Stress test

Day one of the Florida Legislature's Special Session. Day one of deploying Claude to assist with coverage.

First, thank you. The response to “Welcome to FloridAI” has been a reminder that even after more than 15 years of doing this work, readers are still willing to follow me into a new room. That’s something I do not, and will never, take for granted.

That kind of buy-in is the rarest thing in this business. It is the foundation of everything Florida Politics has been able to build.

Not every new venture has earned it. Anyone remember when we launched “Politics of Pot?” Exactly. The graveyard for Florida Politics side projects is a (figuratively) real place.

This endeavor feels different. And not just because I am being honest about what AI can do for (or to) a business like mine. The thing I kept hearing from readers this week was that the vulnerability is shared.

One veteran lobbyist texted me, “Thanks for doing the AI stuff. I have to admit, it makes me feel old, but also scares the shit out of me at the same time!”

People like that are our audience. This is not for the engineers. Or the prompt evangelists. It’s not for the veteran operator who has watched the game change a dozen times and is wondering whether this one is different. It’s for those used to being savvy on the newest tools and trends who, somehow this time, have shaky knees. And that’s a lot of us.

There is going to be a lot of talk in the AI space about code this and prompt that. There are whole Substacks devoted to top_p settings, model rankings, benchmark wars and any number of other seemingly made-up words to meet this technological moment. That’s not what’s happening here.

As I said Monday, I don’t know this any better than you. I am not an expert on AI. But I have a quarter century of expertise on Florida politics and a publication whose dogged reporters and staff are depending on me to get the next few years right. So now I have a Claude window open on my fourth monitor.

This is simply my experience. The hope is that you can learn something from watching me work through it — including, and maybe especially, from the parts where I get it wrong.

There is a scene in “Fight Club” where it lands on the narrator explaining that nobody had to be sold on fight clubs. The hunger was already in the room. Tyler Durden did not invent the appetite — he just gave it a diner.

That, roughly, is what this week has felt like. Almost no one who wrote me framed AI as a new topic I was introducing to them. What they said, in different versions, was: I have been thinking about this. I have been worrying about this. I did not have anywhere to put the worry. Now there is somewhere to put it.

A caveat, because I do not want to overstate this. I have never been about big traffic numbers. If you ask me whether I would rather be the chef of a Michelin-starred restaurant or the owner of eight Outback Steakhouses, you should know my answer.

“Yes, Chef!”

So it is not as if FloridAI has gone viral. But what I have learned about AI is — scratch that — what I am starting to think about AI is that there is no eddy.

If you have spent any time on a river in a kayak, you know what an eddy is. It is the calm pocket of water behind a rock or a bend, where the current curls back on itself and the river temporarily stops trying to kill you. Paddlers use eddies to rest, scout the next stretch of whitewater, and make a plan before committing to the next move. You can park there. You can think.

But AI has no eddies. The current keeps moving. New models, new capabilities, new entrants, new policy fights — there is no rock to duck behind. The minute you stop paddling, you are downstream from where you needed to be.

So, naturally, I launched this publication on the opening day of a Special Session on the state budget. No eddy.

This is one of the highest-tempo weeks on the Florida political calendar. My team was already drowning in work. And then I decided to drop a bowling ball into the bathtub.

I will spare you the play-by-play of why launching a new publication on day one of a budget Special Session is a decision I would not necessarily recommend to others. It is enough to say that neither the calendar nor legislative leadership care about my learning curve.

When the emails from Katie Betta, spokesperson for the Florida Senate, started to hit inboxes, the stress test was underway.

Betta’s emails — and the parallel stream from her counterparts in the Florida House — are multi-pronged.

On one hand, they public-notice a meeting to come. On the other, they outline an offer one chamber has just made to the other. They report on the meeting that just ended.

That usually means the conferees took the other chamber’s offer under notification, gaveled out, and retreated to their corners like the pugilists they are.

That’s where the reporters of Florida Politics come in.

There was a time when the Florida Capitol press corps was the most feared press corps outside the White House.

Obvious to anyone in this game, that is no longer the case. Only a handful of reporters are still covering the budget with any depth.

