speaker: Ulrike Langer

topic: Wie KI in den USA den Lokaljournalismus beflügelt

I am going to tell you a little bit of gloom and doom right now. So some of you have already asked me how many newspapers have or have died in the US. There's more than 3000 since 2005, and that is more than 1/3.

And of the 6000 surviving local papers, more than 80% of them are actually small weekly papers, so probably not the biggest influencers in the local sphere where I live in the Meta Valley, we actually have one of those still functioning papers, but this is an old model and the future is definitely not there. so we can talk about the gloom and doom, but what

I actually want to talk about is about the oases of hope, which also exists. There are many, many new new.

new sites and old, older legacy newspapers who are innovating, who are doing interesting things with AI and today I want to show you 4 local news sites, and 8 lessons to learn from them.

So the first one is the Minnesota Star Tribune.

And they have set up an AI lab.

They are not doing this for themselves. They're doing this for the whole region around Minneapolis and for the for the state of Minnesota.

it launched, it's not very old. It only launched at the end of 2024 with a grant from the Landest Institute, that might sound like a lot of money. It's actually not. It's just to get initiatives off the ground and that is actually a big misunderstanding here in Europe when you hear nonprofit journalism.

So the Minnesota Star Tribune is actually for profit, but they still got this grant. When you hear non-profit journalism that does not mean that they don't need to earn money. They get some money to get off the ground and then they need to innovate and find a business model with this. So, 3 members in this AI lab, they worked to build up this, this AI lab, and one of the first projects they did is

Let us get more value out of our archive.

So, Aaron Pilloer, he's been making his way through several media in the US. He just last week he actually left the Minnesota Star Tribune. He's now with Chicago Public Media, he is one of the key people who built up this lab and said let's focus on the archive. So what they've been doing there is, that was too fast.

First they built

A system that can put a location-based stories in context with each other. For example, let us find all the neural neuralgic intersections that are prone to accidents and let us put them all into, into context so that we can see where do these accidents always happen, and the model that they built is able to find these stories, even if they have only been in print because

it was 50 years ago, or if the street names aren't even mentioned, if the, if the neighbourhood where everything is happening, if there aren't even mentioned, so the, the model can put it into context and find the bigger stories behind the little reports.

But that wasn't even their biggest success. The biggest success is about cooking recipes. So they basically did, you all know the New York Times cooking app, huge success has more than 1 million paying subscribers. I'm one of them. I love that app.

But it took the New York Times.

Years to build it because that was before generative AI and it was hugely expensive. Now the Minnesota Star Tribune, they did the same thing, Minneapolis has a very curious cuisine that's very regional, some people outside of Minneapolis doesn't, don't like it very much, but

Minneapolis is proud of its cuisine, the recipes, they are handed down from generation to generation. It's very immigrant focused, so there is a local element to this, why this is important to them. So the system, the new system found all these recipes from the decades and basically built, they built an app for this, and they did this with a fraction of the cost that it costs the

Times and what took the New York Times years, they did this in a weekend.

That's, that's the power of AI, so if there's one wow example, I think this is this one.

So they don't want to stay where they are right now. They have their target is to help other regional publishers who have limited tech experience, so one of the next step is to take this model and make it more accessible for people in the news industry who do not have the tech expertise, so it needs to be more intuitive, require even less coding. There is still some coding.

evolved, make it completely no code drag and drop, they're working with the Minnesota newspaper association and they're expecting all of this right now, it's still internal we're expecting all of this to go public by the end of the year. And then with the potential to expand even beyond the state of Minnesota Minnesota.

Next example is archive again, Philadelphia Inquirer. Now, what is special about this example is that the Philadelphia Enquirer is almost 200 years old at is one of the oldest newspapers in the US in a city on the east coast that is steeped in history, this is where the Constitution was signed, um.

So if you have a history that goes back.

So far in a city that is so historic for the US, then, of course, you have an archive that is immensely valuable.

If only you can find all the stories in there.

Mostly they are, it's fragmented, if it's 150 years ago, how are you even going to find this if you have to go and look through boxes in the basement.

So what they did is they built an archive system that is so they're going back in time right now, they have gone from present to the 1970s with the goal to go back all the way to their ancient history.

And they're using this internally for their reporters, so, to help their reporters.

Find connections between stories, old and new, and for the moment they're not planning to make their archive accessible to the for the public because of course if you have history that goes back so far, lots of stories have to be put into a new editorial context, the earliest days, there was still slavery in the US. this was, this was before the Civil War, so we

We talk and we think differently about lots of things in society these days, and it was when they were founded, it was 100 years until women could vote for the first time.

So you cannot just put all this old context out there to the public and say, do with it what you want. It has to be put in context.

And

So what they're focusing on instead is helping their own reporters and learn from that, see where the weak points are prove it, make it foolproof, and then go out to public.

The system also understands not only geographical context, I mean that's the same as with the Minnesota example also, what they call smart context is when you when you go back in time, you need to be able to understand which

Periods in history are exactly when, for example, a reporter wants to write about the Reagan and administration, and then the system knows, oh, this was from January 20th, 1981 to January 20th, 1989, even if it is not said. So you need the system to understand this kind of context.

What's also important when they started this, they started with the newsroom asking reporters. What is, what are your biggest pain points like what, what would you like us to be easier for you if you're researching a topic. So with these questions, they then went and started to develop the system. They did not build it and at a later stage, asked the newsroom, how do you like this? But those questions.

was the very beginning.

