I took part in a discussion this week hosted by the legendary Needham analyst Laura Martin around AI and its future in this industry and others. And given the breathless nature of my post-CES LinkedIn feed, I wanted to return to two key points I shared with the audience. Both of which have been identified and named by people far more experienced with all this than I am. The first, and more important, is the Trust Gap. Which is exactly what it sounds like. People have bad experiences with the more popular consumer-facing LLM (Large Language Model) apps and then think “there is absolutely no way I am going to let this thing independently make decisions on $10MM worth of ad buys. Or even $10K.” The second is the “Ouroboros Effect” where AI engines pull data from the open internet on sites that have pulled data from other GPTs, so that Chat is using data created by Gemini and vice versa, and they’re both passing it on as fact, only that data is often a hallucination or a false flag. So there’s that and it’s a big honking problem. Why It Matters Let’s start with the Trust Gap. As much as you hear how “agentic AI” is going to change everyone’s lives, especially everyone buying programmatically sold media, people have to actually, you know, start using it in order for it to have any sort of effect. And if they don’t trust it, then they are not using it. Or they’re using it with an elaborate system of human checks and balances. Which are likely to reveal that they’re not really saving much time and money. If they are indeed saving any time and money. That the industry will eventually get over the Trust Gap is not an issue. It will. The question is when. People get—at some level anyway—that the small confined data sets that the agentic platforms are trained on are very different than the vast ones used by the LLMs and thus more reliable. And then they remember that last week, ChatGPT sent them to a restaurant that closed two years ago and they flinch. Because no matter how many times they hear that the task-focused agentic models are not the same as the consumer-focused large language models, they don’t fully believe it. Or, at the very least, they don’t want to be the test case. Which means that like all of the other changes in the media industry (“we’re going to shift more of our ad dollars to CTV”) the shift to agentic buying systems is going to be a long and slow one, with lots of years of “experimental budget only.” That delay will be compounded by the Ouroboros Effect. Because the data they will be using to place the ads will not be accurate because it was created by one AI, published on the open web, and then scooped up as fact by another AI. That’s one piece and the other is that the LLMs all rely way too heavily on easily gamed sites like Wikipedia and Reddit. As I’ve mentioned previously, there’s a journalist named Ashley Rindsberg who’s done a stellar job of exposing how a group of Bad State Actors have hijacked Wikipedia. Which, to its credit, has done its own internal investigation as a result, and gotten rid of some of the rogue editors. Only the expectation is that this is just the tip of the iceberg, that there is far more of this going on at both Wikipedia and Reddit than we are currently aware of and that we can expect a lot more to come, given that the trick is as simple as hiring a group of people to spend all their time editing Wikipedia entries to reflect a certain POV in the belief that at some time the other side, which is not being paid to do this, will give up. And since LLMs do most of their learnings on the open web, specifically via Wikipedia and Reddit, this becomes a huge problem and raises the Trust Gap yet again, which when combined with the amount of general AI-created MFA (made for advertising) slop out there, makes the data the GPTs are relying on all that much less trustworthy. What You Need To Do About It If you are reading this, do not take away that “AI is just a fad.” It is not, it will change many things and it is not going away—history does not move backwards. Do take away that you should temper your enthusiasm for the degree to which “agentic AI” can change media buying, content discovery or just about anything else related to the business. But understand that there will likely be a tipping point when people start to trust them with larger and larger dollar amounts and the data issues will start to get cleaned up too. And that when this point comes, whether that is in three years or five years or 10, that you will need to be ready for it and act. Or get left on the dustheap of history. (That said, also bear in mind that it is entirely possible that it all just collapses in on itself in a web 1.0 manner, only to be reborn in a different and less fallible formulation a few years hence.) |