Every year, thousands of government audit reports are produced in the Philippines. They’re full of potential stories, but going through them takes time. Jaemark Tordecilla used a custom GPT to help journalists explore those reports and find stories worth investigating. It took him just two days to develop a working approach.
In this AImpactful conversation, Branislava Lovre speaks with Jaemark, a journalist, media adviser and technologist who led GMA News Online for nearly a decade and was later a Nieman Fellow at Harvard University.
Today, Jaemark is building AI tools and helping newsrooms do the same. From investigating public spending to exploring hundreds of hours of parliamentary hearings, he shares practical examples of what journalists can achieve with AI, even without programming skills or dedicated technical teams.
His advice is simple:
“Always start with a journalism problem, not the AI tool.”
We also discuss useful AI tools, verifying information, protecting sensitive documents and why journalism’s real value remains “the integrity of that information.”
And Jaemark has an enthusiastic message for journalists:
“Let’s get greedy. Let’s build more. Let’s find more stories. Let’s find more data. Let’s investigate more.”
What we explore:
- Jaemark’s journey from technology to journalism
- Finding stories in government audit reports and public budgets
- How smaller newsrooms can work with AI
- Building tools without coding skills
- Using Claude Code, Codex and Gemini Notebook
- Making lengthy parliamentary hearings easier to explore
- Why journalists should start with a problem, not a tool
- Checking AI-generated results and protecting sensitive information
- Practical advice for journalists who want to experiment
Episode details:
- Duration: 29 minutes
- Guest: Jaemark Tordecilla, journalist, media adviser, technologist and former Nieman Fellow at Harvard University
- Host: Branislava Lovre, co-founder of AImpactful
- Format: Video podcast with full transcript
Transcript of the AImpactful Vodcast
Branislava Lovre
Hello and welcome to AImpactful. Today, we’re talking about how AI can help journalists with their investigations, work with data, or build their own tools, and what we should be careful about along the way. My guest today is Jaemark Tordecilla, a journalist, media advisor, and technologist. He spent almost a decade as editor-in-chief and was a Nieman Fellow. Today, he develops AI tools and works with newsrooms around the world, helping journalists explore new possibilities.
We’ll talk about his experiences, projects, and practical ideas you might want to try yourself. Jaemark, it’s really nice to have you here.
Jaemark Tordecilla
Thank you. Happy to be here from Manila.
Branislava Lovre
Before we get into AI and all the interesting projects you’ve been working on, I’d love to hear a little more about you. You’ve had a really impressive career, from leading a major digital newsroom to building AI tools and working with journalists around the world. How did that journey begin? And how did technology and journalism come together in your career?
Jaemark Tordecilla
It actually began in tech. I graduated with a CS degree, even did a master’s in tech management. And so I was leading that career in tech before, about midway through, I made what was either the best or the worst decision of my career, which was to become a journalist. And then I ended up becoming a member of the Philippine Center for Investigative Journalism, before eventually becoming editor-in-chief of GMA News in the Philippines, which is the largest TV news organization.
And I was in charge of everything online. So the news website, video publishing, social media alerts, and everything that was related to online. So I did that for nearly a decade before leaving in 2023 to do a Nieman Fellowship at Harvard. And it was the best time to be on campus because everyone on campus was talking about AI. But no one on campus was talking about AI in the context of journalism, because while the Nieman Fellowship is at Harvard, there’s no journalism school per se at Harvard.
No journalism degrees at all. And no one was talking about AI in the context of the Philippines and the larger Global South. Because of my background in tech and in journalism, it sort of fell into my lap. It sort of fell upon me to start thinking about AI in the context of both my industry and my geography.
Branislava Lovre
You mentioned your time at Harvard, and I’d love to go back to that for a moment. If we think about 2023, GPT was still relatively new, and many journalists were just beginning to try it out. And now, only a few years later, so much has changed. I’m curious, what were those early days like for you?
Jaemark Tordecilla
So yeah, in late 2023, I was taking a course under Professor Sabrina Osmany at Harvard, an intermediate course on AI. And one day, near the end of the semester, the professor came in and told us that what we were studying was going to be obsolete because OpenAI just announced they were creating custom GPTs where you can put instructions in words to the AI model and you can connect data sources to it and you can connect whatever files to it.
And so she was like, all this we’re studying right now is sort of going to become obsolete. And so that was the impetus for me to play around with it. In late 2023, I built what was one of the first AI custom GPT applications meant for investigative journalists. So it was a tool for investigative journalists in the Philippines to inspect audit reports that would make their lives easier.
Just as background, in the Philippines there are efforts for transparency. So audit reports are a great tool for journalists to find leads about issues with how our money is being spent by government agencies. An audit report is required for every government agency, every local government. That means every town, every city, every province in the Philippines has an audit report.
