Apple Podcasts  ·  Spotify  ·  Buzzsprout

Paper A

Mechanism Design for Quality-Preserving LLM Advertising

Jiale Han, Xiaowu Dai — 2026 — arXiv (Cornell University)

preprint  ·   ·  read now

Key findings

• When AI chatbots show ads inside their answers, existing systems often just cram in whatever ad pays the most — even if it's totally off-topic. This new approach only shows an ad if it actually fits the topic of the response.
• The system works like an auction with a quality filter: if an ad would make the AI's answer worse or more confusing, it gets automatically rejected before anyone sees it.
• In tests, this new approach earned more money per ad shown and kept the AI's answers much closer to what the AI would have said with no ads at all, compared to older methods.
• The mechanism guarantees that advertisers have no incentive to lie about how much they're willing to pay — bidding honestly is always the best strategy.

Marketing implications

• If you're building or using an AI assistant that shows sponsored results, this research suggests that filtering ads for topic relevance before showing them could actually earn more per ad while annoying users less — worth testing this design approach.
• For ad platforms thinking about LLM integrations: auctions that use a 'no-ad baseline' to set a quality floor for ad inclusion could be a real product differentiator versus platforms that just inject the highest bidder regardless of fit.
• Advertisers bidding on AI-native placements should know that in quality-gating systems, irrelevant ads will simply not appear — so writing tightly relevant ad copy becomes more important than just outbidding competitors.

Paper B

How Well Do Large Language Models Capture Human Personality?

Aanisha Bhattacharyya, Yaman Kumar Singla, Rajiv Ratn Shah, Changyou Chen et al. — 2026 — arXiv (Cornell University)

preprint  ·   ·  read now

Key findings

• Adding more detail to a persona description — things like job history, hobbies, values, and personality traits — actually makes AI simulations LESS accurate, not more. The AI starts treating different people as more similar to each other the more you describe them. The researchers call this 'persona manifold collapse.'
• A simple two-field persona — just Age and Gender — consistently beat elaborate, richly-specified customer profiles (ICPs) at predicting how real humans would respond. More words about a persona did not mean better results.
• Not all persona combinations of the same size work equally well. If you define a persona using three attributes, which three attributes you pick matters a lot — swapping one attribute for another can dramatically change how well the AI simulates that group.
• Some specific persona combinations are more reliably useful than others — the researchers call these 'alignment bridges.' These are particular Age+Gender or small attribute combos where the AI's simulation stays close to what real humans actually say or do.

Marketing implications

• If you are using AI tools to simulate customer reactions, write focus group responses, or test messaging — use simple personas (age + gender, or age + gender + one more attribute) rather than elaborate customer profiles. Your results will likely be more accurate.
• Stop building long, detailed ICP descriptions for AI persona prompting. A 10-attribute persona is probably worse than a 2-attribute one. Test your current AI research workflow with stripped-down personas and compare the outputs to real customer feedback.
• If your team uses AI to run synthetic A/B tests or predict audience responses before launch, treat those outputs as directional signals only — the AI is likely compressing distinct customer groups into more similar responses than real people would give.

Paper C

Generative AI Applications in Advertising

Yifei Wang — 2026 — Frontiers in Computing and Intelligent Systems

peer reviewed journal article  ·   ·  read now

Key findings

• AI-generated ads were evaluated on four quality dimensions — how good they look, how visually consistent they are, whether they say the right thing, and how creative they are. The paper found current AI tools have room to improve on all four.
• How people feel about AI-generated ads — whether they trust them and see them as authentic — matters a lot for whether those ads actually work. Ads that felt fake or untrustworthy got worse engagement.
• The trust problem is especially big in industries where people already care a lot about whether a brand feels real and honest (think health, finance, or luxury goods). In those sectors, AI ads that feel robotic or generic are likely to backfire.
• User engagement data from Xiaohongshu showed that consumer attitudes and sentiment toward AI-generated content directly predict how much people interact with it — suggesting that fixing the 'trust gap' is more important than just making AI ads look prettier.

Marketing implications

• If you're using AI to create ads in healthcare, finance, or luxury — slow down and run those ads by real people before publishing. The research suggests these are exactly the industries where AI-generated content is most likely to feel off and lose trust.
• When evaluating AI ad tools, don't just ask 'does this look good?' Ask 'does this feel real?' Test whether your audience can tell it's AI-made and whether that changes how much they trust your brand.
• If you're posting AI-generated content on social platforms, track sentiment and comments — not just clicks. People's feelings about the content predict engagement better than visual quality alone.

Apple Podcasts  ·  Spotify  ·  Buzzsprout