Effective Interventions Against AI-Enhanced Scams

Kyle Fredrickson

Source-linked AI summary

Effective Interventions Against AI-Enhanced Scams

Kyle Fredrickson

arXiv:2609.00806v1Sep 1, 2026cs.CRcs.CY

TL;DR

Scams impose large and growing losses, while AI increases scam profitability and changes the economics of scam operations. This paper models scam profits under worst-case AI assumptions and finds that reporting interventions can compound to reduce revenue and raise infrastructure costs. The analysis suggests that even modest reporting rates against high-cost infrastructure could substantially affect scam profitability, although effective systems must improve multiple reporting parameters without increasing false reports.

Problem
The paper asks whether reporting scam infrastructure can significantly affect scam profitability as AI makes scams more believable and reduces engagement costs.
Method
The paper develops a simple scam-profit model using communication and financial infrastructure costs under worst-case assumptions about AI.
Results
Reporting rate, reporting centralization, and report accuracy have multiplicative effects that reduce revenue per scam channel while increasing infrastructure costs and demand.
Takeaways & Limitations
Even modest reporting rates against high-cost infrastructure could significantly reduce scam profitability.
Takeaways & Limitations
Effective reporting systems must improve reporting rate, centralization, and accuracy simultaneously, because easier reporting can also increase false reports.

Abstract

from arXiv · show

In 2025, scams were responsible for an estimated $442 billion in direct losses globally. In the United States, reported losses increased by nearly 400% between 2020 and 2025. Though AI in scamming is a relatively new phenomenon, its use significantly changes the economics of scams as well as the bottlenecks in scam operations. In this paper I investigate what interventions will remain effective under this new AI-driven scamming regime. I develop a simple model of scam profits to understand how different interventions asymptotically affect scam operations. I find that three levers--reporting rate, centralization of reporting, and report accuracy--multiply in their effect on expected victims per scam channel, reducing revenue per scam channel while increasing costs. Because effects multiply, interventions affecting all three could have a significant effect on the profitability of the scam business model. My analysis suggests that even modest reporting rates against high-value scam infrastructure could have significant impacts on scam profitability.

1 Introduction

AI changes scam economics by increasing believability and reducing engagement costs, while the paper studies infrastructure interventions that remain effective under worst-case AI assumptions. The model finds that reporting rate, reporting centralization, and report accuracy compound to reduce scam profitability.

Motivation: $442 billion in global direct losses were attributed to scams in 2025, while reported U.S. losses rose nearly 400% from 2020 to 2025.Reported U.S. losses increased from $4.2 billion to $21 billion, and reported losses may undercount actual losses by approximately 7×.
Motivation: AI increases scam profitability by making lures more believable and reducing the marginal cost of engaging with targets.More believable lures are more likely to convert targets into paying victims, while AI reduces engagement costs.
Approach: The paper models scam profits using only communication and financial infrastructure costs, producing a lower bound on scammers’ costs under worst-case AI assumptions.The model excludes other costs and retains only the costs of channels such as phone numbers, social-media accounts, emails, and financial accounts.
Key result: Reporting rate, reporting centralization, and report accuracy multiply their effects on expected victims per channel.Together they reduce revenue per contact channel while increasing scammers’ infrastructure costs and demand for infrastructure.
Key result: Even modest reporting rates against high-cost infrastructure could significantly reduce scam profitability.The paper also derives an upper bound on reporting rates that would make scams unprofitable.

2 Background

Scams depend on communication and financial infrastructure, but the infrastructure is exposed unevenly and relevant intelligence is poorly shared. The paper therefore asks whether reporting scam infrastructure can significantly affect scam profitability.

Scam infrastructure: Scams use communication channels to establish contact, engage targets, and eventually direct victims to financial channels for payment.Examples include SMS lures about missed jury duty and wrong-number messages that lead to investment scams.
Scam infrastructure: Scammers reuse communication and financial channels across thousands of attempts, allowing many people to recognize scam infrastructure before money moves.Low conversion rates mean a channel can be exposed to thousands of people before producing victims.
Scam infrastructure: Communication infrastructure is cheap and disposable, whereas financial infrastructure is harder to acquire and can cost hundreds of dollars.Scammers gate financial infrastructure behind engagement so it is revealed primarily to victims.
Reporting problem: Scammers can reveal financial infrastructure after only a few dozen messages, but the parties holding this intelligence have little incentive to share it.The paper identifies misaligned incentives rather than inaccessible information as the central reporting problem.
Research question: The paper asks whether correcting these incentives and building effective reporting can significantly affect scam profitability.The question focuses on reporting as an intervention against scam infrastructure.

3 Modeling Scam Profits

The model treats each communication or financial account as a scam channel whose lures produce victims or reports before the channel is banned. Under negligible AI-era engagement costs, profitability is determined by expected victims per channel and infrastructure costs.