We’re talking about thousands of audit reports every year. So there’s no way a journalist covering that would be able to read all of that. Before ChatGPT, it would have taken me hundreds of hours of training, maybe to create a model to try to find stories in audit reports. Using a custom GPT, it took me two days to find a workflow that works for journalists in the Philippines. And journalists in the Philippines and newsrooms in the Philippines have been able to use that tool over the past couple of years.
I’ve had the opportunity and the privilege to talk about my work, my other experiments in AI since then. I’ve continued to build tools, and now it’s the most fun I’m having building tools because of coding agents. And I’ve also been active in advising newsrooms around the world about how they can integrate AI into their workflows and how to build things that would be helpful for their journalism as well.
Branislava Lovre
And that brings me to something I think is especially important for smaller newsrooms. We’ve seen impressive AI projects from large media organizations, but they often have dedicated technical teams and much bigger budgets. For smaller, independent newsrooms, things can be quite different. You’ve been working closely with journalists and helping newsrooms find practical ways to use AI.
So, based on your experience, if a small newsroom wants to start using AI more seriously or perhaps build something of its own, where should it begin? What should it focus on first?
Jaemark Tordecilla
I think the first thing that newsrooms listening to this and who haven’t taken the plunge should do is understand the technology. I understand that there’s a lot of resistance or hesitation among many newsrooms about the capabilities of AI. But the way I’d like to think about it is that a good framework when you’re designing a workflow is figuring out what the AI is good at and where the journalistic sense comes in.
For example, AI has gotten really good at working with data, with structured data, spreadsheets, databases. So journalists know to ask the right questions and to find data sources. And that’s something I’ve been working on with independent newsrooms here in the Philippines, where I try to teach them about how to interrogate and investigate data sources that are related to the fiscal space of the Philippines.
So one of the things I taught them was how to inspect the Philippine national budget using the most current AI tools. It’s a paradox because here in the Philippines, we have one of the most open budget processes in the world. It’s one of the most transparent. You can see every line item, like how much each office of the government spends on internet access. You can find that in the budget.
The problem is you’re going to get drowned in numbers if you don’t know where to start. So a lot of journalists, and a lot of newsrooms, actually even the big ones, don’t know where to start. And so one of my projects was to make a reusable skill that you can use on Claude Code and Codex where you just install it, download the budget databases, and it allows you to ask questions in plain, natural language.
You can ask it in English, “How much was spent for flood control projects in my region?” And it will translate that into code, and that code will be used to query those databases. And it’s a lot more effective. There’s a lot less room for hallucinations because the hallucinations come when the AI writes the code, which means that if there’s an error there, it just won’t work and then it will try the query again.
Of course, there are safety things that we take into account. We try to mitigate the risk. We talk about how to spot-check the accuracy of the results and how to interrogate those results. But even that, you don’t need a programming background anymore. You don’t need to have a data scientist or a data analyst in your newsroom to be able to interrogate massive amounts of data.
Branislava Lovre
And once we know what we want to achieve, the next question is how to actually get started. I imagine some journalists watching us are thinking, “OK, this all sounds great, but what can I try myself?” Let’s say I’m a journalist with hundreds of documents or a large dataset, but I don’t have any coding skills. Which tools would you recommend starting with, and is it possible to experiment without spending too much money?
Jaemark Tordecilla
If they’re working a lot with documents, a tool that I usually use and recommend is Gemini Notebook, formerly known as NotebookLM, from Google. It’s the best way to inspect documents and it comes on the free tier of your Google account. So there’s no commitment to pay a monthly subscription fee. I would also encourage people to at least experiment with the new coding agents like Claude Code and Codex and to use it at the first tier of the paid level.
It’s $20 or so, but it allows you to see the capabilities of the newest AI models. A lot of journalists, I found, tried it a few times before. Maybe they tried to use it for generating a report or an essay. It’s really not the best use of AI, in my opinion. But if they go to the first level of the paid tier and try it with coding agents, they’d be amazed at how good it is in interrogating datasets, in making sense out of big documents.
What I like to tell journalists and newsrooms is that this is an opportunity to tell stories like we’ve never been able to before, to tell stories that we weren’t able to before because some things are too technical for us or require expertise in data analysis or data science. And now that’s no longer the barrier. Now the barrier is: can you find relevant data to investigate? And if governments and those in power don’t make that available, what can you do as a journalist to get those documents and datasets available to you?
Branislava Lovre
Can AI help us with that part too? How can we use these tools to find and access public data?
