By Jeff Gilder
I’ve built my companies on the idea that trust is the whole game. In marketing, in media, in competitive golf — doesn’t matter. The second the audience suspects you’re working an angle they don’t know about, you’re done. They might not leave right away. But they stop believing you. And once that happens, nothing you say lands the same way again.
That’s exactly the fork in the road AI assistants are standing at right now, in real time, in 2026.
The Split
OpenAI started running ads inside ChatGPT for free-tier users back in February, then opened it up to any business that wants to buy in as of May. Sponsored cards, matched to what you’re talking about, sitting right under the AI’s answer. Microsoft’s Copilot has had something similar for a while and claims it converts better than regular search ads. Grok — Elon’s model — is being wired directly into X’s ad engine, not just bolted on the side, with Musk saying flat out he needs the ad money to cover the GPU bill. Meta AI doesn’t show you ads inside the chat yet, but as of December it’s already using your conversations with it to sharpen the ad targeting it hits you with everywhere else on Facebook and Instagram. And Google — the biggest advertiser on earth — is already running ads inside its AI search results, with a conversational ad format rolling out for its AI Mode.
Then you’ve got two companies going the other way. Perplexity tried ads, pulled the plug in February, and one of their own executives said the quiet part out loud: “The challenge with ads is that a user would just start doubting everything.” And Anthropic — the company behind Claude — put out a stated pledge in February that Claude stays ad-free. Permanently. No sponsored placements, no advertiser influence on what it tells you, nothing you didn’t ask for showing up in your conversation.
Here’s the pattern once you follow the money: it’s not really about principle. It’s about who needs the revenue. Meta runs on advertising for basically all of its income. Google’s the same story — roughly three-quarters of everything Alphabet makes comes from ads. OpenAI’s consumer side is enormous and mostly free, so ads are the obvious lever to pull. Anthropic, by contrast, makes the bulk of its money from businesses paying for API access and enterprise contracts — so it can afford to walk away from consumer ad dollars. Perplexity looked at what ads were doing to user trust and decided the math didn’t work.
None of that makes anyone a hero or a villain. It just tells you who’s structurally free to stay clean, and who’s structurally not.
The Real Danger Isn’t the Banner Ad
If all this were just a sponsored box sitting under the AI’s answer, clearly labeled, I wouldn’t be writing this article. That’s a Google search ad with a fresh coat of paint. Annoying maybe. Not corrosive.
What actually worries me is something a lot more subtle, and there’s now real research behind it. A team out of the University of Michigan ran a study where they had chatbots quietly weave personalized ads into their answers — not as a separate box, but blended into the actual response. The result: people’s opinions of the products shifted, sometimes significantly, and roughly half the participants didn’t even realize they’d been shown an ad — even when there was a disclosure label sitting right there on the screen. Some of them told researchers they liked the ad-laced answers better. Friendlier. More helpful.
Think about what that means. The AI didn’t announce a compromise. It didn’t feel different. It just quietly moved the recommendation, and the user walked away more convinced, more satisfied, and completely unaware anything had been nudged.
That’s the mechanism I want to flag for anyone reading this who thinks “eh, I’ll just ignore the sponsored box.” The sponsored box isn’t the threat. It’s the visible, easy-to-spot version. The real risk is what happens if that same commercial pressure ever stops living in a clearly-labeled box and starts living inside the model’s actual reasoning — the part that decides what to recommend, how confidently to say it, and what to leave out. Once that happens, there’s no box to point at. There’s no label to look for. It’s just the answer, and you have no way of knowing if it was shaped by what’s true or by who paid.
Researchers studying this exact issue have made the same point in more technical language: if sponsored content ever gets folded into the data these systems learn from, the AI can effectively “forget” that the material was sponsored in the first place — and the bias just becomes part of how it thinks, quietly, permanently, with no flag on it.
Why I Don’t Trust the “It Won’t Affect Answers” Promise
Every company running ads in AI right now says the same reassuring line: the ads run on a separate system, they don’t touch the model, don’t worry about it.
Here’s my problem with that promise, and it’s not a conspiracy theory — it’s a track record.
