arxiv: 2604.04906 · v1 · submitted 2026-04-06 · 💰 econ.TH · cs.AI· cs.CY· cs.SI

Recognition: 2 theorem links

· Lean Theorem

How AI Aggregation Affects Knowledge

Asuman Ozdaglar, Daron Acemoglu, James Siderius, Tianyi Lin

Authors on Pith no claims yet

Pith reviewed 2026-05-10 18:56 UTC · model grok-4.3

classification 💰 econ.TH cs.AIcs.CYcs.SI

keywords AI aggregationsocial learningDeGroot modellearning gapbelief updatinglocal vs globalupdate speed

0 comments

The pith

AI aggregation fails to robustly improve social learning when updates are too fast, though local systems can succeed.

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

The paper extends the DeGroot model of belief updating by adding an AI aggregator that trains on the population's beliefs and returns synthesized signals to agents. It tracks a learning gap measuring how far long-run beliefs deviate from an efficient benchmark that uses all available information. The central result identifies a threshold on aggregator speed: above this threshold no positive-measure set of training weights improves learning across broad environments, while below it such weights exist. Local aggregators trained on proximate or topic-specific data reduce the gap in every environment, but a single global aggregator increases it in at least one dimension of the state.

Core claim

The paper claims that a threshold exists in the speed of the AI aggregator's updates such that, when updating is too fast, there is no positive-measure set of training weights that robustly reduces the learning gap across a broad class of environments, whereas such weights do exist when updating is sufficiently slow. Local architectures trained on proximate or topic-specific data robustly improve learning in all environments, so replacing them with a global aggregator worsens learning in at least one state dimension.

What carries the argument

The AI aggregator in the extended DeGroot dynamics, which trains on population beliefs and feeds synthesized signals back to agents, with update speed and training weights as parameters that control the size of the learning gap relative to the efficient benchmark.

If this is right

When the aggregator updates slowly enough, there exist training weights that bring long-run beliefs closer to the efficient benchmark in a broad class of environments.
Local aggregators trained on proximate or topic-specific data reduce the learning gap in every environment.
Switching from specialized local aggregators to one global aggregator increases the learning gap in at least one dimension of the state.
The architecture and speed of AI feedback jointly determine whether social learning converges closer to or farther from the truth.

Where Pith is reading between the lines

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

Platforms using fast global AI models for information synthesis may need to slow update rates or segment by topic to avoid widening belief distortions.
Empirical studies could compare belief convergence in networks with fast versus slow AI interventions to test the threshold prediction.
The result limits how far a single universal AI aggregator can be scaled without careful controls on speed and data locality.

Load-bearing premise

Agents continue to update beliefs according to the extended DeGroot rule using the AI's signals, and a well-defined efficient benchmark exists for measuring the learning gap.

What would settle it

Simulate the extended model under fast update rates across many environments and check whether any training weights produce a positive reduction in the learning gap; if none do, the threshold claim holds.

Figures

Figures reproduced from arXiv: 2604.04906 by Asuman Ozdaglar, Daron Acemoglu, James Siderius, Tianyi Lin.

**Figure 1.** Figure 1: Global aggregator architecture. The two-island model is the simplest network that features within-group reinforcement, crossgroup information flow, and systematic asymmetries in representation in training data. These features are central to the operation of the AI aggregator in practice, where training data often overrepresent some groups and adoption varies across the population. Various qualitative prop… view at source ↗

**Figure 2.** Figure 2: Local aggregator architecture. information about the other topic θj ′ with j ′ ̸= j. The assumption is not that only one island cares about a topic, but that first-hand signals and specialized expertise are concentrated locally (e.g., local health systems vs. local industries/labor markets), making initial information topic-specific. We normalize initial beliefs so that agents place zero belief on topics a… view at source ↗

read the original abstract

Artificial intelligence (AI) changes social learning when aggregated outputs become training data for future predictions. To study this, we extend the DeGroot model by introducing an AI aggregator that trains on population beliefs and feeds synthesized signals back to agents. We define the learning gap as the deviation of long-run beliefs from the efficient benchmark, allowing us to capture how AI aggregation affects learning. Our main result identifies a threshold in the speed of updating: when the aggregator updates too quickly, there is no positive-measure set of training weights that robustly improves learning across a broad class of environments, whereas such weights exist when updating is sufficiently slow. We then compare global and local architectures. Local aggregators trained on proximate or topic-specific data robustly improve learning in all environments. Consequently, replacing specialized local aggregators with a single global aggregator worsens learning in at least one dimension of the state.

