According to Inc, senior tech journalist Harry McCracken detailed his 2025 shift from an AI dabbler to a daily user, with his workflow now heavily reliant on Google products like the new Gemini 3 Pro and NotebookLM. He uses AI for research, citing its effectiveness over traditional keyword searches, and for “vibe coding,” building custom apps like a PDF renamer in 20 minutes using Google’s AI Studio. However, he criticizes AI features in established tools like Google Docs and Microsoft Word as intrusive and unhelpful, and notes failures like ChatGPT’s Agent feature being unable to perform its own demo tasks in August and a confusing Perplexity Email Assistant launch. His conclusion is that while the best AI productivity tools are transformative, the industry’s legacy of overpromising continues with rushed, poorly-integrated features.
The Good Stuff is Actually Good
Here’s the thing that gets lost in all the AI noise: when a tool is designed from the ground up to solve a specific problem, it can be magical. McCracken’s praise for Google’s NotebookLM is a perfect example. It doesn’t try to know everything. It just knows your stuff. For anyone drowning in documents, transcripts, or research, that’s a revelation. It’s a focused assistant, not a boastful oracle. And his move to Gemini 3 Pro as his default chatbot shows how fast the landscape is shifting—loyalty is dead when a better model drops every few months.
Then there’s “vibe coding.” I mean, that term alone captures the shift. Eight months ago, it wasn’t even a thing. Now, a non-coder can dream up an app and, with some clever prompting, basically will it into existence. That’s not just a productivity boost; it’s a change in what’s possible for individuals. Building a custom tool to fix a niche problem (like terrible auto-generated PDF names) used to require hiring a dev or learning to code. Now? It’s a 20-minute conversation with an AI. That’s profound.
The Bolted-On AI Tax
But for every NotebookLM, there are a dozen half-baked AI features crammed into software we already use. And it’s exhausting. McCracken nails it by calling it a “mental tax.” Every time you open Gmail or Word now, you’re not just looking at a blank page. You’re looking at a blinking cursor and a blinking AI button begging you to let it write for you. The pressure is subtle but real. It’s like your software is constantly second-guessing your ability to compose a sentence.
Why is this so bad when the dedicated tools are so good? It’s about intent. Google and Microsoft are in a features arms race, terrified of being seen as laggards. So they ship AI integrations that feel rushed, intrusive, and often useless. Remember the Chicago Sun-Times publishing AI-generated book reviews of non-existent books? That’s the same energy. It’s not about making your product better; it’s about being able to say you have the feature. When McCracken found ChatGPT’s Agent couldn’t even get the date right, it just confirmed that quality control is often an afterthought.
A Bicycle, Not a Crystal Ball
The core lesson here is about managing expectations. The most successful uses McCracken describes treat AI as what it is: a powerful, sometimes flawed, tool. It’s a research starting point, not the final answer. It’s a code generator that needs a human to define the problem. It’s a document analyzer, not a mind-reader. This is the classic “bicycle for the mind” idea—it amplifies human effort, it doesn’t replace human judgment.
So where does that leave us? The divide is widening. On one side, you have sharp, focused AI-native tools that are becoming indispensable. On the other, you have legacy software with AI glitter glued on, creating clutter and frustration. The winners in 2026 won’t be the ones with the longest list of AI features. They’ll be the ones, like NotebookLM, that use the technology to quietly solve a real problem without all the fanfare and flashing buttons. Basically, the best AI should feel like it’s working for you, not trying to convince you it exists.