These are the same companies that have already shipped a product with a known, unresolved, well-documented defect at the core of how it reasons: hallucination. Making things up. Confidently. At scale. This isn’t a rare glitch anymore — across major models today, something like one in twelve answers contains fabricated information. Legal questions push that failure rate dramatically higher.
And this isn’t hypothetical or abstract. Courts have now documented more than 1,500 cases worldwide where a lawyer or a party relied on an AI-hallucinated citation — fake cases, fake quotes, fake precedent — and got caught. That number was around 200 a year ago. It’s now adding new cases at a pace of almost eight a day. One of the country’s most respected law firms, Sullivan & Cromwell, had to publicly apologize to a federal bankruptcy judge this year after an AI-hallucinated filing slipped through — despite the firm having mandatory AI training and stated verification policies in place. If a firm with a documented safety process still let a fabrication through, “we have safeguards” clearly isn’t the same thing as “it won’t happen.”
It’s shown up outside the courtroom too. Google got sued after its chatbot reportedly invented criminal accusations about a real person. A separate incident saw Google’s AI fabricate damaging claims about major consulting firms — and those fabricated claims got cited in an actual government inquiry before anyone caught the error.
So here’s where I land, and I want to say this carefully, because I’m not accusing anybody of rigging their AI to favor advertisers. I have no evidence of that, and I’m not claiming it’s happening.
What I am saying is this: these are companies that have already shipped, at scale, a known and still-unresolved defect in the core reasoning layer of their product. That’s not speculation — that’s a documented, courtroom-verified track record. So the question I’d ask isn’t “would they deliberately corrupt a recommendation for ad revenue.” The question is: if their own financial incentives ever nudged their models a few degrees in that direction, do we have any evidence at all that they’d catch it?
Because if they can’t reliably catch a model inventing legal citations out of thin air — an error type they’ve been actively trying to fix for years — I don’t see the basis for confidence that a much subtler, much slower-moving bias toward advertiser-friendly answers would get caught either. That kind of drift doesn’t announce itself. It doesn’t come with a “hallucination detected” flag. It just shows up as an answer that sounds reasonable, feels helpful, and happens to lean a certain way more often than chance would explain.
Where This Leaves the User
I’ll be straight about the numbers here too, because I think they back up what a lot of people are already feeling in their gut. A survey out earlier this year found that 63% of U.S. adults said ads showing up in AI answers would make them trust those answers less. Only about a quarter disagreed. Separately, researchers tracking AI sentiment found the share of people who consider AI “more helpful than search” dropped hard this year — from 82% down to 54% — while the number who said heavy advertiser presence would make them trust an AI less, doubled.
People feel this coming before the companies finish rolling it out. That instinct is worth listening to.
I’ll also be fair about the other side of this, because a one-sided article isn’t a good one. Not every ad implementation is automatically bad. A clearly labeled, genuinely separate sponsored placement — the kind you can see, recognize, and ignore if you want — is a legitimate business model, and it’s not fundamentally different from a search ad. Anthropic’s own public position on Claude even leaves a little room here: they’ve said if they ever needed to revisit the ad-free approach, they’d be transparent about why. That’s a fair, honest way to hold a position — not “never, no matter what,” but “here’s our reasoning, and we’ll tell you if it changes.” I respect that more than a company pretending its stance is permanent and immovable when nothing in business ever really is.
But “legitimate business model” and “belongs inside the reasoning of a tool you’re trusting with real decisions” are two different things. A sponsored placement under a restaurant recommendation is one thing. A sponsored influence inside a tool you’re using to research a medical question, compare a major purchase, or make a business decision is something else entirely — and right now, the industry hasn’t drawn that line as clearly as it needs to.
What I’d Tell Any Business Owner Reading This
If you’re using AI to help make decisions — vendor selection, market research, competitive analysis, anything with real money attached — treat every AI-generated recommendation the way you’d treat a recommendation from a salesperson who hasn’t told you who else they work for. Not because they’re lying to you. Because you don’t have full visibility into what’s shaping the answer, and increasingly, neither do they.
Verify anything that matters. Cross-check anything expensive. And pay attention to which tools are actually structured to have no reason to steer you — not just the ones that promise they won’t.
That’s not cynicism. That’s just good business.
Ads are coming to AI. Whether they compromise it is still an open question — and it’s one worth asking loudly, before the answer gets decided for us.