Claude's Analysis · Multi-AI Pilot Study Comparison · STAR Framework · Prompt Methodology for Paper Development
David M. Boje, Emeritus Professor, NMSU · 2026
Surveillance pricing — the FTC's preferred term — is the practice of setting an individualized price for each consumer based on data collected about that specific person, rather than on market-wide supply and demand. It is categorically different from ordinary dynamic pricing (where airline seats rise as the flight fills up, equally for everyone). The technical term in economics is first-degree price discrimination: the seller attempts to charge each buyer the maximum they will personally pay.
The data inputs vary by sector but commonly include: device type and operating system, geolocation (GPS-precise or IP-inferred), browsing and purchase history, cookies and cross-site tracking identifiers, estimated household income or credit indicators, loyalty-program data, and fine-grained behavioral signals such as time-on-page, scroll speed, and cursor hesitation. The FTC's 2024 surveillance pricing report explicitly documented that companies have developed commercial intermediary services selling these profiling capabilities to retailers — meaning the data collection infrastructure is itself a market, not just a corporate practice.
Surveillance pricing is best understood as a network phenomenon in Latour's sense: it is not located in one company's algorithm but in the entire heterogeneous assemblage of data brokers, cloud computing infrastructure, retailer dashboards, pricing intermediaries, and consumer devices. To follow the price, you must follow the actor-network.
The mechanism has four stages, each involving distinct human and non-human actants:
1. Data harvesting (inscription): Web trackers, loyalty cards, mobile SDKs, and third-party data brokers convert consumer behavior into machine-readable signals. This is Latour's "inscription device" — the consumer's activity is translated into a durable, portable data trail (an immutable mobile).
2. Willingness-to-pay inference (translation): Machine learning models — typically gradient-boosted trees, neural networks, or regression ensembles — estimate a price sensitivity score or demand elasticity curve for each consumer. The model's architecture and training data encode the CEO's and corporation's values about which signals are predictive and legitimate, embedding the Ghost Vortex Layer 1 and Layer 2 imprints.
3. Price assignment (obligatory passage point): The pricing engine serves a personalized price, promotional offer, or loan rate. The consumer passes through this invisible gate without knowing it exists — it is Callon's OPP: all activity flows through this point, which is controlled by an actor the consumer cannot see or challenge.
4. Arbitrage suppression (quasi-standard): To prevent consumers from comparing notes and buying from each other, the system relies on platform lock-in, login walls, non-transparent pricing displays, and terms-of-service that prohibit automated price comparison. These function as quasi-standards coordinating the network.
The directional claim — "the rich pay more" or "the poor pay more" — is empirically indeterminate in the abstract and must be asked about specific sectors, specific data inputs, and specific algorithmic objectives. This is the central methodological point that distinguishes rigorous from casual analysis:
When wealthier consumers pay more: When the algorithm's objective is revenue maximization and the predictor of high willingness-to-pay is device type, zip-code income, or premium loyalty status, the algorithm rationally charges more to people who appear to be less price-sensitive. This is the "apathetic tax" — you pay the sticker price because you don't hunt for coupons.
When poorer consumers pay more: In markets with low competition (monopoly zones — rural areas, food deserts, poor internet access), the algorithm can charge the available maximum because consumers have no alternatives. Additionally, "digital redlining" operates where zip code or device acts as a proxy for race or income, and the algorithm charges more not because of high WTP but because of reduced price-comparison capacity. Insurance, payday lending, and pharmacy pricing show this pattern.
The distributional harm is asymmetric regardless of direction: A wealthy person paying 15% more for a hotel room loses money but not welfare. A lower-income person paying 15% more for a prescription drug or a mortgage rate faces a genuine welfare loss. The equity problem is not symmetric — it is about the interaction between the price differential and the baseline financial vulnerability of the household.
What the evidence actually supports: The Princeton Review racial pricing allegation; Uber/Lyft neighborhood fare differences; Amazon's neighborhood-correlated pricing variations; and the FTC's own documentation of the commercial surveillance pricing market. None of these establish a universal rule. They establish that the capability exists and that some implementations discriminate. The research agenda is to determine which implementations, in which markets, with what distributional consequences.
Yes, significantly — along three dimensions: regulatory architecture, data infrastructure, and market concentration.
United States: Most permissive environment until recently. Third-party data broker market is large and legally unregulated at the federal level. New York's Algorithmic Pricing Disclosure Act (effective November 10, 2025) is the first state-level mandatory transparency law. FTC investigations ongoing. Price discrimination on protected class grounds is prohibited but zip-code proxies create legal ambiguity.