Florida Politics has always taken a different view of the process. I view every line in a budget spreadsheet as a story.

We hyper-cover the back-and-forth of conference. We compare one chamber’s offer against the other, flag where the two are in conflict, and detail every line as it closes out.

In the early stages of budget negotiations, that means 45 to 60 budget insight items a day.

So, as it applies to this Substack, the question is: can AI help?

The answer is decidedly yes, though with some caveats.

Without setting up parameters — which drove Drew Wilson crazy — I dropped the spreadsheets from each committee’s meeting into our Cowork space and asked, simply, for Claude to identify five areas where the House and Senate diverged.

Cowork is Anthropic’s desktop tool. You drop files into a shared workspace and Claude reads and works through them. It’s the AI-for-people-who-do-not-write-code version of what developers have been doing for the better part of a year.

The Transportation, Tourism and Economic Development silo — TED, in Florida politics parlance — had just dropped its first House offer. I handed Claude the spreadsheet and asked it to read the document the way I would — tell me where the chambers are actually fighting.

It immediately came back with five insights.

The headline is the Florida Department of Transportation Work Program, where the Senate opened nearly a billion dollars higher than the House. The other four lines are cleaner fights over whether the state should be funding housing affordability and rural economic development at this scale — SAIL Innovative Housing, the Florida Job Growth Grant Fund, Hometown Heroes, and a roughly $94 million rural prosperity package — all funded by the Senate, zeroed by the House.

Five lines, five stories.

In budget sessions past, that kind of scan took too long — or didn’t happen at all. You could argue lawmakers structure these documents to be purposefully difficult to track. Two-hundred-page spreadsheets land at end of business; conference reconvenes at 9 a.m.

Let me be clear about what is and is not happening here. Florida Politics reporters are not using AI to write budget stories.

Every one of those five lines was in the spreadsheet, in plain sight, for anyone willing to read 200 pages of appropriations on a Tuesday morning. The difference is that it took about four minutes to find using AI as a tool.

On day one of a Special Session with multiple currents and no eddy, that is the difference between covering conference and being covered by it.

But here is where the river I am kayaking diverges.

The Florida Politics reporters do not need me to tell them what is interesting. As soon as the spreadsheets dropped and the billions of dollars were proposed, five or six of them began hunting for stories in the subject matter they have been covering for years.

That is not the role I play during budget conference. One of the assignment editor functions I have performed for more than a decade is to quickly scan the budget documents and reach out to 50 to 100 sources — yes, that many — familiar with the appropriations process, reminding them, if you see something, say something … to us.

And boy do you all say something.

Now, AI is adding another stream.

So what did AI get right? And does it matter?

The synthesizing capability is beyond useful. It is, in fact, exactly what some lawmakers have already messaged me they do not want AI doing — identifying the specific differences between the two chambers, surfacing where one chamber may have parked dollars it did not want the other chamber to see.

That work normally takes us hours.

Take the TED offer. The House dropped its first offer on a timeline that, in past Sessions, would have meant only the highest-level staff knew what was inside before the rest of us caught up. With Claude on the spreadsheet, Florida Politics had a story and a text alert out almost immediately. In its opening offer, the House had held the position it staked out during the Regular Session, leaving Ben Albritton‘s Rural Renaissance at zero.

That is not AI writing the story. The reporter wrote the story. AI compressed the analysis layer underneath it from hours to minutes — which, on day one of a Special Session, is the entire ballgame.

So what else did AI get right?

• A funding gap at the Florida Gaming Control Commission — The Senate landed at $40.48 million, the House at $32.34 million — with the disagreement centered on $3.29 million for new FGCC law enforcement squads and $4.5 million for a modernized licensing and enforcement system, both fully funded by the Senate and zeroed by the House. The dispute lands as FGCC enforcement activity has been climbing and Chair Julie Brown has publicly asked the Legislature for more resources.

• The Hillsborough-Pinellas rundown. Home counties of the two Budget Chiefs — Lawrence McClure in the House and Ed Hooper in the Senate — had dozens of local projects threaded through both chambers’ offers. There were shared priorities at different amounts, House-only earmarks, Senate-only items. It’s the sort of cross-chamber disparities that normally take a Florida Politics reporter half a day to find, yet with the help of AI, it took mere minutes.

