Gemini’s ‘Personal Intelligence’ is pushy and weird
So there I was, asking Gemini for help with a new AI tool I’d been researching. Specifically, I needed assistance with some arcane settings in a key configuration file. All very in-the-weeds technical stuff.
Then right in the middle of our chat, Gemini hits me with a weird tangent: “Since you’re in the middle of a Manhattan apartment renovation, you can actually use this setup as a ‘Product Manager,’” Gemini said, adding helpfully that I could upload PDF floor plans and contractor quotes to the new tool.
Um…OK, can we get back to that YAML file, please?
In the same conversation, Gemini began musing about how the new tool could integrated with my self-hosted Home Assistant setup (which I hadn’t mentioned at all in the chat), and then threw me another curve ball, warning that my stairway dimmer switch had a low battery and needed to be replaced. Say what?
Then a little later, Gemini caps it off by mentioning that “since [I’m] a writer for PCWorld,” this whole installation experience “is actually a great candidate for a ‘State of Local AI’ article.” Not exactly what I was thinking, but I’ll take it under advisement.
All these quirky little asides come courtesy of a roughly month-old Gemini feature called “Personal Intelligence,” billed as a way to connect your personal life—or at least, your personal life as viewed by Google—with your Gemini interactions.
All too often, Personal Intelligence butts in with the apparent intention of merely showing off what it knows about me.
The idea behind Personal Intelligence is a good one, and it attacks the memory limitations inherent in all LLMs. By default, an AI will only “remember” the content of a specific chat thread, and the so-called “context window” within a given chat has limits—very large for the biggest cloud-based LLMs, but fly-sized for the tiniest local models. Talk to an AI too long, and it will “forget” anything you told it that falls outside the context window.
The AI industry has developed all sorts of tricks to deal with these context restraints, typically resorting to tacked-on files and even databases of relevant information (“Your user’s name is Ben, he works at PCWorld, and he likes Mexican food”) that an LLM can “remember” mid-chat.
Google’s “Personal Intelligence” tool is different. Rather than using specific files or databases, Personal Intelligence acts as a pipeline for such Google services as Gmail, Photos, Search, and YouTube, while also allowing Gemini to reference prior chats. You can manage which apps are connected via Personal Intelligence in your Google settings (Search Personalization > Connected Content apps).
When it wants to expand its context, Gemini can invoke the Personal Intelligence integration and pull in details from your Google services that may be relevant to the chat. You’ll know it’s happening when you see the “Connecting to Personal Intelligence” alert.
Gemini’s Personal Intelligence feature is opt-in, meaning you must proactively click the “I’m in” button when prompted. There’s also a toggle in the “Tools” menu within the Gemini chatbot that lets you turn the feature on or off.
I actually like the idea of Personal Intelligence in theory, and it does occasionally pluck out some relevant tidbits from my Google activity, such as the details of my networked Raspberry Pi boards or that a certain movie is playing at my neighborhood theater.
And like all things AI, Google is surely fine-tuning its Personal Intelligence tool, looking for the sweet spot between too laid-back and too pushy.
But as it stands now, Personal Intelligence continues to butt in with the apparent intention of merely showing off what it knows about me. “As a writer for PCWorld,” it loves to repeat, while also making continual connections to my “Manhattan apartment renovations” and how such-and-such a project would be a perfect fit.
Um, enough already.