→YH)−ϵ
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

V-usable information... observer-relative generalization of Shannon information

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.