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 shows that fast AI aggregator updates in an extended DeGroot model block positive-measure training weights from robustly improving collective learning, while slow updates allow it and local training beats global.

read the letter

The main takeaway is straightforward. In this extension of the DeGroot model with an AI aggregator, fast updating prevents any positive-measure set of training weights from improving learning across environments. Slower updating allows such weights, and local aggregators trained on nearby data improve outcomes in every case while a global one does not. They lay out the model clearly by adding the aggregator that trains on population beliefs and feeds back signals, then define the learning gap against an efficient benchmark. The threshold on update speed and the local-global comparison follow from restricting training support. These pieces connect without loose ends. The derivations check out once the architecture is fixed. The existence and non-existence results track the dynamics, and the local case is a direct restriction. No internal contradictions appear in the setup or the measure claims. One limitation is that the findings stay tied to the exact definitions of robust improvement and the learning gap. Different choices for signal synthesis or the benchmark could shift the threshold. The paper stays within theory and does not test empirical fit, which keeps the scope tight but limits how far the design principle travels to real platforms. Readers working on opinion dynamics, social learning, or information economics would get the most from this. It supplies a concrete angle on how aggregation speed and scope affect belief formation. The work deserves peer review. The question is timely and the result is sharp enough to merit checking the details. I would send it out.

Referee Report

0 major / 3 minor

Summary. The paper extends the DeGroot model of social learning by introducing an AI aggregator that trains on population beliefs and feeds synthesized signals back to agents. It defines the learning gap as the long-run deviation from an efficient benchmark and establishes a threshold result on aggregator update speed: when updates are too fast, no positive-measure set of training weights robustly reduces the gap across a broad class of environments, whereas such weights exist for sufficiently slow updates. The analysis further shows that local aggregators trained on proximate or topic-specific data improve learning in all environments, while replacing them with a global aggregator worsens learning in at least one dimension of the state.

Significance. If the central threshold result holds, the paper offers a clean theoretical account of how AI aggregation can either stabilize or destabilize collective beliefs depending on update speed and training scope. The self-contained extension of DeGroot dynamics, with explicit free parameters for training weights and update speed, yields falsifiable comparative-static predictions and avoids ad-hoc axioms. The local-versus-global comparison supplies a concrete mechanism for why specialized aggregators may outperform a single global one, which is directly relevant to the design of AI-mediated information systems.

minor comments (3)

[§2.2] §2.2: the definition of the efficient benchmark (relative to which the learning gap is measured) should explicitly state whether it remains invariant when the aggregator's training support changes; a short remark would remove any residual ambiguity.
[Theorem 1] The proof of the non-existence result for fast updates (Theorem 1) would benefit from a brief remark on whether the positive-measure claim continues to hold if the environment class is restricted to finite-state Markov chains.
[Notation] Notation: the symbol for the aggregator's update-speed parameter is easily confused with the standard DeGroot weight matrix; a distinct symbol or explicit reminder in the text would improve readability.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of our manuscript, the clear summary of our results, and the recommendation for minor revision. We are pleased that the threshold result on aggregator speed and the local-versus-global comparison were viewed as providing a clean theoretical account with falsifiable predictions.

Circularity Check

0 steps flagged

No significant circularity; derivation is self-contained within the extended model

full rationale

The paper extends the classic DeGroot framework by adding an AI aggregator that trains on population beliefs and feeds signals back. It defines the learning gap as long-run belief deviation from an efficient benchmark and derives a threshold result on aggregator update speed for existence/non-existence of positive-measure training weights that reduce the gap uniformly. The local-vs-global comparison follows directly from restricting the aggregator's training support. All steps are explicit mathematical constructions inside the fixed architecture and benchmark; no step reduces by definition or by self-citation to its own inputs. The model is self-contained against its stated assumptions and does not rely on fitted parameters or load-bearing prior results from the same authors.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 1 invented entities

The central claim rests on the mathematical extension of the DeGroot updating rule with an AI training component whose weights and update speed are the key free parameters.

free parameters (2)

training weights
The set of weights the AI uses to synthesize signals from population beliefs; the main result concerns existence of weights that improve learning.
update speed parameter
The speed at which the aggregator retrains on new beliefs; the threshold result is defined with respect to this speed.

axioms (2)

domain assumption Agents update beliefs according to the DeGroot averaging rule
The paper explicitly extends the DeGroot model.
domain assumption The learning gap is measured relative to an efficient benchmark
Defined in the abstract as the deviation from the efficient benchmark.

invented entities (1)

AI aggregator no independent evidence
purpose: Trains on population beliefs and feeds synthesized signals back to agents
New component introduced to study AI effects on social learning

pith-pipeline@v0.9.0 · 5457 in / 1517 out tokens · 49767 ms · 2026-05-10T18:56:00.109142+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
We extend the DeGroot model by introducing an AI aggregator... learning gap as the deviation of long-run beliefs from the efficient benchmark... threshold in the speed of updating... robust improvement set Λ_ρ
IndisputableMonolith/Foundation/RealityFromDistinction.lean reality_from_one_distinction unclear
augmented transition matrix Γ... closed-form consensus via perturbation (Schweitzer 1968)

Reference graph

Works this paper leans on

7 extracted references

[1]

Ifρ < ρ ⋆, thenµ(Λ ρ) = 0
[2]

