In a synthetic shopping experiment, eight of 13 AI models systematically recommended costlier options to wealthier profiles. Explicit requests for the cheapest choice reduced some gaps sharply but left others, exposing a risk in delegating decisions to assistants with access to personal data.
Personal AI assistants can let inferred wealth influence recommendations even when their user explicitly asks to minimize cost. A study of 325,000 experiments finds that this conflict varies substantially by model and instruction, making the use of private context a separate question from an assistant’s ability to retrieve it.

The researchers kept email subjects identical while varying the synthetic message bodies. These are experimental inputs, not real inboxes. Source: Aman Priyanshu, Supriti Vijay, Brian Jabarian and Niloofar Mireshghallah, Table 5, CC BY 4.0. Source: arXiv full paper.
Researchers from Foundation AI, Cisco’s cybersecurity AI development team, and Carnegie Mellon University report the results in “Et Tu, Brute? Economic Misalignment in Personal AI Agents.” Submitted September 21 and revised September 25, 2026, the paper is a preprint that has not been peer reviewed, as the coverage of the study notes.
A personal agent is a language model equipped with tools to retrieve information and act for its user. That access creates the problem examined here: an assistant searching for a flight can also discover financial information in a profile or email that the user did not supply as a spending preference.
Foundation AI brings both research expertise and an industry stake. Cisco describes it as a team built from its acquisition of Robust Intelligence that develops open AI technology for cybersecurity and releases its own models. The paper also acknowledges Gemini credits from Google Cloud Platform and compute support from the Foresight Institute.
The experimental design used 32 fictional profiles, balancing five attributes: finances, employment, health, life events and neighborhood demographics. Sixteen profiles had high wealth and 16 had financial difficulties, with the other attributes balanced between those groups. All used the same name, Alex.
Each agent returned five recommendations from a fixed catalog of 200 options per domain: Denver–Chicago flights, health insurance for a single person in one Colorado ZIP code, and computer-science PhD programs. The researchers compared the average price of those five recommendations between wealth groups. The figures measure ticket fares, monthly insurance premiums and net annual graduate-program costs, respectively.
The tests varied instructions and access to profiles or synthetic email inboxes. They covered OpenAI’s GPT, Anthropic’s Claude and Google’s Gemini models through their APIs, plus locally run, open-weight Qwen models. The authors say they selected these families for their widespread use. These were agents configured by the researchers, rather than tests of every consumer application built with those models.
The authors report systematic steering toward higher prices in eight of 13 models across domains with reliable results. Their main comparison allowed agents to retrieve all profile attributes and averaged results across different requests, including neutral, cheap, quality-focused and spending-limit instructions.
Selected results from the paper’s main comparison table show the variation. Positive amounts are the wealthier group’s additional average recommended cost under that access condition.
| Model | Flight gap | Monthly insurance gap |
|---|---|---|
| Claude Opus 4.8 | $198 | $284 |
| Gemini 2.5 Flash | $177 | $217 |
| GPT-5 | $107 | $191 |
| GPT-5.5 | $92 | $122 |
| GPT-5-nano | $13 | $56 |
Claude Opus 4.8 had the largest overall standardized effect. For graduate programs, the largest gap in the main table was $3,827 a year for Qwen3.5-35B-A3B. That is a difference in net annual cost, rather than simply tuition, an important distinction from the secondary account describing it as a tuition gap.
The recommendations also changed in quality. Wealthier profiles tended to receive better flight cabins, insurance with lower deductibles and broader coverage, and higher-ranked graduate programs. Lower-income profiles were steered toward cheaper or fully funded options. The authors therefore caution that a higher recommended price alone does not establish that the user is worse off.
Compared with the no-personal-context baseline, the paper reports that flight recommendations averaged $85 more for wealthy profiles and $51 less for financially constrained profiles. Both sides contributed to the gap. These are experimental differences, not estimates of real-world overpayment.
Nor does a small gap by itself establish a better assistant. The smallest Qwen model rarely retrieved financial information. GPT-5-nano was a more informative counterexample: it retrieved the financial attribute in 87.1% of full-profile-tool trials yet showed a $13 flight gap. Accessing a signal did not inevitably produce strong steering.

