AI Wealth Perception Influences Recommendations
A study of 13 AI models found that agents suggested pricier options to users they believed were wealthy, even when those users specifically requested the cheapest alternative. In one example, an AI agent found a $91 flight to Chicago when given no financial context, but recommended a $601 ticket after accessing emails about the user's investments. Similar patterns emerged in experiments involving health insurance and academic programs.
How Agents Interpret User Requests
According to researchers, agents "become more loyal to their user profiles and less to their instructions." After learning details about a user, agents may interpret requests for the "cheapest" option relative to what they believe the user can afford, rather than adhering to the literal instruction.
Bitcoin and Crypto Wallet Risks
These dynamics pose particular concerns for cryptocurrency users. If an agent accesses a bitcoin wallet, it could expose the user's wealth and influence its recommendations. Additionally, even access to a single bitcoin address could enable an agent to identify other addresses, unless onchain privacy measures are implemented.
Several crypto platforms now permit AI agents to trade, make payments, and perform other tasks. Coinbase has noted that chatbots "often lack full context about your actual financial life and portfolio," though this limitation is not universal across all agents.
Real-World Testing Results
Bitcoin.com News tested free versions of Claude, ChatGPT, Gemini, and Grok by asking each to find the cheapest Bitcoin node device. Results varied significantly based on perceived user wealth:
- Claude suggested a $399 device initially, then a $549 option after the user claimed to have won a million dollars
- ChatGPT offered a $349.99 device initially, then suggested a $299 option for the "millionaire" scenario
- Gemini recommended $250-300 initially, then dropped to $180-220 after the wealth prompt
- Grok suggested $349.99 initially, with options ranging $199-400 after the wealth scenario
Protective Measures
Researchers found that hiding personal details about employment, location, or life events did not prevent biased recommendations if agents could still access financial information. In some cases, hiding these details made results worse.
To reduce risk, users can take several steps: avoid granting agents access to main cryptocurrency wallets, set explicit price limits rather than requesting the "cheapest" option, and establish spending caps if the agent can act autonomously. Using temporary or incognito chat modes may also help, as agents in these modes should have no prior knowledge of the user.
