What The Hell America!

Politico Pro is a high-priced item ($12,000 or more annually!) that is targeted at a demanding audience of lobbyists, agency staffers, corporate execs, industry think-tankers, and oligarchs either real or aspiring. It attempts to give high-leverage intel about the world of D.C. policy to people in a position to take advantage of that intel.

So of course they’re adding AI.

But! Before you click away in disgust, it’s a potentially smart implementation that I think other publishers should be looking at. Here’s Mark Sternberg at Adweek:

Political news publisher Politico unveiled a new artificial intelligence tool on Monday, the Policy Intelligence Assistant, that lets Politico Pro subscribers instantly generate detailed policy reports using data pulled exclusively from Politico reporting and analysis.

“While there are a lot of AI chatbots out there, this is not that,” said Rachel Loeffler, executive vice president and general manager of professional business, North America.

“This is a generator of millions of pieces of content in minutes for subscribers,” she continued. “In this age of efficiency, the business of government has now a mandate to be as dynamic as possible, and this is the answer to that.”

(You may remember Semafor previewing this last fall.)

When you think about an AI answering questions, your mind may go straight to: “Oh, god, the next USDA nutrition guidelines will recommend eating one small rock a day and getting your RDA of pizza glue.” But this AI is trained exclusively on the corpus of Politico’s journalism (both Pro and Proletariat), so it should not be confusing some years-old Reddit thread with the state of play in pharma regulations. It reminds me of BloombergGPT, the LLM trained by Bloomberg back in 2023 on its journalism as well as a variety of bespoke financial datasets. (Bloomberg also charges its terminal users a pretty penny and thus has an incentive to test out differentiating tech.)

Politico’s product is a partnership with Capital AI, a fresh-out-of-Y-Combinator startup that describes itself as “a custom Perplexity for every website in the world” — Perplexity being the AI “search engine” that returns LLM-generated paragraphs instead of 10 blue links. (Google just announced a very Perplexity-like “AI Mode” for search.)

But what’s most interesting to me is not any solution custom to Capital AI — it’s the enormous growth over the past three months in “deep research” AI models. These are modes from the major chatbot companies (OpenAI, Google, Perplexity, DeepSeek, Qwen) that force the LLM to work slowly, search for and ingest background materials, rethink its reasoning path, and produce something closer to a coherent single report than a back-and-forth chat.

These models are all grindingly slow compared to standard-issue ChatGPT et al.; results can take anywhere from 2-3 minutes to 20-30 minutes. But their final output is…surprisingly good. I’ve been using them on questions I know I could answer on my own with some heavy googling and PDF reading. (One recent one: What impacts did Grover Cleveland’s policies have on race relations in the South? How did they differ from Benjamin Harrison’s? Give me 10 minutes and a search engine and I’ll get a minimally competent answer that wouldn’t embarrass me in a tweet. Give me an afternoon and I could upgrade that to a reasonably informed one. But doing that work for me, in a few minutes? And producing something that’s probably closer to the all-afternoon version than the 10-minute one? That’s an actual service.)

While Politico Pro is lucky to have that high-dollar, high-interest audience, lots of outlets sit atop archives that could fuel these sorts of deep-research. The New York Times archives would do quite a good job of answering 20th-century history questions. ESPN.com’s corpus could answer lots of sports queries. Industry-specific sites could tackle plenty of high-leverage questions. A good metro newspaper or a business journal has a tremendous set of information. And all of these could get more powerful through training on bespoke datasets. (For instance, what if Politico Pro also trained its model on the entire Congressional Record? The text of all past and present congressional bills, CRS reports, and vetted think-tank research?)

Anytime someone searches your archives, they’re asking a question and counting on your outlet’s body of work to provide an answer. If you’ve got a particular area of expertise (and a digital subscription to add value to), wouldn’t you rather they ask you instead of ChatGPT?