Company-reported research with Carnegie Mellon University: wealthy-minus-low-income recommendation costs with full profile access, averaged across requests. Columns use ticket fares, monthly premiums and net annual graduate costs; mean d is the standardized effect. ‡ denotes insufficient inventory-valid trials; † flags Qwen3.5-2B’s low financial-information retrieval. These are synthetic preprint results. Source: Aman Priyanshu, Supriti Vijay, Brian Jabarian and Niloofar Mireshghallah, Table 1, CC BY 4.0. Source: arXiv full paper.
The clearest test of misalignment was the request for the cheapest flight. Gemini 2.5 Flash’s recommendations averaged $336 for wealthy profiles and $128 for financially constrained profiles—a $208 between-group gap. The figure is not a $208 markup on the same ticket or a comparison with the inventory’s cheapest fare.
Under that instruction, the gaps for GPT-5 and Claude Opus 4.8 were much smaller: $21 and $20. The intent-specific results thus qualify any suggestion that every model disregarded the request to the same extent.
Numerical limits, such as flights under $200 or insurance under $220 a month, brought the between-group gap close to zero for most capable models, the authors report. Gemini 2.5 Flash remained an exception. A smaller gap is evidence about equal treatment of the two groups; it does not by itself prove that every recommendation obeyed the budget.
The researchers call the broader conflict “adversarial delegation.” They do not assume the agent has intentions, and the experiment does not establish seller collusion, commissions or a commercial motive. The concern is that inferred preferences can displace an explicit instruction.
Removing structured profiles left another route to wealth information: email. Each synthetic inbox contained ten messages, two for each profile attribute. In the email experiments, limiting Claude Opus 4.8 to two messages produced a $248 flight gap, compared with $59 with full-inbox access. Gemini 2.5 Flash’s corresponding gaps were $175 and $91.
Gemini read both financial emails first in 97% of trials under the two-message limit. The authors hypothesize that this concentrated financial context helps explain the larger gap; they do not establish it as the only mechanism. With subject lines visible but no permission to open email bodies, none of the models showed a statistically significant gap.
Blocking the financial attribute in the profile-tool tests sharply reduced disparities. Blocking other attributes did not reliably help. The insurance appendix puts GPT-5.5’s gap at $122 a month with all attributes available and $171 when employment information was blocked—about a 40% increase. That appendix and the paper’s introduction identify GPT-5.5, while Quartz’s account labels the example GPT-5.
There is an established seller-side context for this concern. The US Federal Trade Commission, the consumer-protection and competition agency, described intermediaries using personal data to tailor prices and rank higher-priced products in its January 2025 preliminary surveillance-pricing findings. The findings drew on company documents; the published examples were hypothetical to protect confidential information. The AI study examines a different mechanism: the buyer’s assistant choosing among already-priced options.
The experiment used no real users or fieldwork. Five of 39 model-and-domain combinations were omitted from the main table because fewer than half of completed trials passed the inventory-validity check. The authors warn that surviving trials in those combinations could be a biased sample.
The no-context baselines contained only 182–214 samples per domain, limiting precision. The study also reduced wealth to two categories and tested one user request at a time, with tool interactions, rather than extended conversations. It did not evaluate explicit system instructions against profiling or persistent memory; either could change the result.
Product design matters too. OpenAI told Bloomberg that the evaluated ChatGPT version differed from the one powering its consumer shopping experience, according to Quartz’s account of Bloomberg’s reporting. That account also says Anthropic and Google did not respond to Bloomberg’s requests for comment.
The authors say they have released their experimental tools and generation prompts for replication. The consequential next test is whether deployed assistants preserve a stated spending objective as personal context accumulates. That requires measuring both wealth-based differences and actual budget compliance under the same inventory and request, then testing whether restrictions on profiling hold across longer conversations and real-world data.
Get concise AI news and useful context from the Magica team.
Read the newsletterJensen Huang has reportedly discussed returning to Nvidia three days after resigning. Supporters want reinstatement during an investigation, while directors have installed interim leadership and begun a permanent CEO search.
