Apple released the next wave of Apple Intelligence on Wednesday, which brings ChatGPT integration in iOS 18.2 to supported iPhone models. As part of its 12 Days of ChatGPT event, OpenAI held a live stream to demo the new features.
It shouldn’t be surprising to see Google drop the big Gemini 2 announcements on the same day. While it could have been a coincidence, it sure looks like Google was looking to steal Apple and OpenAI’s thunder. Everyone, including Google, knew iOS 18.2 would come out this week with ChatGPT in tow.
As a longtime iPhone and ChatGPT user who still doesn’t have access to the early Apple Intelligence features (I live in Europe), I will say that I’m already envious of Gemini 2. It’s not about the better, faster abilities you expect from the next-gen Gemini models. It’s about the AI agents that Google unveiled, including Project Mariner, which lets you use Gemini to browse the web and complete tasks for you.
Early reports told us that Google would unveil such a tool, which was dubbed Jarvis at the time. I explained that I wanted AI agents to browse the web for me but Google Jarvis wouldn’t be it.
Now that Project Mariner is officially the name of Google’s Chrome-browsing AI, and we know what it can do and how it behaves, I’ll say that it looks amazing. There’s nothing like it in ChatGPT, and it’ll be a while until Siri can browse the web for me.
However, my privacy concerns remain, as Google didn’t cover enough details about what happens with my data once I tell Project Mariner to browse the web for me.
How Project Mariner works
Project Mariner is built on top of Gemini 2. It can browse the web for you and read the information on the screen, including pixels, text, code, images, and forms. Then, it can take commands and perform tasks for you. Simply type the prompts in a Chrome extension, and then Project Mariner gets to work.
The Gemini 2 AI agent can browse the web for you, but only if you keep the Chrome tab open and in focus. That’s a good thing, as it ensures Project Mariner doesn’t browse the web in the background without your knowledge.
The downside is that you won’t be able to work with in a different window (yet), which is what I want from such AI agents. Imagine telling ChatGPT or Gemini to research topics in the background while you work on something else.
Back to Project Mariner, the AI agent will type, scroll, and click in that active tab to complete the tasks in the prompt. It can reason to determine the steps required to perform its job.
You’ll also see a log of performed steps so you know what Project Mariner is doing all the time. You can stop it if you need to. Google said in a blog post that Mariner will also ask for confirmation before taking sensitive actions like buying something on your behalf.
The video above shows the user asking Project Mariner to find contact emails for a list of companies from a spreadsheet. That’s a great example of what AI agents will be good for. We do this tedious research job all the time while browsing the web. It’s also time-consuming, as we need to visit various websites, scroll, click, and save information.
The Project Mariner demo proves that. The job takes about 12 minutes to complete. Google sped up the video for the demo, but it didn’t fake it, as in the early days of Gemini demos.
Google is thinking about security too, and says that it’s building tools in Project Mariner to ignore attacks from hackers:
With Project Mariner, we’re working to ensure the model learns to prioritize user instructions over 3rd party attempts at prompt injection, so it can identify potentially malicious instructions from external sources and prevent misuse. This prevents users from being exposed to fraud and phishing attempts through things like malicious instructions hidden in emails, documents or websites.
However, Google never said anything about user privacy when it comes to Project Mariner.
The privacy issues
We might be in the early days of AI agents, but Google has to make the privacy implications clear to the user from the get-go. It should follow Apple’s lead, which did exactly that when unveiling Apple Intelligence at WWDC. Apple made sure we understand how AI privacy works, and it built strong privacy protections into Apple Intelligence that cover both on-device and cloud AI processing.
With Project Mariner, I have no idea what happens to my data. It’s not just about my prompts training Gemini, which I’d want to opt out of from the start. It’s about how Project Mariner works.
I’d want the AI agent to work on the computer rather than Google’s cloud to minimize data exchanges between the browser and Google’s server. After all, it only has to browse the web for me. The PC should be powerful enough to handle on-device data.
If that’s not the case, Google should explain how it protects the data that reaches its servers and what it does with it.
Also, since the AI would browse websites for me, I’d want to be able to customize how Mariner works. I would want the AI agent to always reject cookies when opening a new page. Also, I wouldn’t want Google to use AI agent browsing data to enhance the profile it has on me for advertising purposes.
Put differently, I don’t want to pay for AI agents with my data, whether it’s Project Mariner or a future AI agent from OpenAI or Apple. I’d rather it was a premium feature that cost money.
Finally, I’ll also address the Google Chrome and Google Search elephants in the room. Project Mariner works in Chrome and searches the web via Google Search. I wouldn’t want any of that. The video above is the full-length demo of the previous clip. It lets you see exactly how Mariner works in Chrome and how it uses Google Search to find the websites.
Say I do end up using Project Mariner; I’d want to be able to use the extension in a different browser. Also, Project Mariner should use whatever default search engine I rely on, which won’t be Google’s.
That might seem like wishful thinking right now. Google wants Gemini 2.0 to be built into various apps and services, including Search, Maps, Docs, and others. That’s great for people who want to make the most of Google’s ecosystem, of course. It also gives Google a competitive advantage over rivals.
However, remember that Google lost a big antitrust case that targeted Google Search. The company is a monopolist in the US. It wouldn’t look good if Project Mariner were to work only in Chrome and only with Google Search. Then again, if Google sells Chrome to settle that antitrust case, that will be a different matter.
Finally, I’ll also say that Google can always claim that the same privacy rules it has in place for Gemini will govern Project Mariner, Astra, Jules, and all the other AI agents it’s working on. But that’s not good enough. We really need to know what happens to our data in chats with the AI agent from the moment we send out the prompts to the moment we get answers.