Yeah, I already mentioned all of this, um.

User needs is the only point I have not mentioned yet, so not only for minimising the risk, but also becau because they asked the public, having access to your 200 year archive was nothing somebody in the audience ever said they will probably like this once they have the chance, but right now they want someone who tells me what's the most important story today, so this is why another reason why they

didn't go straight into public.

Next, 3rd example is the Baltimore Banner. This time we're not talking about the archive. We're talking about a problem of growth. So, the Baltimore Banner is a digital only on news publication, they're only 3 years old. They're already the biggest local media company in Baltimore, and they have the challenge that they were, they are very successful. They are growing very fast with their membership.

model, they started with some foundation grant money for philanthropy money, but they're actually,

The the they're pulling in revenue 13 million $13 million last year and they're expecting to be profitable next year.

so their challenges, they went from publishing 2 to 5 articles per day initially, now they are publishing 30 to 45 articles per day. That is an explosive growth and their existing CMS is in no way compatible to handle this, this growth. So they needed a new taxonomy system.

And they actually did this by what I'm a little bit too fast here.

They did this by having different large language models compete with each other. So the, the winner was Claude Sonnet with an 85% rate of success, and when they went back and added a second category of taxonomy, then it was 93% correct, which is of course a huge

huge, um.

Progress from what they had before.

Somehow I'm a little confused here, but OK, doesn't matter. Um.

What stood out is that as a side effect, they discovered that it's not all about statistics, so the, the example which they fed their, their taxonomy system with to test the whole of it was a baseball example. the Baltimore Orioles, which is a beloved baseball team, um.

And they thought, well, we just need a system to better find all the statistics, what they found out is that's not what the public actually wants. They want more human interest stories, so the stories about the players give us more background about them, what do they do when they're not playing baseball. this is what Emma Patti, the managing editor, a few weeks ago at a conference said, the

experiment told us that we need to focus less on stats on more on people behind the plate or behind the baseball plate. So that's what they're doing now, writing, freeing up their, their reporters to focus more on people stories. People are interested in people's stories. And then in the next step they want to take this example and also use it for other data-driven stories like local economy, um.

real estate, anything that is basically data driven.

Last example, village media. So, this morning we already heard about how Amboga Ablett is going in this direction, and a lot of this sounded very familiar to me with one difference, village media, they are already doing this, they started 11 years ago with one local news site in Canada. Now they have 34 sites still mainly in the province of Ontario.

And they are so successful that they are now licencing the AI technology to other publishers, 120 of them.

And they, they are profitable. They make money from, from advertising, but also from licencing.

And they think of themselves as

Community focused first, and then a media company 2nd. How are they using AI, they are.

They, use AI for surveys to find out what the community actually wants. And then into this community, the system suggests community polls, it draws the most valuable feedback from communities.

And

So that the newsroom is data driven informed about the user needs.

they have comment engines in their communities, and these communities are, think of them as a as a mixture of

Reddit, Facebook, and local neighbourhood communities, that that's, that's the base of what they look like.

often when there is not enough engagement, the AI suggests what kind of comments to write to get the conversation going because the first comment often sets the tone. If the first comment is something constructive, the discussion will go on in that manner. If it is something mean spirited, it might, it might continue mean-spirited or just die, and that's what they don't want.

What they also make money with is they sell not only ads, but they sell their data anonymized, so it is data conform, data privacy law conform. they call this, our gold mine dashboard, so it's an AI powered tool that advertisers have access to post jobs into the right communities or to post event.

notices in the right communities. And the fourth pillar is a so-called community leaders programme. So in this community, the communities can be sponsored. For example, if there's a discussion, from local, about bicyclists, about bike paths, then there might be a bike shop who's sponsoring that and is taking part in the in the in the discussion, so they are adding value and our

In grading themselves into local communities.

yeah, I'm running out of time, so 8 strategic lessons for these

From these four sites, um.

Lesson one, focus on local operations.

Do not focus on local, on technological complexity for its own sakes, but, focus on how this can help you in your own local situation.

launch with strong

Partnerships, this

These local examples, they do not try to reinvent the wheel. I mean, the minute the Minnesota partnership shows that if you band together, you can be stronger. Not every local newsroom needs their own AI lab.

You can have mutual lapse and pull your expertise and your costs and learn from your other from your experiences.

3, minimise your risk with phased rollouts. this is what both Minnesota and Philadelphia are doing with their internal first, archive.

AI informed archives.

AI for business intelligence, that is what both the Baltimore Banner and Village Media are doing, they do not they do not ask any longer so much what should we cover? they ask what should we cover that drives subscriber value that our membership value. What are actually what is actually valuable content that are local communities want to be.

part of

0.5 is community responsive AI strategy.

That

Is the local advantage if you are part of your community in this

In this model where the community is first and not the newspaper is first, then you get valuable advice from your communities and data driven that can be reinforced.

And be driven back and inform your editorial strategy.

Revenue diversification through data that is something that some of them are doing. If you're successful with, with how AI has worked for you on the local model. Maybe there's a business model behind it. If it works for your newsroom, why wouldn't it also work for others?

embrace data and then

Critic human in the loop. We've been hearing, hearing this all day long. You should be.

Always, always checking what AI tells you.

yeah, thank you very much.

I know

Thank you, Oli.