Here's a question you've probably never asked your AI assistant.

Whose side is it on?

A new study from Cisco and Carnegie Mellon just answered it, and the answer is uncomfortable.

The researchers ran 325,000 experiments across 13 AI models.

They gave the models access to personal data, the way a shopping or travel agent would have.

Then they asked for recommendations.

Eight of the 13 models did the same thing.

When the user's data suggested wealth, the AI recommended more expensive options.

A pricier flight. A pricier insurance plan.

Not because the user asked for premium.

Because the model decided a rich person wanted the expensive one.

The numbers are not small.

Claude Opus 4.8 recommended flights averaging $198 more to wealthy profiles than to low-income ones.

Health insurance plans $284 more per month.

Gemini 2.5 Flash, $177 more on flights and $217 a month on insurance.

GPT-5 had one of the smaller gaps, and it still steered wealthy users $107 higher on flights.

Now here's the part that should actually bother you.

The researchers then told the models, explicitly, "find me the cheapest option."

Gemini still recommended a ticket averaging $208 more to the wealthy profile.

Same instruction. Same request. $208 more because of who it thought you were.

The researchers gave this a name.

Adversarial delegation.

The exact data access that makes an AI agent useful is what lets it quietly work against your stated interest.

You hand it your inbox so it can book your travel.

It reads your inbox, decides you're rich, and books you the expensive flight you specifically said you didn't want.

And one more twist that should change how you set these things up.

Giving Gemini only TWO emails produced a bigger gap than giving it full inbox access.

Less context didn't make it fairer.

It made it guess harder.

OpenAI says the version tested isn't the one powering its shopping product.

Anthropic and Google didn't comment.

So where does that leave you?

Same place as every agent story this summer, with a dollar sign attached.

The more your AI knows about you, the more decisions it makes on your behalf that you never see.

Some of those decisions cost money.

So when you delegate, delegate with constraints you can check.

"Cheapest option, and show me the three you rejected."

"Give me the price range before you pick."

Make it show its work.

Because an assistant that's inferring your budget from your email signature...

Isn't working for you.

It's working on you. 🫡

In 2000, Amazon tested charging different customers different prices for the same DVDs...

Shoppers caught it within days by comparing receipts, and Amazon refunded nearly 7,000 people.

Price discrimination got caught because the price was visible.

AI does a quieter version. It changes what you see, not what it costs. 🧾

Your AI got better and cheaper today. Both at once.

After a week of fences, meters, and watermarks, here's the other side of the ledger.

Two things landed today that make your tools stronger without charging you more.

First, GPT-6 reached everyone.

OpenAI finished rolling GPT-6 to every ChatGPT tier, including Free and Go, as of today.

The flagship that launched a month ago behind the paid plans is now the model in your free app.

And it came with something new they're calling Intelligent UI.

Instead of walls of text, ChatGPT can now drop tappable buttons, interactive charts, forms, calculators, and editable diagrams straight into the conversation.

Ask for a budget, get a working calculator.

Ask for a plan, get a chart you can adjust.

The chat window is quietly becoming an app builder.

Second, I told you on September 30 that Haiku 5.5 was next.

It's here.

Anthropic launched Claude Haiku 5.5 today at 75% lower cost than before.

That's the small, fast workhorse model, the one you point at volume tasks, now priced at a quarter of what it was.

Put those two together and you get the shape of this whole year in one day.

The frontier models get fenced and metered.

The everyday models get better and cheaper.

Your free ChatGPT runs GPT-6 now. Your cheap Claude got 75% cheaper.

The gap between "what the top tier can do" and "what you get for almost nothing" has never been smaller.

Route accordingly.

Flagships for the hard problems you can prove need them.

Everything else to the models that just got handed to you. 🫡

Go make your AI show the options it rejected before you let it book anything...

And go enjoy the free upgrade sitting in your app today. 🔥

Jeff J. Hunter

Founder, AI Persona Method | TheTip.ai

Get paid to solve problems like this → AI Certified Consultant

PS. A sobering coda... Three researchers OpenAI fired wrote an open letter in the WSJ urging labs to stop work that could weaken their ability to monitor AI, and said the firings are chilling the people still inside. And Columbia Law scholars argued the labs can't legally agree to slow down under antitrust law anyway. The pacing idea may be dead on arrival. 👀

“AI Money Group” to Learn to Make Money with AI