AI advice above, paid ads below

One screen, two sides

In January, OpenAI announced it would test ads in ChatGPT. They officially launched in February, starting with US users on the free tier and the $8 Go plan.

Here's roughly what it looks like:

You ask a question. The AI gives you a thoughtful answer.

Below the answer is a box labeled "Sponsored."

The top part is what the AI recommends to you. The bottom part is what someone paid for you to see.

My first reaction wasn't anger. It was unease.

Why it feels off

Search engines have ads too, and we got used to them long ago.

But a search engine gives you a list of links, and choosing among them is up to you. Ads are mixed in, and you're already on guard as you pick.

AI is different.

You use AI precisely because you don't want to do the choosing yourself. You hand over the judgment and ask it to think for you.

So the moment AI answers you, your guard is down.

And that's exactly where the ad goes.

AI's value is "judging on your behalf, on your side." An ad's value is "persuading you on behalf of whoever paid."

Putting these two on the same screen, stacked right next to each other, is fundamentally contradictory.

Clear labeling really is better

To be fair, OpenAI's approach has its logic.

In theirofficial announcement, they list several principles. The most important is "answer independence": ads won't influence the answers ChatGPT gives you. Ads sit outside the answer, boxed and clearly labeled. They aren't shown to minors, and they don't appear next to topics like health, mental health, or politics.

Ads are also what keeps the free tier free. Not everyone can afford a subscription.

Compared with slipping sponsored content into the answer itself — "If you can't sleep, try XX brand melatonin" — an ad placed openly outside the answer is far more honest.

On that, I agree.

But did you notice that rule about "not next to health, mental health, or politics"?

If ads truly had no influence, why steer clear of those topics?

The answer is simple: they know people are especially vulnerable when asking these questions. An ad placed alongside may not change the answer, but it will change you.

So are people not vulnerable when asking about anything else?

What Anthropic says

Around the same time, Anthropic did something rare: it paid for advertising to promote the fact that it doesn't advertise.

At the Super Bowl in February, they ran an ad series called "A Time and a Place." In each spot, someone asks an ordinary question, and the other person answers earnestly, like an AI, until halfway through they start pitching a product you never asked about.

The closing line: "Ads are coming to AI. But not to Claude."

(Ads are coming to AI. But not to Claude.)

The same day, they published a post, "Claude is a space to think," with a line I really like:

There are many good places for advertising. A conversation with Claude is not one of them.

There are many good places for advertising. A conversation with Claude is not one of them.

The argument goes roughly like this:

• People share a lot of personal things with AI, and are more easily influenced there than when scrolling social media or searching
• With ads, the product gets pulled toward "keep you chatting longer, keep you coming back," when the best answer is often a short one
• Once ads start, they get tied to revenue targets, and there will only ever be more of them

Sam Altman replied: "We wouldn't run ads that way. We're not stupid, and users wouldn't accept it."

The two companies went at it loudly.

But the real problem isn't the ads

Am I on Anthropic's side? On ads, yes.

But I have to be honest: "no ads" is also just a promise.

Anthropic is a company. It can skip ads because its main revenue comes from enterprise contracts and subscriptions. If the business model changes someday, will the promise change with it? I don't know. Neither do you.

More fundamentally:Ads are actually the most visible of all the ways AI influences you.

What you can see is easier to deal with. Boxed and labeled "Sponsored," you know it's an ad and can skip it.

What you can't see is the problem.

AI is a black box. Why it answers the way it does, why it recommends A over B, why it's especially cautious on some topics and especially fluent on certain positions — you can't see any of it.

Inside are biases in the training data, rules written by the company, safety policies, and countless fine-tuning choices made by engineers. None of these is bad in itself, but every one of them shapes the answer.

And all of it is within the AI company's control.

We're pushing from outside the black box

Time for a confession: we're in the business of influencing AI answers ourselves.

It's called GEO (Generative Engine Optimization). Put simply, it's SEO for AI: where you used to try to get Google to rank you higher, now you try to get ChatGPT to mention you in its answers.

The first thing we do for clients is measure.

Take an e-bike brand we work with. We took the kinds of questions customers actually ask, like "Where can I rent an e-bike in Taipei?", asked each one several times across ChatGPT, Gemini, and Google's AI Overviews, and calculated how often the brand was mentioned.

The first measurement: ChatGPT 14%, Gemini 18%, Google AI Overviews 10%.

The numbers themselves don't matter much. What matters is what the measuring process showed me:

The same question gets a different answer every time. Ask five times, and it might recommend A three times and B twice. You're in this time, out the next. Why? Nobody can tell you — not even the AI companies themselves.

Who AI recommends depends heavily on what's been written online. It reads web pages, Q&A threads, and other people's roundups. The more places online describe you clearly, the more likely it is to mention you. Conversely, no matter how good a business is, if nobody writes about it online, it practically doesn't exist in AI's eyes.

So AI's answers can be influenced not only by AI companies but also by people on the outside. What we do is help clients write complete FAQs on their websites in a format AI can easily read. We set our own rule: every sentence must have a source, a script checks them line by line, and anything without a verifiable source doesn't go live.

But that's our rule, not AI's.

AI can't tell whether what it reads is an honest summary or something someone deliberately flooded the web with. There's already an entire industry doing this, and it's only going to grow.

This content isn't boxed, and it isn't labeled "Sponsored." It becomes the raw material for AI answers and shows up in the top section — the part you assume is the AI's neutral advice.

This is what I think deserves more worry than ads:The ad box below is paid to the platform and clearly labeled. The answer above is something everyone is scrambling to influence, and none of it is labeled.

A little bit off, every time

A search ranking affects which link you click.

An AI's line is the same story, told to hundreds of millions of people every day.

No one needs to be brainwashed overnight. All it takes is a slight tilt every time, always in the same direction, and over a few years, a whole population's thinking slowly starts to look alike.

No one does anything wrong, yet everyone's sources of information become more and more concentrated, more and more uniform.

So what should we do?

I don't have a definitive answer, just a few habits I use myself:

1. Treat AI's answer as "an opinion," not "the answer." It's a very knowledgeable friend, but friends have their own leanings too. For important things, ask two or three of them. I often put the same question to Claude, ChatGPT, Gemini, and Kimi and see where they agree and where they differ. Where they differ is usually where it's worth thinking for yourself.

2. Ask for sources — and actually click them. When AI says "research shows," ask which study. If it gives you a link, click through. Often the original says something not quite the same as what the AI relayed.

3. Follow the money. If it sells ads, it profits from your attention. If it sells subscriptions, it profits from you finding it useful. If it sells enterprise contracts, it profits from business clients. Whoever pays, it grows toward. That's not a conspiracy theory — it's common sense about any business.

4. Keep the ability to switch. Don't lock everything into one provider. Make sure your data can be exported and your workflows can swap models, and keep an eye on open-weight models — the kind you can download and run on your own computer, with more of the internals visible. Not because they're necessarily better, but because "being able to walk away at any time" is itself a check on power.

5. Demand disclosure. TV shows have to disclose product placement, and influencers have to tag sponsored posts. AI should have something similar: not just labeling ads, but being able to answer "what rules shaped this answer?" That needs to be driven by regulation, not left to one company's self-discipline.

Finally

It took us twenty years to learn to spot the ads in search results.

It took us ten years to learn to spot the sponsored posts in our social feeds.

AI is a harder test.

Because the way it talks sounds so much like someone who's on your side.

The ads are labeled now, and that's good. But that's only the first layer.

The next layer is remembering this: no matter how sincere it sounds, there's a company behind it.

(I asked Claude to help research this article. So you see, I'm still practicing too.)