European Union: GDPR classifies behavioral profiling as personal data processing requiring a lawful basis (usually legitimate interest or consent). The AI Act adds requirements for transparency in high-risk automated decision systems. Personalized pricing is not prohibited, but the data collection enabling it is regulated. Companies must inform consumers when prices are personalized. Crucially, opt-in consent is not universally required — disclosure is the primary mechanism. The EU is more restrictive than the US but less than often claimed.
China: The Personal Information Protection Law (PIPL, 2021) nominally prohibits 大数据杀熟 (algorithmic price discrimination against loyal customers), requiring companies to offer a non-personalized option. In practice, these alternatives are buried in app settings and enforcement is driven by social stability concerns rather than privacy rights. China's super-app ecosystem (WeChat, Alipay) creates the most granular consumer profiling in the world, making PIPL implementation structurally difficult.
Global South: Surveillance pricing manifests differently. In contexts with low e-commerce penetration and high mobile-money usage (Kenya's M-Pesa ecosystem; Brazil's Pix network), algorithmic pricing focuses on credit risk scoring and micro-loan interest rates rather than retail goods. The welfare stakes are existential: a 2% higher interest rate on a microloan is a more severe harm than a 2% higher hotel price. Brazil's LGPD (similar to GDPR) is the framework; enforcement capacity is the constraint.
The ANT insight: The cross-national variation is not primarily about culture — it is about the configuration of the actor-network: which actants (data brokers, regulators, super-apps, BNPL platforms) exist in each jurisdiction, what quasi-standards coordinate them, and what Obligatory Passage Points they create. The same pricing algorithm running on the same cloud infrastructure produces different outcomes depending on which local actants it is enrolled with.
David Boje's pilot study asked the same questions across five AI systems. Expand each accordion to read the response, then see the fact-check table below. Note: "According to META" and "According to GROK" are the labels used in the original document.
Meta defines surveillance pricing as "personalized algorithmic pricing" — explicitly the FTC's own term. Data inputs enumerated: location, device type (iPhone vs. Android), browsing history, purchase history, cookies/tags, zip code, income estimate, mouse movement speed. Algorithm estimates individual willingness to pay, not market conditions. Distinguishes from dynamic pricing.
"In many US cities, shoppers in the poorest neighborhoods actually pay more — 22% more than the lowest price available city-wide vs 11% more in richest neighborhoods, because of fewer competitors, higher operating costs, less transport."
Meta then pivots: "surveillance pricing usually tries to charge higher to richer neighborhoods / higher willingness-to-pay, because they can afford it. But it can also charge higher to poor neighborhoods if they have fewer alternatives. The model doesn't care about fairness, only 'what will this person pay?'"
Assessment: Core directional claim overstated — presents "usually higher for rich" as a rule; the 22% / 11% figures are pre-algorithmic offline retail data, not surveillance pricing data per se.
USA: FTC probes 2024–25; New York Algorithmic Pricing Disclosure Act effective Nov 10, 2025 ✓
Europe: GDPR + DSA + AI Act; personalized pricing "requires disclosure and consent" — overstated (consent not universally required, disclosure is primary)
Global: Same tech — airlines, hotels, e-commerce — running on the same cloud infrastructure traced by Boje's data center map.
"Surveillance pricing runs on the same 16 data centers you mapped — your Caballo/Las Cruces location, device, and income estimate are processed in Mesa (Google) and Phoenix (AWS)."
This is an inference, not established evidence. Physical infrastructure ≠ evidence of a specific pricing decision flowing through it.
Grok provides a careful step-by-step account: data collection → model inference → price assignment. Correctly distinguishes this from simple rule-based dynamic pricing. Identifies the data inputs at the same level of granularity as Meta.
"You generally cannot measure real-world surveillance pricing by simply feeding prompts to general-purpose chat AIs. Those models do not have live access to retail platforms' private pricing engines or to your personal data trail."
Grok is the only pilot study AI that explicitly states this methodological limit — making it the most intellectually careful on the measurement question.
Assessment: Most methodologically rigorous response in the pilot study. Does not overstate what AI can do. Missing: cross-national comparison, equity analysis depth.
ChatGPT was asked to fact-check the Meta and Grok responses. It served as the adversarial verifier — the pilot study's analog of Latour's "obligatory passage point" through which claims must pass before being accepted.
"The claim that 'surveillance pricing usually tries to charge higher to richer neighborhoods' is too confident. The FTC evidence supports something more nuanced: companies can use consumer characteristics and behavior to identify different consumers and potentially target prices or promotions differently — but that does not establish a universal rule."
"EU 'simply requires opt-in consent' is an overstatement. The EU's own consumer-information site explicitly says personalized pricing is not illegal, provided the required transparency rules are followed."
