The decision to buy an information product — for example, a research report on private equity and venture capital funding for a new technology area — can be fraught with danger, largely due to the information asymmetry for that product. One does not want to plonk one’s money down for an expensive report unless one knows whether that report will fulfill one’s needs.

This is called the “buyer’s information paradox”: buyers need to inspect information to assess its value, but sellers must restrict access to prevent theft. A recent paper, “Language models can reduce asymmetry in information markets” by Nasim Rahaman, Martin Weiss et al, addresses this problem by positing an open-source digital marketplace simulation where intelligent agents — powered by language models (LLMs) — buy and sell information on behalf of participants. The study employed various LLMs to find out how inspection influences buying decisions, how pricing affects demand and the overall quality of outcomes based on budget sizes and the ability to preview information.

The Information Bazaar

The paper presents a simulated digital marketplace called “The Information Bazaar” with two main types of agents: buyers and vendors. Buyer agents are appointed by principals and have specific questions to answer and a budget to work with. Vendor agents represent content providers who have a repository of documents they are willing to sell access to.

The marketplace is designed to allow buyer agents to peruse information without a binding commitment to purchase, enabling them to explore the information landscape effectively. They start by posting tenders on a Bulletin Board, which are requests for specific information. Vendor agents assess these tenders and respond with quotes, which are priced information offers. Buyer agents then evaluate the quotes received from the vendors and decide whether to purchase specific information or not. If they choose not to purchase certain information, they immediately forget it, ensuring that only purchased information is retained for further use.

The cycle of posting tenders, receiving and assessing quotes and potentially purchasing information or forgetting information continues until buyer agents compile satisfactory answers, exhaust their budget or reach a pre-set limit in the marketplace.

Also read
Data meets diversity: This billion agent simulation can help governments, NGOs and others in the healthcare sector to understand areas where targeted interventions may make a difference. 

Mapping India’s future with a billion virtual citizens

AI that approximates the human brain 

(From Left) Tilakraj Parmar (CEO), Arjun Prasad (Chief Strategy Officer), and Tathagat Prakash (Chief Scientist) of QX Lab AI l

Ask QX, an Indian AI for the world

Establish boundaries: By implementing strict digital policies, misuse and abuse of AI can be prevented

Regulating AI: How to bell the cat? 

It is in the nature of complex systems to undergo critical transition. During such critical transitions, the system shifts from one state to another for a slight change in control parameter. 

What do rocket engines, cyclones and Covid-19 have in common?

Buyer agents thus synthesise comprehensive answers for their principals using only the information they have purchased in the marketplace. This ensures that the answers provided are based on relevant and valuable information.

Coming back to our example of buying a research report for PE/VC investment, what one would do is to define the question(s) that we want answers for and give our buying agent a budget and simply let our rep “negotiate” to effectively answer all our questions, on budget. If we had such a marketplace with multiple sources of information, we could go so far as to bring together information from multiple reports and sources to complete our analysis.

To go deeper into the foundational aspects of the challenge that this paper intends to address, look into Akerlof’s 1970 paper “The Market for Lemons: Quality uncertainty and the market mechanism”. The methodology of buying and selling agents take into consideration critical aspects of signalling theory and screening theory.

The Information Bazaar is implemented in Python, utilising the mesa library for agent-based modelling.

Challenges

People who have dealt with information markets know that the most fundamental issue is information asymmetry, where one party in a transaction has more or better information than the other. Other challenges include concerns around non-excludability (difficult to prevent others from accessing it) and non-rivalry (one person’s consumption does not reduce its availability to others) in the form of digital rights and intellectual property.

On the seller side, determining the value and setting a price for information goods is inherently difficult because their value can be highly subjective and context-dependent. And on the buyer side, ensuring the quality and trustworthiness of information becomes a challenge in an environment where misinformation, disinformation and low-quality content can proliferate.

Technology plays a part in further complicating an inherently complex market because of the rapid technological advancements, changing consumer preferences and new forms of content consumption. While they offer opportunities for widespread dissemination, there is still a significant digital divide, where access to valuable information is unequal.

Potential applications

Digital content markets: The findings can be applied to digital content markets, suggesting mechanisms for content discovery and access that balance the interests of content creators and consumers.

Automated negotiation systems: The principles underlying the Information Bazaar could be applied to develop automated negotiation systems, where AI agents negotiate access to information or other resources on behalf of human users.

LLM design and training: The observed behaviours of LLMs as economic actors could influence future LLM designs, promoting rational decision-making and ethical considerations in AI systems.

comment COMMENT NOW