.
Image: dlyastokiv/Adobe Stock; Cristina Gaidau/Getty Images
Thank you once again for reading Fast Company’s Plugged In. A quick programming note: We will be taking the next two Fridays off. Happy holidays to all, and I look forward to resurfacing in your inbox next year.
For any number of reasons, 2025 has hardly been my favorite year. But if I were to make a list of things that went well, my relationship with AI would be on it.
This was the year I went from being an AI dabbler to a daily user. And while some of that usage still amounts to messing around—hello, Sora!—even more involves tasks that make me more productive. More importantly, it brings me better results, a goal I hold dear. (Sadly, not every AI enthusiast agrees.)
Here, then, is a look at how I’m using AI as 2025 winds down. I covered some of this ground in a September Plugged In. But since I wrote that, the technology has become even more core to my workflow, and my AI A-team has shifted pretty dramatically to Google products for the first time. So a year-end update seemed worthwhile.
First, I’ve finally figured out how to use chatbots such as OpenAI’s ChatGPT, Anthropic’s Claude, and Google’s Gemini as research tools. I remain wary of accepting anything they say as the truth, since AI still has a devious knack for hallucinating fantasies that sound like fact. But it’s dawned on me that I don’t need to take AI at its word.
Starting a research quest with a detailed AI prompt is often more effective than trying to boil it down into keywords of the sort I would have typed into a search engine in the past. And every self-respecting chatbot now provides citations for its work, at least when I ask for them. They lead to web pages written by actual humans, which are far easier to assess than wordage extruded by an LLM.
After spending most of 2025 weaving between ChatGPT and Claude as my chatbot of choice, I was (mostly) wowed by the new Gemini 3 Pro-powered version of Gemini that debuted in November. It’s become my default bot. But the frenzied pace of competition in the category argues against long-term loyalty: I need to spend more time with the new GPT-5.2 version of ChatGPT, which arrived last week.
More than any garden-variety chatbot, I have found Google’s NotebookLM utterly essential this year. Instead of trying to be an expert on human knowledge in its entirety, it just digests files you feed to it. Then it lets you ask questions about them and responds with startlingly useful summaries and citations. They frequently lead to insights I wouldn’t have managed if left to my own devices, and have never mischaracterized anything or otherwise led me astray.
For me, NotebookLM is most valuable as I spelunk through transcripts of the interviews that provide raw ingredients for articles I write. (In the case of our five-part oral history of YouTube, there were dozens of them, about 168,000 words in total.) For you, the source material might be internal documents, white papers, or something else relating to whatever you’re working on. Either way, this free tool, like most of history’s best software, is a bicycle for the mind.
(Disclaimer: I’m not talking about NotebookLM’s best-known feature—podcast-like “audio overview” synthetic conversations based on your sources, which are an astounding magic trick but have never left me feeling smarter about a topic.)
Finally on the AI good news front, there’s vibe coding—coming up with ideas for apps and having AI do nearly all the work of turning them into functioning software. When 2025 started, it didn’t even exist as a thing, at least under that name. Now I can’t imagine working without it.
That started back in April, when I used a vibe coding tool called Replit to build the note-taking app of my dreams. The project required dozens of hours of effort and hundreds of dollars in usage fees. But eight months later, I use the app I created every day, and it still makes me unreasonably happy.
Lately, I have been vibe coding with Google’s AI Studio, which is powered by Gemini 3 Pro. So far, the results have been less quirky and buggy than Replit’s sometimes are, making whipping up my own apps even more irresistible.
Case in point: Last month, I bought a ScanSnap document scanner and soon discovered that its cloud service gave the resulting PDFs incomprehensible names. With Gemini’s help, I constructed a smart PDF-naming utility. It reads the files and renames them with clearer descriptions than I’d write myself. Problem solved, in about 20 minutes.
For all the ways AI speeded my work in 2025, it’s been far from an unalloyed blessing. Notably, all the tools I praise above are newish and AI-first. When existing products are retooled to emphasize AI, the technology often feels bolted on. It’s not just that it isn’t dependably helpful; sometimes, it’s an obstacle to progress.
For example, Google Docs, Microsoft Word, Gmail, and Outlook would all be delighted to compose text for me, a feature that has become as prominent an element of their user interfaces as the 58-year-old blinking cursor. I have no interest in turning that job over to them. And yet I can’t ignore the various icons, widgets, and promos dedicated to these tools, which stare me in the face every time I sit down with these products. It’s an ongoing mental tax levied for alleged benefits I’d prefer to avoid.
advertisement
In other cases, it’s obvious that AI features have been rushed to market without sufficient quality control, as if the bragging rights for having shipped them were all that mattered. I have learned to tamp down my expectations, or even assume that new functionality will perform as advertised at all.
In August, for instance. I discovered that ChatGPT’s new Agent feature couldn’t perform some of the tasks in its own list of things I should try. It was also incapable of reliably determining the current date. A month later, I was intrigued enough by Perplexity’s Email Assistant to briefly spring for a $200-per-month Perplexity Max account. I never got it up and running, in part because Perplexity’s own explanation of its new tool was notably short on, you know, explanation. I might have felt less lost if it had just included a screenshot or two.
Whether or not there’s an AI bubble, the industry responsible for the technology is still in the process of confronting its legacy of overpromising and underdelivering. But with the good stuff getting really good, anything that fails to live up to its own hype—or simply meet reasonable standards of utility—will only look more ridiculous. May the momentum recently seen in AI productivity’s best products continue in 2026 and beyond.
You’ve been reading Plugged In, Fast Company’s weekly tech newsletter from me, global technology editor Harry McCracken.
If a friend or colleague forwarded this edition to you—or if you’re reading it on fastcompany.com—you can check out previous issues and sign up to get it yourself every Friday morning. I love hearing from you: Ping me at hmccracken@fastcompany.com with your feedback and ideas for future newsletters. I’m also on Bluesky, Mastodon, and Threads, and you can follow Plugged In on Flipboard.
ABOUT THE AUTHOR
Harry McCracken is the global technology editor for Fast Company, based in San Francisco. He writes about topics ranging from gadgets and services from tech giants to the startup economy to how artificial intelligence and other breakthroughs are changing life at work, home, and beyond.