Revenue model: Each scam channel sends lures that convert into victims or reports, and the channel is banned after reaching a report threshold.The model represents channels as phone numbers, social-media accounts, emails, or financial accounts.
Cost model: The cost model assumes AI reduces lure-sending and victim-engagement costs to zero and retains only channel costs.Actual scamming costs are higher because they also include money laundering, labor, real estate, and other expenses.
Revenue model: Revenue equals average loss per victim multiplied by expected victims per channel and the number of channels.The model writes this as ℓ·E[S]·m, where E[S] is expected victims per channel and m is the number of channels.
Revenue model: For n lures, victims and reports are modeled as independent conversions with probabilities s and r, and banning occurs at t or more reports.The report count follows a Binomial(n, r) distribution, and victims can occur only while reports remain below the threshold.
Revenue model: E[S] = sn · Pr[R < t], so expected victims equal potential victim conversions multiplied by the probability that the channel avoids banning.This expression connects reporting risk directly to expected victims per channel.
Profit maximization: Maximizing profit reduces to maximizing expected victims per channel because channel costs are fixed overhead and marginal lure-engagement costs are assumed negligible.The optimal lure count is the value n* that maximizes expected victims per channel.
Profit maximization: Theorem 1 states that the maximum of expected victims per channel is attained at an approximate optimal lure count n* ≈ λ*(t).The theorem defines E*(s, r, t) as the maximum value of E[S | n].

4 Consequences of Interventions

Interventions reduce victims and revenue per scam channel, forcing scammers to acquire more infrastructure under the paper’s model. Effects across reporting, centralization, thresholds, and conversion can multiply, but false-positive reduction is comparatively weak.

Combined effects: Multiplicative effects across intervention levers can substantially reduce revenue per scam channel.The paper states that modest effects on multiple parameters compound in their total effect on scam operations.
Revenue and infrastructure: A revenue reduction by A× requires scammers to acquire at least A× as many channels to restore profits.Greater demand can also increase the price per channel and infrastructure costs.
Revenue and infrastructure: When channel supply is inelastic, increased infrastructure demand can raise prices and reduce scam revenue by up to A×.With highly elastic supply, scammers may instead acquire more channels and keep revenue near pre-intervention levels.
Reducing conversion: Halving scam conversion rate approximately halves expected victims and requires doubling channels and total lures to restore profits.Unlike reporting-rate interventions, conversion-rate reductions do not change the optimal number of lures per channel.
Increasing reporting: Doubling reporting rate approximately halves expected victims and optimal lures per channel, requiring twice as many channels while leaving total lures unchanged.The figure shows the reporting intervention’s maximum and peak occurring at approximately half the baseline values.
Centralizing reporting: Centralizing reporting acts like increasing the reporting rate by multiplying the effective reporting rate through the centralization factor c.The model uses n*≈λ*(t)/(cr), and interventions increasing centralization affect scams in the same way as reporting-rate increases.
Reducing the ban threshold: Halving the ban threshold approximately halves expected victims and optimal lures, requiring twice as many channels to restore profits.The red curve’s maximum is slightly less than half the baseline because of the e^-δ(t) factor.
False positives and thresholds: Reducing expected false reports is a weak lever because its asymptotic effect enters only as a logarithmic addition.For pfp=0.001 and μ=1, halving false reports changes the exact threshold from 6 to 5 and requires only 1.2× as many channels to restore profits.

5 Discussion

The discussion argues that reporting is the more effective intervention than education, especially when centralized and directed at costly financial infrastructure. Its model suggests multiplicative effects can reduce scam profitability, with modest reporting rates potentially sufficient against high-cost channels.

Intervention effects: Reporting, centralization, and report accuracy combine multiplicatively to reduce revenue per channel while increasing infrastructure costs and demand.The paper identifies these interacting levers as its most important result for intervention design.
Prioritize reporting over education: Education is difficult to target and sustain, and prior evidence found little or no reduction in phishing click-through rates.AI may make scams more believable, further complicating education-based reductions in scam conversion.
Prioritize reporting over education: Reporting is structurally easier because most people can recognize scams; the remaining challenge is truthful reporting to the relevant authority.This motivates improving reporting rather than relying primarily on education.
Focus reporting on financial infrastructure: Reporting disposable communication channels is unlikely to work because phone numbers, social accounts, and emails are cheap to replace.Fresh phone-verified Gmail accounts cost about $1, while US SIM-based account verification costs $0.26.
Focus reporting on financial infrastructure: Financial accounts are more promising targets because they are costly and difficult to acquire, and reporting can increase their required number and price.Scammers may reveal financial accounts and crypto wallets about 70 messages into a conversation.
Increase centralization of reporting: A centralized reporting repository could address decentralized signal-sharing and the 36% of respondents who did not report because they were unsure where to report.The paper treats centralization as a high-leverage intervention across scam infrastructure types.
A bound on scam profitability: For YouTube crypto scams, the model estimates that unprofitability requires reporting by 85% of viewers for a $3 account, 0.06% for a $4,000 account, or 0.34% for a $759 median account.These estimates assume timely action on reports and favor scammers through assumptions that overestimate the required rates.

6 Conclusion

The conclusion presents a simple worst-case model showing that reporting and centralization can offset AI-driven increases in scam conversion. It also identifies a design challenge: improving reporting rate, centralization, and accuracy simultaneously without increasing false reports.