• Breadth across silos. Claude helped Florida Politics move on the House zeroing out Ron DeSantis’ Florida State Guard and the Senate’s opening environmental offer on Everglades wastewater funding — among other beats. The standout was the Tampa Bay Rays stadium story, where Senate Appropriations Chair Hooper publicly paused the $50 million Hillsborough College request tied to the Rays ballpark plan. Other outlets, including the Tampa Bay Times, ran with the public version. Florida Politics sources said the money would probably still arrive — routed through the Public Education Capital Outlay report. That is not the story Claude wrote. That is the breathing room Claude gave a reporter to make one more phone call.

But, as excited as I was about the stories AI was helping to produce, my colleague Drew took a more reserved approach.

Per Drew:

What Florida Politics ran into on Day One were not simple errors from Claude, but a misunderstanding — by us — of what the system actually is.

The problem is not whether the model can understand politics. The problem is whether the model can understand this moment in politics with the same prioritization and nuance that the minds at Florida Politics naturally apply.

A language model can absolutely understand what a Special Session is. It can understand that a Regular Session failed. But publication-grade political reporting requires another layer entirely: the constantly shifting hierarchy of what matters right now.

That hierarchy is built from accumulated institutional memory, live reporting, sourcing, procedural awareness and instinctive editorial weighting developed over years of covering the same ecosystem every day.

Replicating that inside an AI system is not as simple as “adding context.” By the time enough chronology, negotiation status and political nuance are supplied to reliably prevent false assumptions, the humans involved are already doing much of the interpretive work themselves.

And when that contextual scaffolding is incomplete, the model does not pause and admit uncertainty. It attempts to complete the pattern.

What makes that dangerous is that the missing context often does not produce factual errors. It produces editorial ones.

A read that made sense at 2 p.m. during budget conference may be misleading by 3:30 p.m. and wrong by 5 p.m. The risk is not simply that AI will invent a fact. It is that the model will inherit an outdated frame, smooth over the uncertainty and produce copy that sounds publishable even though its editorial judgment is already behind the story.

And if Drew was circumspect of the AI modeling, Janelle, who describes herself as our resident AI critic, saw real flaws in the data provided and conclusions reached by Claude.

Per Janelle:

As the resident AI critic, my take on Day One of this artificial intelligence experiment is not as rosy as Peter has portrayed. It is true that AI has proven a useful tool in quickly distilling data, comparing the House and Senate side-by-sides and identifying gaps. Indeed, that useful tool can save as little as minutes, and as many as hours.

However, our current applications’ benefits stop there. While there was plenty of data analyzed correctly — the House offered this, the Senate offered that — there were also critical errors in that analysis that took a human eye to catch. In one of them, the AI itself admitted error.

When asked broadly to identify gaps in funding offers between the House and Senate in agriculture and environment related silos, Claude provided inaccurate information I caught. When asked to fact check itself, Claude fessed up to the error and attributed it to, ironically, reading the wrong line. That’s the sort of error AI enthusiasts claim the technology eliminates, not perpetuates.

In that same analysis, there appears to be what has become known as “AI hallucinations;” that is, there was information provided in the analysis that simply did not exist in the document Claude was asked to analyze. In this case, the information Claude provided signaled there was a story, but when removing information provided arbitrarily, it suddenly became a “there is no there, there” kind of a situation. In this case, time was wasted.

Finally, there is the issue of writing. While we are not using AI to create content, we are testing it just to see what it can do. Here is where reasonable minds can disagree. As a self-described hater of AI (in most but not all applications), I find AI “writing” to be fundamentally terrible. It creates weird analogies and doesn’t seem to understand grammar. While it was clear conversational was the goal, attempts to cobble together stories using AI included numerous incomplete sentences, several run-on sentences and prose that felt forced. While that can all be edited, I continue to ask myself the same question – could I have written this faster than I edited it? The answer for me, though admittedly not for all, is a resounding “yes.”