A more accurate account of how I use AI for Northfield.org

Part 2 of an occasional series on using AI as a Northfield citizen

I’ve never been interested in being a news reporter. The hard work of creating the zero-to-one draft after conducting research, interviewing, and assembling the facts always felt like a homework assignment or like taking notes at a meeting.

But when blogging became a thing in the early 2000’s, I immediately gravitated toward its ability to offer me a first-person, independent publishing platform where I could write casually about a news event or a public issue, sprinkling in my opinion whenever I felt like it, while—and this is important—linking to the news stories published by actual reporters. (And in case you’re wondering, those em dashes are mine.)

In ten years of blogging at Locally Grown Northfield, I did two in-depth multi-post series. In the six months since the second relaunch of Northfield.org last October, I’ve done six, three of them multi-post. (Examples: #datacenters, #ice, #amesmilldam.)

And what I’m starting to realize is that I’m much more interested and willing to publish those types of posts because of how I’m using AI in the process.

Why am I telling you this? Transparency matters to me personally. Independent journalists are increasingly public about similar processes, though their contexts differ from local civic journalism in ways that matter. You're here for my judgment, and here's what that judgment looks like in practice.

How AI affects your thinking

When I wrote Part 1, I described my AI rules of engagement, some of which didn’t quite match what I was actually doing.

I'm a Wary AI Enthusiast. Here's What That Means

Here’s a more accurate account.

There’s a useful set of terms for thinking about what AI actually does to your thinking. I've been learning about a distinction researchers make between AI-related cognitive offloading, outsourcing, and surrender.

Cognitive Offloading is what you do when you use a calculator or a spell-checker. The thinking is yours; the tool handles execution. You still decide what to calculate, what to do with the result, and whether the answer makes sense. The cognitive work stays with you.

Cognitive Outsourcing is different, apparently, because that’s when the tool does the reasoning, not just the execution, and you consume the result. You get the output without having done the thinking that produced it. This isn’t always bad. We outsource cognitive work all the time when we defer to experts, to institutions, to trusted sources. The question is whether you’re doing it deliberately and whether you can still evaluate the results that come back.

A good example is asking AI to summarize the transcript of a public meeting, which can be quite helpful yet can go awry. The Local Lens AI website published a summary of the March 17 Northfield City Council meeting, focusing on the Ames Mill Dam grant decision. It includes a mistake in the meeting overview (“The council’s decision aims to secure $2 million in funding for the dam’s removal and environmental restoration”). But more importantly, it didn’t mention the issue around the $500,000 Post Consumer Brands ‘donation.’

Cognitive Surrender is a term that gained traction earlier this year due to a working paper titled, “Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender.” From what I understand, it’s what happens when you’ve outsourced so consistently that you stop asking questions and don’t bother evaluating the results of your outsourcing. You lose the ability to notice you’re doing it, and gradually lose the underlying capacity to evaluate what came back, even if you wanted to. Nobody decides to surrender. It happens incrementally, while you’re feeling productive.

Am I still doing the thinking?

I think about where my process sits on that scale. When I’ve spent time digging deep on an issue, formed my own view of what matters, and then ask AI to draft something that represents what we’ve already worked through together, that’s closer to offloading. When I’m stuck and ask AI to suggest a frame before I’ve formed my own instinct about it, that’s closer to outsourcing, and I try to catch myself doing it.

What I think provides some protection: several decades of my living here and paying close attention to the cultural and political climate. That’s not a guarantee against getting something wrong. But it does mean I can usually tell when an AI-generated frame doesn’t fit the actual situation: when it’s too tidy, missing the subtext, or using language that would land the wrong way with the people involved. That’s the layer I try not to outsource.

The ICE/local police series I published between December and February is a useful test case.

The series grew out of a straightforward observation: after an ICE arrest in Northfield last November, residents were shouting at police officers, and nobody seemed to know what the department’s actual role was supposed to be during federal immigration operations. The policy existed, it was in writing, but what it meant in practice was murky.

I spent time with the policy document, the Chief’s KYMN Radio interview, and his Facebook statement before I brought any of it to AI. By the time I did, I had a specific list of questions the public record hadn’t answered. AI helped me research the national context, draft sections of the posts, and sharpen the language. But the questions I sent to Chief Schroepfer across two formal rounds of Q&A, and the decision each time to give him a deadline and publish regardless of whether he responded, were mine.

Across seven posts, I made different editorial calls at different moments. In December, after the Chief answered six specific questions fully and on the record, I ran his complete response without commentary. Getting clarity was the mission; he’d provided it. In January, when drone deployment at the student walkout raised an obvious inconsistency with his stated policy for actual ICE operations, I said directly: “That doesn’t make sense to me.” In February, I pressed on what remained operationally vague: the threshold problem, a documentation gap, a claim about community input I couldn’t verify. Each decision about what a given moment warranted was mine throughout.

In the middle of the series, I ran into Chief Schroepfer at the January student walkout, where I was covering and he was present in an official capacity. We had a friendly chat about the drone and traffic management during the walkout. That’s the texture of doing this work in a town of twenty thousand people. The judgment calls about tone and fairness weren’t abstract. They had immediate real-world stakes. That’s the layer AI didn’t touch.

What it looks like in practice

Here's a screenshot of a representative exchange from the February post on Chief Schroepfer's responses. See the explanation in the caption:

I'm working out how to frame the accountability timeline across the series. Claude, an AI assistant made by Anthropic, maps the sequence and suggests emphasizing that the Chief took a month to answer January questions, which would have sharpened the accountability angle. I decide against it. He'd been dealing with the ICE surge, the student walkout, and the hotel protest simultaneously. The timeline stays in the lede, but without the prosecutorial emphasis. The collapsed sections show Claude's reasoning before it responded.

A more accurate description

So what I’m describing isn’t really “AI drafts, I edit.” It’s more like I use AI to externalize and pressure-test my thinking, then react critically to what comes back. The engagement I feel when I’m arguing with a draft or redirecting it is the cognitive work. My process is more of a continuous back-and-forth, with judgment happening throughout, not just at the end when I’m deciding what to cut. I’m not editing AI’s work so much as interrogating it in real time.

What stays mine:

• What questions get asked and who gets contacted

• What the story’s frame is

• All verification

• Every editorial call about tone, timing, and emphasis

• Publication decisions, including whether to run something at all

What AI handles:

• Zero-to-one drafting (after I've done initial thinking)

• Research and national context

• Structural suggestions to accept or reject

• Pushback that sharpens my thinking

Doing local civic journalism is more interesting to me now than it's ever been. Not because AI does the journalism, but because it handles the parts I was always bad at and never liked, and because it pushes back in ways that sharpen my thinking. It's also making me more interested in getting better at the craft itself.

None of this is settled. The process I’ve described is working well enough that I keep using it, but “working well enough” and “figured out” aren’t the same thing.

Cognitive surrender happens incrementally, which means I can’t rule it out by pointing to the ICE series and calling it evidence.

What I can do is stay alert to when I’m letting AI set the frame, keep running the dual-AI feedback loop that forces disagreements into the open, and keep posting about Northfield matters in ways no model can replicate.

If something in that balance shifts, I’ll tell you.

In the meantime, if you have questions, comments, or criticisms, add a comment or contact me here.

See the latest Northfield.org posts in the archive.