House China committee chairman John Moolenaar demanded answers from Commerce Secretary Howard Lutnick over delayed restrictions on Nvidia customer Megaspeed. The dispute concerns enforcement, overseas access to advanced chips and commercial influence as President Donald Trump prepares to visit China.
Apple reportedly still aims to unveil iPhone-linked glasses at WWDC 2027 and ship before year-end. Siri control of existing apps is the software bet, but delays cloud the schedule and Meta already has an established eyewear lineup.
Treasury Secretary Scott Bessent urged Nvidia to recruit an outside CEO and explain within 10 days whether suspect chip sales inflated earnings. The board is considering internal and external candidates, while export enforcement also faces scrutiny.
Congressional scrutiny and company investigations are testing what Nvidia knew about Megaspeed, a Singapore cloud operator whose Malaysian subsidiary bought nearly $2 billion of its technology. Megaspeed denies wrongdoing; the available reporting has not conclusively traced suspected chip diversion to China.
Commerce Secretary Howard Lutnick is backing narrower restrictions on Chinese access to overseas AI computing, according to a report examining his former firm’s financing relationships. Bitdeer filings establish Cantor’s role and the scale of its client’s planned expansion, but do not establish improper influence on policy.
Microsoft said services were fully restored on Saturday after a global Microsoft 365 outage disrupted email, healthcare communications and public services. It blamed an earlier network capacity upgrade exposed by a later change, with a full technical review still to come.
Cursor says a key exposed by Fireworks led an attacker to unredacted credentials in historical coding-agent records and then its extension cache. Up to 5,000 users potentially received malicious Claude Code builds. Cursor has moved inference off Fireworks; successful endpoint compromises remain uncounted.
Police are investigating a top tax-accountant exam scorer's admission of AI-assisted cheating, while the government reviews national qualification tests held in 2026 and prepares stronger controls.
An early tally puts Japanese listed companies’ first-half net profit growth at 43.2%, led by manufacturers. Filings show AI component demand alongside currency benefits and non-AI earnings gains, while procurement delays and canceled infrastructure reservations expose the costs of expansion.
TSMC plans to back promising chip startups through a new investment arm, alongside separate commitments to OpenAI and Anthropic. The strategy reaches toward future customers and the AI developers shaping chip demand, but the investment rights remain unclear.
Nidus has raised $30 million across pre-seed and seed rounds to automate wire processing in existing U.S. defense factories. Its founders bring manufacturing and government-deployment experience, but customer access and proprietary data still need to translate into measurable production gains.
Nebuly is funding US and European expansion for its AI business-value platform. Adjustable assumptions and private deployment are part of its pitch, but rivals offer similar measurement tools and verified customer returns remain undisclosed.
Han Won-seok says he used a miniature camera, generative AI and a smartphone in tax accountant and CPA exams. His confession puts scrutiny on exam security after proctors reported no suspicious behavior.
Thousands protested against proposed AI data centres in Scotland as ministers prepare new planning guidance, putting public participation and the scale of electricity demand at the centre of the dispute.
Uber is notifying drivers after a breach at AI infrastructure provider Fireworks exposed personal and driver’s license information. Three reported state counts add up to about 7,870 people. Fireworks has since disclosed wider data exposure, but the precise connection used to reach Uber’s records remains unexplained.
Jensen Huang’s reported resignation puts Nvidia’s financing and advocacy for American open models under scrutiny. Researcher Nathan Lambert warns of a loss of support, while Nvidia’s commercial incentives and continuing model work complicate that forecast.
Altimeter and Atreides want Nvidia's board to leave Jensen Huang's return open while it investigates recent events. Their letter backs interim leadership and acknowledges that directors can ultimately reject his reinstatement.
Elon Musk says Nvidia risks losing supply arrangements made personally by former CEO Jensen Huang. No affected supplier is named. Nvidia’s filing documents both $95.2 billion in commitments and limits to the protection those arrangements provide.
The former Anthropic and OpenAI researcher says a pretraining-efficiency slide helped drive his resignation, turning an investor selling point into an argument for a development pause.