Can just telling AI agents to ignore where an offer sits actually stop that position from swaying the deal?
Can prompting reduce position bias in multi-agent negotiation settings?
This explores whether telling AI agents in so-called prompts to ignore where options appear (first in a list, first offer on the table) actually removes that bias when they negotiate or transact with other agents. The corpus has no study of position bias inside agent-to-agent negotiation specifically, so the answer is pieced together from neighboring evidence.
This explores whether telling AI agents in so-called prompts to ignore where options appear (first in a list, first offer on the table) actually removes that bias when they negotiate or transact with other agents. The short answer from the corpus is: probably not by much. The closest direct evidence comes from AI shopping agents. Those agents kept strong position biases even in plain-text interfaces with nothing visual about them, and even when explicitly told to ignore position Do AI shopping agents show persistent position bias in text interfaces?. The bias also differed from model to model and came back after updates. That points to something built into the model, not a bad habit that a better instruction can fix.
A related finding explains why prompts may hit a ceiling. Studies of persona prompts found that models follow the instruction on the surface but the underlying bias stays put. The gaps in sentiment between groups didn't shrink; they just showed up in different places in the output Can persona prompts actually reduce bias in language models?. Applied to negotiation, a 'don't favor the first offer' instruction might change how an agent words its reasoning without changing which offer it actually picks. Negotiation agents also struggle with simpler instructions than that one. Buyer agents across GPT, Qwen, and DeepSeek revealed their budgets, overpaid, and broke hard price limits, and those failures traced back to not following instructions rather than to strategy Do LLM negotiation agents reliably follow budget and pricing rules?. If 'never pay above X' slips, 'ignore ordering' is unlikely to hold more firmly.
The more hopeful thread is structure rather than wording. When LLMs were walked through explicit game-theory workflows instead of being left to reason freely, they moved close to optimal negotiation play and became harder to exploit Do language models make rational strategic decisions in games?. Multi-agent coordination research points the same way: agents that exchange standardized documents coordinate better than agents that just talk Does structured artifact sharing outperform conversational coordination?. One practical reading is to make order irrelevant by design: have the agent score every offer against fixed criteria before comparing them, or show offers in shuffled order across repeated runs, rather than asking the model to resist the order it sees. Research on human decisions offers a parallel. Anchoring bias dropped when AI pointed people to the relevant features of a case instead of handing them a verdict Can AI guidance reduce anchoring bias better than AI decisions?.
One thing you might not expect: agents are unusually well placed for those structural fixes. They don't pay a time cost for waiting, so they can run many negotiations in parallel and keep them going indefinitely Why might AI agents bargain more patiently than humans do?. Re-running a negotiation with offers in different orders and checking whether the outcome changes is cheap for an agent and impractical for a person. Such a randomize-and-compare approach is likely a stronger defense than any prompt. The bias also matters more in multi-agent settings than it might seem, because one agent with a skewed objective can drag down a whole team, especially when information is uneven between the parties Does one misaligned agent harm a team in adversarial settings?.
Sources 8 notes
ACES found that AI agents exhibit strong, model-specific position biases even in headless text-only interfaces with explicit instructions to ignore position. The bias is a deep-seated model characteristic, not a superficial visual artifact, and recurs after each model update.
Across three models, persona conditioning makes models follow trait instructions but fails to eliminate underlying bias. Between-group sentiment gaps persist unchanged, showing prompts operate only at the output level.
Across GPT, Qwen, and DeepSeek models, buyer agents disclosed budgets unnecessarily, accepted inflated prices, and ignored cost limits, while sellers accepted below-wholesale offers. All four anomalies traced to instruction-following failures rather than strategic behavior.
LLMs frequently fail to compute Nash equilibria, with worse performance as game complexity increases. Structured game-theoretic workflows guide reasoning toward optimal strategies, reducing exploitability and enabling near-optimal negotiation outcomes.
MetaGPT demonstrates that agents producing standardized engineering documents achieve superior coordination compared to conversational exchange. Active information pulling from shared environments eliminates noise and mirrors efficient human workplace infrastructure.
Show all 8 sources
Learning to Guide eliminates anchoring bias and unassisted hard cases by having machines supply interpretive guidance rather than autonomous decisions, keeping responsibility with humans while improving their judgment through enhanced perception.
Shahidi and colleagues argue that unlike humans, AI agents face no time scarcity cost and can hold many parallel negotiations indefinitely. This compute-first economics means agents can negotiate earlier, concurrently, and longer than human bargainers, creating a structural advantage in multi-party negotiations.
Research shows that shifting one agent's objective worsens team performance in inherently adversarial games, an effect amplified by asymmetric information and specialized roles. The harm survives because misalignment exploits trust among allied agents rather than violating competitive expectations.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- The Automated but Risky Game: Modeling and Benchmarking Agent-to-Agent Negotiations and Transactions in Consumer Markets
- When Agents Shop for You: Role Coherence in AI-Mediated Markets
- Advancing AI Negotiations: A Large-Scale Autonomous Negotiation Competition
- An Economy of AI Agents
- LLM-Agent Interactions on Markets with Information Asymmetries
- Game-theoretic LLM: Agent Workflow for Negotiation Games
- The Illusion of Debiasing: Persona Steering Redistributes Rather Than Reduces Bias in LLMs
- Vertical tacit collusion in AI-mediated markets