Jaemark Tordecilla
Right, so with these coding agents, I’ve been able to demonstrate a lot of ways to be able to get data. For example, I’m coaching a bunch of newsrooms from Mongolia this year. I visited Ulaanbaatar to conduct an AI seminar there, and I never would have guessed I’d end up in Ulaanbaatar to do a seminar there, but that’s where the work took me. And I discovered that their national statistics office actually publishes an API, which is an application programming interface that allows you to download the data.
And for the longest time, the barrier to using that API was just that you needed to have a programmer to build an application to access that API. And I would demonstrate how Claude Code or Codex would be able to figure out how to access that API with just a few prompts. And now all their stories could have a data layer in it. Of course, it’s not going to do the job of the journalists of finding the stories in the data, but it makes access to the data and making sense of the data very easy.
So journalists no longer have an excuse of not having a data layer in their stories or not having data to support their stories.
Branislava Lovre
We’ve talked about some really interesting ways journalists can use AI to work with public data, and I know you’ve been building some impressive AI tools yourself. I’d love to hear about one of your recent projects. What was the problem you wanted to solve, and how does AI help?
Jaemark Tordecilla
Right. So one of my latest projects is about the impeachment trial of the Philippines’ vice president. She’s being impeached right now and the trial is being held at the Senate. And it’s a marathon, sessions of several days, I think. So I wanted to create a map for it, and I used AI to help me build that. I would take the transcript of those hearings, which you can create using machine learning models, and identify the different speakers.
It will identify the different voices of those who are speaking. And the novel thing about the work is that I used agentic AI to identify who actually said what. And so it’s really clever. It reads the context of the whole transcript, so when someone is introduced, then the AI figures out that, oh, that someone is this lawyer or that senator or this witness. And so I was able to put structure into the whole transcript, which used to be unstructured.
And so at the end of it, I could generate a digest of what happened during the day, who said what, and what the important points were. And again, like I said, putting structure into what used to be unstructured data. I am able to create an audio overview or a podcast out of those summaries. I can find a witness list because I’m able to identify who are the witnesses who appeared in any day.
And so I have a witness list for the whole trial. I’m also able to track documents that are introduced into evidence during the trial, as well as jurisprudence of rulings from other courts and, you know, issuances and memos, all that legal stuff. And I’m able to extract that from the transcript and even link to it or make it available for download on my website. And so it became a great use case not just for journalists who are covering the trial, but also for lawyers and law students who are following the trial, and also for regular citizens.
And so I took that same model and I’m now applying it to parliamentary hearings about the Philippine budget, which is like twice as long and twice as complicated. And now I have a map of the whole budget process. Like, if someone asks for an additional fund for his district, we can track that from the transcript of the budget deliberations. And so it’s added a whole layer of transparency. And the most impressive thing, I think, is that I’m doing it on my own.
Basically, I’ve covered hundreds of hours of parliamentary hearings on my own, only with the help of AI.
Branislava Lovre
I have to say that’s impressive, especially when you think about how much information you’re working with. And when you explain it, it almost sounds easy. But I’m sure there were some challenges along the way. How long did it actually take you to go from the first idea to something that worked?
Jaemark Tordecilla
Right, so let me talk about that in the context of the impeachment hearing. So it took me not that long to be able to come up with the first prototype. I did that in about a day. So it was just me experimenting with whether AI could identify speakers using their voices, the context of the conversation, and even their appearance.
Funny thing, ChatGPT and Codex don’t allow you to do biometric searches of speakers. So you just rely on their outfits. So it’s able to identify the different speakers if they’re the same, if the outfit remains the same throughout the hearing. And so I did that in about a day and then it took me about a week to catch up on a few weeks’ backlog of sessions, the hearings themselves.
And now, an eight-hour session would take maybe about three hours to process. And the bottleneck is me because I check everything before I publish. I double-check; I do spot checks of the AI’s work. The goal is not transcript fidelity. What it does is give you a rough transcript, and then you can check directly in the video whether what the transcription said was accurate.
But the goal really is to put structure into those eight-hour hearings and to be able to summarize them effectively. So AI is very good at summarization. It’s very good. It’s gotten very good at figuring out context from what was said so that when it figures out that the end of the conversation between the lawyer and the judge is marked by a phrase like “Yes, Your Honor.” Right?
So it’s able to figure that out. What it’s still bad at, or at least not up to human judgment, is the judgment part. So, for example, during the impeachment trial, there are many senators. We have 23 senators right now in the Philippines. Not all of them attended, but maybe five to 10 senators speak up every day during the trial. And to start with, I let AI pick which senator to include in the summary.
And it was doing a bad job at figuring out who had the most important questions. And so I didn’t want to do it myself, right? Because I only have so much time in the day and I don’t want to go through all those transcripts. So, to mitigate the lack of judgment of AI when it comes to who to include and who to exclude, I just revised the system to include everyone. And then journalists who use it can use their judgment to figure out which is the most important part of those questions.
