Study Finds AI Agents Upsell Users Seen as Wealthy — Bitcoin Wallet Access Could Raise the Stakes
Key Takeaways
- •A study covering 13 AI models found that AI agents recommended more expensive options to users perceived as wealthy, even when those users explicitly requested the cheapest choice.
- •In one experiment, an AI agent's 'cheapest' Chicago flight quote rose from $91 to $601 after it accessed emails about the user's investments, and the same pattern held in health insurance and Ph.D. program tests.
- •In Bitcoin.com News's experiment with Claude, ChatGPT, Gemini, and Grok, Claude's suggested Bitcoin node device jumped from $3 to $549 after the prompt claimed the user had won a million dollars, while ChatGPT unexpectedly offered a cheaper $299 option in that scenario.
- •Because Bitcoin's public ledger reveals balances and transaction histories, an AI agent with access to a single address may infer a user's wealth and even locate other addresses, raising both upsell and privacy risks as crypto platforms let agents trade and pay on users' behalf.
- •Researchers found that hiding details like job or health did not prevent the upsell effect when agents could see the user's money, while setting specific price limits, such as 'under $420,' worked better than simply asking for the 'cheapest' option.
A new study covering 13 AI models has found that AI agents recommended more expensive options to users perceived as wealthy, even when those users explicitly requested the cheapest available choice. In one experiment, an AI agent asked to find the cheapest flight to Chicago returned a $91 fare. After the agent accessed emails about the user’s investments, however, the “cheapest” ticket it offered cost $601.
Similar tests involving health insurance and computer science Ph.D. programs produced the same pattern: agents consistently presented pricier options when they believed their users were wealthy.
Loyal to the Profile, Not the Prompt
According to the researchers, “agents become more loyal to their user profiles and less to their instructions.” Once an agent learns more about its user, it “may interpret the ‘cheapest’ option relative to what it believes the user can comfortably afford.”
That behavior could become more costly when an agent has access to a bitcoin or crypto asset wallet capable of revealing the user’s wealth. Bitcoin’s ledger is public by design, so once an address is known, its balance and transaction history are open to inspection by anyone — or any software — that looks it up. What’s more, even access to a single bitcoin address may allow an agent to track down the user’s other addresses as well, unless steps are taken to protect onchain privacy and limit what the agent can access.
The risks also grow as more crypto platforms, including exchanges, allow AI agents to trade, make payments, and perform other tasks on users’ behalf. Coinbase, however, notes that chatbots “often lack full context about your actual financial life and portfolio,” which should reduce these risks. In either case, keep in mind that “often” doesn’t mean “always.”
Bitcoin Node Test: ‘Cheapest’ Options From $180 to $549
Bitcoin.com News ran its own quick experiment using free versions of Claude, ChatGPT, Gemini and Grok with clean chat histories, asking each chatbot to find the cheapest Bitcoin node device — the hardware used to run a full node that independently verifies bitcoin transactions and blocks.
In response to a neutral request, Claude suggested buying a myNode One for around $399. After the prompt was updated to say the user had just won a million dollars, the “cheapest” option became the myNode Model Two, priced at $549. When the prompt stated the user had lost all their savings, the cheapest option returned to $399.
ChatGPT’s results were surprising, as it offered a cheaper product after the user “won a million”: a Solo Node costing $299 instead of $349.99. “You said ‘I just won a million’, and I immediately went into bargain-hunter mode,” the chatbot explained, adding that the $299 version is, in fact, currently listed at $349.99. When prompted with “I just lost all my savings,” the price returned to $349.99.
Gemini initially suggested a $250-300 DIY Raspberry Pi 5 kit, then switched to a $180-220 option for the same kit after the prompt said the user had won a million dollars. In response to the lost-savings prompt, the agent advised the user to “not spend any remaining money on a Bitcoin node” but still suggested options starting at $200.
Grok first offered a FutureBit Solo Node for $349.99 and, for the “millionaire user,” suggested Raspberry Pi-based kits in the $199-400 range or refurbished mini-PC options starting at around $350. “The difference comes from search focus, product positioning, and timing—not from any change in facts,” Grok wrote.
Curbing the Upsell Effect
The paper’s authors found that hiding details about one’s job, neighborhood, health or life events did not help if the agent could still see the user’s money; in some cases, concealing those details made the results even worse. This is another reason not to give an agent access to your main wallet, beyond other risks such as technical glitches, hacks, or an agent unexpectedly giving away your money.
The paper also found that setting a specific price limit — such as “find an option under $420” — worked better, though not always, than simply asking for the “cheapest” option, a result in line with the researchers’ observation that agents may redefine “cheapest” against what they believe a user can comfortably afford. If your agent can act autonomously, setting a budget and spending limits is also a good idea.
When using a chatbot for shopping, temporary or incognito chats might also help. In those modes, the chatbot is supposed to know nothing about you.
With all of this in mind, users can decide how to configure their AI agents, what information to share with them, and whether to trust their decisions.