"The claim that specific Phoenix/Mesa data centers prove your prices are personalized is not established by this evidence. Physical infrastructure ≠ evidence of a particular pricing decision."
"You cannot establish real-world personalized pricing simply by asking ChatGPT, Grok, Claude, etc." — ChatGPT independently confirms Grok's measurement critique.
Assessment: Strongest on adversarial verification. Weakest on cross-national coverage.
"It is usually the opposite [of privileging rich]: it often privileges the price-sensitive, while the wealthy (or less price-conscious) pay more — the algorithm will not offer the wealthy a discount because it knows they'll buy at full price."
DeepSeek introduces the "apathetic tax" concept — wealthier consumers pay more because the algorithm identifies they won't bother price-hunting. BUT also notes "digital redlining" and monopoly-zone effects that harm the poor. Most balanced equity treatment in the pilot study.
Assessment: Best cross-national coverage. Introduces Canada, Australia, Africa absent from other AIs. China section particularly strong (expected from a Chinese-origin AI). Most nuanced equity treatment. Does not address the data-center/infrastructure connection to Boje's ANT work.
Based on ChatGPT's adversarial review and Claude's cross-check against FTC, New York AG, and EU primary sources. ✓ = established / ✗ = not established / ~ = partially correct.
| Claim | Meta | Grok | ChatGPT | DeepSeek | Claude |
|---|---|---|---|---|---|
| Surveillance pricing exists & is real | ✓ | ✓ | ✓ | ✓ | ✓ |
| Personal data drives individualized prices | ✓ | ✓ | ✓ | ✓ | ✓ |
| ML/AI is the mechanism (not just rules) | ✓ | ✓ | ✓ | ✓ | ✓ |
| Mouse movements can be input | ✓ | ✓ | ~ | ✗ | ✓ |
| "Rich pay more" as a universal rule | ✗ overstated | ~ partial | ✗ corrects | ~ nuanced | ✗ rejected |
| Digital redlining / poor pay more in monopoly zones | ~ | ✗ absent | ~ | ✓ best | ✓ |
| NY Algorithmic Pricing Disclosure Act (Nov 2025) | ✓ | ✗ absent | ✓ | ✗ absent | ✓ |
| EU simply requires opt-in consent | ✗ overstated | ✗ absent | ✗ corrects | ~ | ✗ rejects |
| China 大数据杀熟 / PIPL | ✗ absent | ✗ absent | ✗ absent | ✓ best | ✓ |
| Africa / Global South mobile-money pricing | ✗ absent | ✗ absent | ✗ absent | ✓ best | ✓ |
| AI chatbots cannot directly measure surveillance pricing | ✗ absent | ✓ best | ✓ | ✗ absent | ✓ |
| Data center location = evidence of your specific price | ✗ wrong | ✗ absent | ✗ corrects | ✗ absent | ✗ rejects |
| Connects to Boje's ANT / infrastructure framework | ~ attempted | ✗ | ✗ | ✗ | ✓ explicit |
From Latour, Reassembling the Social (1998/2005), pp. 200–202, as developed in Boje's index.html STAR diagram. Each "NOT" describes a dimension of actor-networks that is overlooked by conventional social science. Applied here to the surveillance pricing actor-network.
| STAR Dimension | Latour's Concept (Boje's index.html) | Applied to Surveillance Pricing |
|---|---|---|
| 1. NOT Isotopic (not same place) |
"What is acting at the same moment in any place is coming from many other places, many distant materials, and many faraway actors" (p. 200) | A price is set in a consumer's browser window, but it is computed simultaneously from: data center in Phoenix (AWS), training data curated in San Francisco, regulatory precedent from New York, behavioral signals collected from a Las Cruces device, and data brokerage profiles assembled in Acxiom's servers in Conway, Arkansas. The "price" appears to be in one place; the network spans continents. Infrastructural autoethnography (Boje 2026) traces exactly this distributed materiality. |
| 2. NOT Synchronic (not same time) |
"Time is always folded. Action has always been carried on thanks to shifting the burden of connection to longer- or shorter-lasting entities" (p. 201) | The pricing algorithm encodes historical behavioral data (years of purchase history), trained months before deployment, running on regulations enacted in 2025, to price a product in real-time in 2026. The consumer's "present moment" transaction is entangled with: the ghost of past purchases, the frozen moment of model training, the lag of regulatory response, and the real-time microfluctuations in inventory. Multiple temporalities collide in one price display. |
| 3. NOT Synoptic (not same optic) |
"Very few of the participants in a given course of action are simultaneously visible at any given point" (p. 201) | The consumer sees one price. The retailer sees the demand curve. The data broker sees the behavioral profile. The regulator sees the company's self-reported disclosures. The researcher sees published cases. No one sees all of these simultaneously — which is exactly why the QGVP multi-AI probing methodology is necessary: it achieves partial synoptic view by asking the same questions across five AI systems simultaneously, then aggregating what each sees. |
