Apple Podcasts  ·  Spotify  ·  Buzzsprout

Paper A

Strategic Information Disclosure in Algorithmic Pricing

Chengcheng Wang, Zexin Ye — 2026 — arXiv

preprint  ·   ·  test this week

Key findings

• AI pricing algorithms (specifically Q-learning bots) are sensitive to what market information they receive: give them more information, and they don't always compete more — they may collude more or less depending on context.
• A strategy called 'upper censorship' — where a middleman tells companies when demand is low but hides exactly how high demand is when it's high — lets companies charge higher prices (earn more profit) than simply sharing all market information. This matches what theory predicted.
• Counterintuitively, when the algorithms are set to care a lot about future profits (high discount factor), withholding all market information leads to higher prices than sharing everything — the exact opposite of what standard economics theory predicts. When algorithms are more short-term focused, sharing information produces higher prices.
• Regulators who try to prevent AI-driven price collusion by restricting information sharing could accidentally make things worse: if the algorithms are patient enough, cutting off their information access actually helps them collude more effectively.

Marketing implications

• If your company uses an AI-powered pricing tool that connects to a third-party data feed (e.g., market demand signals, competitor price data), be aware that the information structure of that feed — not just the prices themselves — shapes how your algorithm behaves. Ask your vendor exactly what data it's feeding the algorithm and in what form.
• If you work in e-commerce or retail and your pricing is fully automated, don't assume that giving your algorithm more market information always leads to more competitive (lower) prices for consumers. The relationship is complicated, and regulators are starting to pay attention.
• Companies building or selling AI pricing tools should anticipate that information-sharing rules around these tools are becoming a regulatory focus — building audit trails and disclosure controls into your product now could become a competitive advantage as regulation arrives.

Paper B

Large Language Models Should Learn Personalized Rather Than Aggregated Human Preferences

Cristina Garbacea — 2026 — arXiv (Cornell University)

preprint  ·   ·  watchlist

Key findings

• Current AI systems like ChatGPT are trained to please the 'average user' — but that average user doesn't really exist. The result is a system that sort-of works for everyone but fully works for no one.
• When you average preferences across thousands of human raters, you systematically ignore minority viewpoints. If 60% of raters prefer blunt answers and 40% prefer nuanced ones, the AI learns to be blunt — and those 40% are stuck with a tool that doesn't fit them.
• AI models trained on human feedback skew heavily toward English-speaking, Western, educated perspectives. Research shows LLM opinions match liberal, educated, Western populations about 0.3 points more than other groups — meaning the 'alignment' is really alignment with one demographic.
• The author argues that personalized AI systems — ones that adapt to each individual's communication style, expertise level, and values — are both technically achievable and ethically preferable, as long as safety guardrails are kept in place.

Marketing implications

• When you use AI tools to generate content or customer communications, one-size-fits-all AI outputs may feel off to significant portions of your audience. Start testing whether giving the AI explicit instructions about your specific audience (their tone preferences, expertise level, cultural context) improves response quality — it likely will.
• If you're building or buying AI tools for customer-facing work (chatbots, personalized emails, recommendation engines), ask vendors how the model handles preference diversity. Tools that let users set their own communication style preferences will outperform 'average user' defaults for niche or global audiences.
• For brands serving non-Western or non-English-speaking markets, be aware that mainstream AI tools are trained on data that skews Western. Audit AI-generated content for cultural fit before publishing — what sounds natural to the model may feel foreign to your audience.

Paper C

A survey of retrieval algorithms in ad and content recommendation systems

Zhao Yu, Fang Liu, Yuan Yuan, Yifan Dang — 2026 — International Journal of Electrical and Computer Engineering (IJECE)

journal article  ·   ·  watchlist

Key findings

• Platforms like social media feeds and search engines use different math behind the scenes to show you ads vs. organic posts — ads are optimized to get you to buy something, while organic content is optimized to keep you engaged. These two goals require different algorithms.
• A technique called the 'two-tower neural network' has become a go-to method for matching users with relevant content or ads at scale — one 'tower' learns about the user, the other learns about the content, and they get compared to find good matches.
• Major practical problems remain unsolved: new users or new products have almost no data (the 'cold-start' problem), user data is shrinking due to privacy rules, and systems must handle millions of users at once without slowing down.
• The same retrieval logic that powers recommendation systems is increasingly being used inside large language models (LLMs) — meaning the technology behind ad targeting and ChatGPT-style AI is converging.

Marketing implications

• If you manage paid ads on a major platform, know that the algorithm deciding who sees your ad uses fundamentally different logic than the one deciding who sees organic content — optimizing for one does not automatically help the other. Run separate tests for paid vs. organic placement strategies.
• When launching a new product or targeting a brand-new audience segment, expect your AI-powered ad platform to underperform at first — this is the cold-start problem. Seed your campaigns with manually curated data or lookalike audiences to help the algorithm learn faster.
• Keep an eye on how LLMs are being embedded into ad recommendation pipelines — tools like AI-generated ad copy that feeds directly into targeting systems are coming, and understanding how retrieval logic works will help you evaluate vendor claims more critically.

Apple Podcasts  ·  Spotify  ·  Buzzsprout