This completes the proof

Ifρ > ρ ⋆, thenµ(Λ ρ)>0. This completes the proof. A.4 Proofs from Section 5 Proof of Proposition 2.We have ∆⋆ ≡∆ 1(ρ, α, β, β, h, π)−∆ 0(h, π) =| ¯∆1(ρ, α, β, h, π)| −∆ 0(h, π), where ¯∆1(ρ, α, β, h, π) = (1−ρ)(hπ2+π)+α(β(1+β−ρ)h 2π+βhπ2+βh+β(1−β+ρ)π) (1−ρ)(hπ2+2π+h)+(β(1+β−ρ)h 2π+βhπ2+βh+β(1−β+ρ)π) − π π+1 . 36 Fixingρ∈(0,1)andπ >1, we have ∂ ¯∆1 ∂h (ρ,...
[3]

Becauseα < hπ2+π hπ2+2π+h , we have ∂ ¯∆1 ∂β (ρ, α, β, h, π)<0, implying that − ¯∆1(ρ, α, β, h, π)−∆ 0(h, π)as a function ofβis increasing over[0,1]

Indeed, we show whyβ ⋆ 1 exists and is unique. Becauseα < hπ2+π hπ2+2π+h , we have ∂ ¯∆1 ∂β (ρ, α, β, h, π)<0, implying that − ¯∆1(ρ, α, β, h, π)−∆ 0(h, π)as a function ofβis increasing over[0,1]. Fixing anyh >1, we have − ¯∆1(ρ, α,0, h, π)−∆ 0(h, π) = 2π π+1 − 2(hπ2+π) hπ2+2π+h <0. In addition, we have − ¯∆1(ρ, α,1, h, π)−∆ 0(h, π) = 2π π+1 − hπ2+π hπ2+2...
[4]

This implies− ¯∆1(ρ, α, β, hβ, π)−∆ 0(hβ, π)<0

Indeed, ifβ < β ⋆ 1, then by the definition of supremum there exists someh β >1such thatβ < β ⋆ 1(hβ). This implies− ¯∆1(ρ, α, β, hβ, π)−∆ 0(hβ, π)<0. Thus, we haveP(h β)<0, as desired. Lemma A.4.Fixingρ, β∈(0,1),α∈(0, 1 2)andπ >1. For eachh >1, letβ ⋆ 2(h)∈(0,1)denote the (unique) threshold such that− ¯∆1(ρ, α, β, h, π)≤0if and only ifβ∈[β ⋆ 2(h),1]. The...
[5]

Becauseα < hπ2+π hπ2+2π+h , we have ∂ ¯∆1 ∂β (ρ, α, β, h, π)<0, implying that ¯∆1(ρ, α, β, h, π)as a function ofβis decreasing over[0,1]

Indeed, we show whyβ ⋆ 2 exists and is unique. Becauseα < hπ2+π hπ2+2π+h , we have ∂ ¯∆1 ∂β (ρ, α, β, h, π)<0, implying that ¯∆1(ρ, α, β, h, π)as a function ofβis decreasing over[0,1]. Fixing anyh >1, we have ¯∆1(ρ, α,0, h, π) = hπ2+π hπ2+2π+h − π π+1 >0. In addition, we have ¯∆1(ρ, α,1, h, π) = (1−ρ)(hπ2+π)+α((2−ρ)h2π+hπ2+h+ρπ) (1−ρ)(hπ2+2π+h)+((2−ρ)h2π+...
[6]

This implies ¯∆1(ρ, α, β, hβ, π)>0, as desired

Indeed, ifβ < β ⋆ 2, then by the definition of supremum there exists some hβ >1such thatβ < β ⋆ 2(hβ). This implies ¯∆1(ρ, α, β, hβ, π)>0, as desired. Back to the original claim of Proposition 3. We setβ ⋆ = min{β ⋆ 1, β⋆ 2, π−1 2π−2α(π+1) } ∈(0,1). By 40 Lemma A.3, we have that there exists1< ¯h1 < ¯h1 <∞such that − ¯∆1(ρ, α, β, h, π)−∆ 0(h, π) ( ≥0,if1≤...
[7]

For the former case, we have ∆1(ρ, α, β, h, π)−∆ 0(h, π) = ¯∆1(ρ, α, β, h, π)−∆ 0(h, π)<max n α, hπ2+π hπ2+2π+h o − hπ2+π hπ2+2π+h = 0

Otherwise, we consider: ¯∆1(ρ, α, β, h, π)≥0or ¯∆1(ρ, α, β, h, π)<0. For the former case, we have ∆1(ρ, α, β, h, π)−∆ 0(h, π) = ¯∆1(ρ, α, β, h, π)−∆ 0(h, π)<max n α, hπ2+π hπ2+2π+h o − hπ2+π hπ2+2π+h = 0. For the latter case, we have ∆1(ρ, α, β, h, π)−∆ 0(h, π) =− ¯∆1(ρ, α, β, h, π)−∆ 0(h, π)<0. Putting these pieces together yields ∆⋆ ≡∆ 1(ρ, α, β, h, π)−...