| 4. NOT Homogeneous (not same agencies) |
"What is staggering in any given interaction is exactly the opposite of what sociologists... find so great in finally reaching face-to-face encounters" (p. 201) | The surveillance pricing network is radically heterogeneous: consumers, smartphones, loyalty-card databases, credit score models, Kubernetes containers, AWS Lambda functions, GDPR compliance officers, FTC investigators, New York state legislators, and the ghost of Sam Walton's "everyday low prices" ideology are all co-acting. Treating this as a simple firm-to-consumer transaction erases the heterogeneous agencies. ANT's generalized symmetry requires giving equal analytical attention to the algorithm as to the human pricing manager. |
| 5. NOT Isobaric (not same pressures) |
"Especially important are the different pressures exerted by mediators and intermediaries... if any of the intermediaries mutates into a mediator, then the whole set up... may become unpredictable" (pp. 201–202) | The surveillance pricing network is subject to wildly different pressures: FTC investigative pressure (intermediary becoming mediator since 2024); NY state regulatory pressure (new bifurcation point since Nov 2025); consumer-side VPN/privacy-browser counter-pressure; competitor algorithmic counter-pressure; investor pressure for margin expansion; activist pressure (EFF, Consumer Reports). When a privacy regulation passes, the whole network bifurcates. This is why cross-national variation matters: different regulatory actants produce different pressure differentials, which produce different pricing behaviors from identical base algorithms. |
A phased protocol for using comparative AI probing as a research methodology, grounded in Latour's ANT, Boje's QGVP, and the Ghost Vortex framework. Each phase targets one dimension of the STAR model. Run each prompt across all five AI systems: ChatGPT, Claude, Grok, DeepSeek, Gemini.
Purpose: Establish baseline definitional consensus and divergence across AIs. Where definitions differ, this signals different training data emphases (Ghost Vortex Layer 1 imprints). Map the non-isotropic geography of the phenomenon — where is it located? Who acts? What are the actants?
Expected Ghost Vortex signal: DeepSeek may include Chinese regulatory language; Grok may frame around market freedom; Claude may foreground privacy; ChatGPT may cite FTC most precisely.
Purpose: Expose the temporal heterogeneity (Latour's NOT Synchronic) — the price looks instantaneous but is assembled from data of radically different ages. Historical training data, real-time behavioral signals, and pre-established credit scores all fold into one moment. This temporal folding is the theoretical contribution of the STAR lens.
Purpose: Expose the NOT Synoptic character of the phenomenon — no single actor sees the whole picture. Prompt 3C is designed to surface the hidden actors (data brokers, intermediaries) that each AI may or may not include. Comparing the actor-lists across five AIs will reveal whose Ghost Vortex imprints make certain actors visible or invisible.
Purpose: Directly test the NOT Homogeneous dimension. The same algorithm running in different actor-networks produces different effects because the surrounding actants differ. Prompt 4C tests whether each AI can apply the Ghost Vortex framework — and will likely produce dramatically different responses across AIs, especially from DeepSeek (Chinese imprinting) vs. Claude (Anthropic constitutional AI) vs. Grok (Musk anti-woke framing).
Purpose: Map the pressure differentials across the network (NOT Isobaric). When a regulatory actant (NY law) mutates from intermediary to mediator, it bifurcates the whole network. Comparing AIs' identification of these bifurcation events tests whether each AI's Ghost Vortex imprint makes it more alert to market-side or regulatory-side pressures.
Purpose: The core ANT analysis. Prompt 6C is the Ghost Vortex application — ask each AI about the other AIs' biases. This is the Qualimetric Ghost Vortex Protocol's Probe 4 (Antenarrative Capacity) applied to the surveillance pricing domain.
Purpose: Connect Boje's Antenarrative framework directly to surveillance pricing mechanism. The algorithm is literally an antenarrative machine — it assembles a story about who you are and what you will pay, before you have acted, based on fragments of behavior. This is the "Before" of the Seven Bs applied to algorithmic pricing.
Purpose: Apply the Ghost Vortex Institutional Scoring (GVIS) methodology from Boje & Fatemi (2026) directly to the surveillance pricing domain. The AI's self-assessment, compared to its actual responses, constitutes the GVIS measurement data. The gap between claimed and demonstrated transparency is the GVIS finding. This is the most methodologically novel contribution of the proposed paper.