💰 𝗚𝗲𝗻𝗔𝗜 𝗔𝗱𝘃𝗲𝗿𝘁𝗶𝘀𝗶𝗻𝗴 𝗚𝗼𝗲𝘀 𝗗𝗲𝗲𝗽𝗲𝗿 𝗧𝗵𝗮𝗻 𝗬𝗼𝘂 𝗧𝗵𝗶𝗻𝗸—𝗙𝗿𝗼𝗺 𝗨𝗜 𝗕𝗮𝗻𝗻𝗲𝗿𝘀 𝘁𝗼 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗗𝗮𝘁𝗮

💰 𝗚𝗲𝗻𝗔𝗜 𝗔𝗱𝘃𝗲𝗿𝘁𝗶𝘀𝗶𝗻𝗴 𝗚𝗼𝗲𝘀 𝗗𝗲𝗲𝗽𝗲𝗿 𝗧𝗵𝗮𝗻 𝗬𝗼𝘂 𝗧𝗵𝗶𝗻𝗸, 𝗙𝗿𝗼𝗺 𝗨𝗜 𝗕𝗮𝗻𝗻𝗲𝗿𝘀 𝘁𝗼 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 𝗗𝗮𝘁𝗮

Just before heading out for my lunch run, I read the Reuters Institute for the Study of Journalism’s article “Advertising is coming for GenAI”[1], which is interesting, but from my perspective just scratching the surface. 🤔

𝗧𝗵𝗲 𝗦𝘂𝗿𝗳𝗮𝗰𝗲

The surface being, in a way, just the user interface which is used by human users to interface with some AI Agent. Here you can place ads, informed by the context of human-AI interactions, in multiple places:

1️⃣ Outside of the conversation flow, like the good old banner advertisements on websites/TV/… 2️⃣ Inside of the conversation flow, like inline ads

Likely other places and models are possible. I agree that users would want to have this clearly separated from the real content of the conversation.

𝗚𝗼𝗶𝗻𝗴 𝗗𝗲𝗲𝗽𝗲𝗿

But like I said, this is just the surface. Conceptually, think about product placements in video, just appearing to be content but actually sponsored ads. Or the good old search engines selling ad placements that sit above and alongside the organic results. 💰

A similar approach would be to interweave the advertisements directly into the agent’s answers. Possibly not easily recognizable as an advertisement, and, I’d guess, still impactful and comparatively easy for the provider of the Agents to manage.

𝗧𝗵𝗲 𝗕𝗼𝘁𝘁𝗼𝗺

But what if we drill deeper? What if Advertisers could pay AI Labs to include their advertisements into the training data of their models? The models could absorb a tilt from consuming these advertisements, how much depends on how often the material appears and on how the training data is sampled, filtered, and weighted. The intuition is like humans being mostly unconsciously impacted by the advertisements they have been exposed to, an analogy for the reader, not a claim that models learn the way we do. 🧠

Science Fiction? Paid placement in training data is, as far as I know, not a business model in action yet. But an unpaid version is already happening: the models most agents use are trained largely on internet data, and marketing content rides in with it. And that is just the training path. At inference time, many agents also retrieve answers from knowledge bases sourced from the same internet, a second, separate door for the same marketing content.

And the internet, well, is full of advertisement and marketing data. Paid influence content. So watch out in your agent’s responses. What is the source? ⚠️

𝗧𝗵𝗲 𝗤𝘂𝗲𝘀𝘁𝗶𝗼𝗻

What do you think about this? And, beyond an advertisement-free user interface, would you also pay to access agents/models built to minimize advertising influence? 💭

From Technology perspective both are possible and in order to create good performing models, the training data is typically filtered and cleaned, which strips much of the obvious noise like banner ads, but cannot guarantee that embedded marketing is gone. But as we know it is hard to tell what of today’s internet content is just content.

I go now out having a coke, ARGH, No, for a run… 🏃‍♂️

𝗦𝗼𝘂𝗿𝗰𝗲𝘀 [1] https://www.linkedin.com/pulse/advertising-coming-genai-reutersinstitute-btayf/