Conclusion: The paper develops a simple model of scam profits under worst-case assumptions about AI-driven changes in scam costs.The model focuses on infrastructure costs rather than attempting to represent every scammer expense.
Conclusion: Increasing reporting and centralization while reducing false reports has multiplicative effects on revenue per channel, infrastructure costs, and infrastructure demand.The conclusion frames this multiplicative effect as an advantage even if AI raises scam conversion rates.
Conclusion: The central implementation challenge is improving reporting rate, centralization, and accuracy simultaneously while rapidly converting reports into actionable intelligence.Making reporting easier can increase both true and false reports, potentially diminishing system efficacy.
Conclusion: A system achieving these goals could provide a realistic path toward making scams unprofitable.The paper presents this as a supported possibility rather than a guaranteed outcome.

A Proof of Theorem 1

This proof characterizes the channel-lure level that maximizes expected victims under a reporting threshold. It uses the binomial-to-Poisson approximation and establishes asymptotic properties of the maximizing function.

A Proof of Theorem 1: The proof defines g(t) through the optimized Poisson expression and uses λ*(t) as its maximizing value.The theorem’s optimization is expressed through this auxiliary function.
A Proof of Theorem 1: Theorem 1 states that the maximum expected victim count is attained at n* ≈ λ*(t).Here n is the number of lures per channel, while λ*(t) is the optimizing Poisson parameter defined by the model.
A Proof of Theorem 1: For large n and small r, the report count R is approximated by a Poisson random variable with mean rn.This approximation converts the binomial reporting process into the Poisson form used in the proof.
A Proof of Theorem 1: The maximizing function satisfies g(t) < t and approaches t asymptotically as the threshold increases.The proof establishes the upper bound directly and derives the limiting ratio g(t)/t → 1.
A Proof of Theorem 1: The proof substitutes the asymptotic bound on g(t) back into the expression for the maximum expected victims.This completes the link between the auxiliary Poisson optimization and the theorem’s expected-victim result.

A.1 Proof of Lemma 2

The lemma proof analyzes how the optimized Poisson objective changes when the reporting threshold increases. It derives a one-step ratio and combines those ratios across thresholds.

A.1 Proof of Lemma 2: Lemma 2 studies the scaling of the Poisson maximum g(t) when the threshold changes from t′ to t.The result is formulated for t′ < t.
A.1 Proof of Lemma 2: The proof introduces A(t) = g(t+1)/g(t) · (1 + 1/t), with A(t) ≥ 1.This ratio captures the normalized change in the optimized value between adjacent thresholds.
A.1 Proof of Lemma 2: It lower-bounds g(t+1) using the optimizer λ*(t) for the preceding threshold.The argument expands the Poisson CDF at t+1 into its value at t plus the probability mass at t.
A.1 Proof of Lemma 2: The proof uses the Poisson probability ratio and the first-order condition at λ*(t) to relate the point mass to the cumulative probability.These identities yield the adjacent-threshold ratio needed for the lemma.
A.1 Proof of Lemma 2: The result for arbitrary t′ < t follows by multiplying the adjacent ratios across all intermediate thresholds.This telescoping product extends the one-step relation to the full threshold interval.

B Centralization

The model centralizes reporting by requiring reports to reach the authority that can ban a scam channel. This makes the effective reporting probability the product of filing and reaching that authority, replacing the original reporting term in the model.

B Centralization: Centralization requires a report to reach the single authority able to ban the scam channel.The paper gives Instagram-managed accounts as an example where Instagram alone can ban the account.
B Centralization: The centralized-reporting probability substitutes for r throughout Theorem 1, including the optimal lure-count calculation.
B Centralization: The resulting maximum expected-victim expression uses cr under the same assumptions as §4.1.

C Proof of Lemma 1

Lemma 1 bounds the probability that an innocent channel receives at least t false reports under a Poisson model. The proof applies Markov’s inequality, optimizes the exponential bound, and solves the resulting threshold condition using a monotone transformation.

C Proof of Lemma 1: Lemma 1 models false reports received by an innocent channel as Rfp ∼ Poisson(µ) and seeks a threshold t with Pr[Rfp ≥ t] ≤ pfp.
C Proof of Lemma 1: The closed-form threshold uses the principal branch W0 of the Lambert W function.
C Proof of Lemma 1: The proof converts the tail event into an exponential event and applies Markov’s inequality with the Poisson moment-generating function.It uses E[e^(θRfp)] = e^(µ(e^θ−1)).
C Proof of Lemma 1: The exponential bound is minimized over θ, yielding θ = ln(t/µ) when t > µ.
C Proof of Lemma 1: Solving the optimized inequality requires e^(-µ(e^µ/t))t ≤ pfp, followed by logarithmic rearrangement.
C Proof of Lemma 1: For K = ln(1/pfp) with K > µ, h(t) = t ln(t/(eµ)) is positive and increasing on t > eµ, so thresholds above t* satisfy the bound.The threshold t* is defined by h(t*) = K − µ.