Branislava Lovre
We’ve talked about some really interesting ways AI can help journalists, but there are so many tools available now and new ones appear almost every week. When you’re working with a newsroom, how do you help them figure out which tools they actually need? Where should they start?
Jaemark Tordecilla
Right, so this is one thing I always tell the newsrooms and the journalists that I work with. Always start with a journalism problem, not the AI tool. Figure out what the problem is. What problem are you trying to solve? What feature are you trying to build? What story do you want to tell your audiences that you weren’t able to do before? Do that before figuring out what AI tool is out there, this shiny new thing, instead of starting with a solution looking for a problem.
Start with the problem. Do you want to retain users better? Do you want to reduce the workload on your day-to-day staff? Do you want to be able to do investigative journalism at a scale that you were never able to do before? Do you want to make your readers understand complex numbers and complex topics in explainers and short videos? Figure out what the problem is, and then from that, backtrack and choose what AI tool would be right for solving that problem.
Branislava Lovre
I’ve heard about so many exciting possibilities, but of course there are also risks, especially when journalists work with sources, public records, or sensitive information. From your experience, what should journalists be careful about? And what would you say are the most important things to keep in mind when using AI responsibly?
Jaemark Tordecilla
Yeah, I think fundamentally the first thing is that you still need to understand the input and the output. So you need to be able to verify whether the output is correct. If you don’t understand the process, if you don’t understand the story, if you don’t understand the source, then you’ll get into a lot of trouble. And so usually my advice for journalists is to bring your own data into AI and use AI to process it.
AI has gotten very good at fact-checking if you give it a ground truth source. It’s terrible at fact-checking if you just have it browse the whole web, right? Because you don’t know whether the sources that it will use would be trustworthy. But if you give it a ground truth like a document or a database or the transcript from a long parliamentary hearing, it’s gotten very good at fact-checking that.
And it also gives you the opportunity to audit and check whether or not the final output is something that you could publish, something that you can verify. After that, I think it’s also noteworthy to mention that journalists, newsrooms, tech teams need to read the label in terms of what these AI models will do with the data that you will upload. So some of them have explicit language where they say they might be able to use your data, your prompts, your documents and spreadsheets, if you’re only at the lowest paying tier.
If you’re just working with public documents, then that should be good, but if you’re working with sensitive documents, you have to read the label and figure out whether or not it is safe to upload into these models. If you’re not sure, talk to your editor, talk to your legal team, or just don’t upload it. But yeah, I think, in the same way that we don’t trust everything that the reporter passes, right?
The editor goes through it and does fact-checking and verification. I don’t think we should trust the output of AI willy-nilly. We still need to exercise that judgment. At the end of the day, I think this is something I always say in our workshops. The product that we offer the public as journalists is not content. Anyone can make content. Content creators make content all the time.
But the product that we offer our audiences, our viewers, our readers, is our guarantee that whenever we publish something or we air something or post something on social media, that artifact, whatever form it may be, has been vetted by us, has been fact-checked by us, has been rigorously reported by us. And so our product really is the integrity of that information. And I think fundamentally that wouldn’t change even though our tools have become so advanced and so different from what we had in the past.
Branislava Lovre
I really like what you said about the integrity of information. I think that’s something we should never lose, no matter how much technology changes. And we’ve heard some really interesting examples of what AI is already making possible. So when you look ahead, what would you love to see journalists and newsrooms do with AI?
Jaemark Tordecilla
I think it’s going to be an exciting time. In one of the programs I’m advising on, I told newsrooms, “Let’s get greedy. Let’s build more. Let’s find more stories. Let’s find more data. Let’s investigate more.” Because these tools now allow us the capacity to build things and to make stories about things that we weren’t able to do before. And so I’m excited about what we can come up with if we apply ourselves to these kinds of tools and workflows.
Branislava Lovre
And before we wrap up, I’d like to bring this back to the journalists watching us. You built your own AI tools and worked with newsrooms around the world. So, based on everything you’ve learned, what would you encourage journalists to explore or try in their own world?
Jaemark Tordecilla
I think the one piece of advice I could give to everyone listening right now is to get your hands dirty and build and try it out for yourself. Try it with a document you’re working on that you’re investigating or a dataset that you’re investigating and see what comes out of it. Now is the best time to experiment because it takes literally hours to build things that used to take weeks or months to build.
Just see for yourself how well it can suit your workflow. And I think you’ll be surprised by what you find.
Branislava Lovre
And I think that’s a great message to end on. Jaemark, thank you so much for sharing your experience and all these interesting ideas with us. And to everyone watching or listening, thank you for being here. Don’t forget to subscribe to AImpactful for more conversations about AI and its impact. Thank you for watching, and see you next time.


Leave A Comment