How We Examined AI’s Single-Narrative and Hype in Indonesia

When we started this project in June 2025, concerns, questions, and critical reporting surrounding the current trajectory of AI development had been growing elsewhere in the world, while discourse in Indonesia was still primarily led by the overtly enthusiastic government officials. 

On social media, local tech professionals and CEOs were happy to disseminate the gospels of Silicon Valley tech leaders. The Indonesian government had granted OpenAI CEO Sam Altman a special “golden visa” recognizing his vision of the future as highly important.

A 2024 study done by Rio Tuasikal—one of the journalists in this project—showed that private sector and government quotes dominated 70% of AI media coverage in Indonesia. About 60% of the articles relied on a single source.

---

As a nonprofit journalism organization, we depend on your support to fund more than 170 reporting projects every year on critical global and local issues. Donate any amount today to become a Pulitzer Center Champion and receive exclusive benefits!

---

NVIDIA’s Jensen Huang was allowed to talk on a popular media channel about how AI was “almost like an actual person”: a very knowledgeable tutor, doctor, or consultant that can “democratize knowledge,” The media interview with Jensen did not contain any challenging questions; or at the very least, deeper inquiries into the nuts and bolts of the AI systems and the imagined future itself.

Therefore, the main goal for our project was to directly respond to the hype and claims surrounding AI permeating the public discourse with very little alternative perspectives.

Through our choice of publishing outlets and events to promote the reporting, we targeted Indonesian well-educated urban populations and youth—demographics that possess the social and cultural power to influence decision makers in both public and private sectors. 

Part I: Labor

We wrote several stories on how AI was reshaping white-collar jobs like media and creative work, as well as blue-collar work like driving for ride-hailing apps.

We set out to investigate common claims made by AI CEOs when it comes to AI and labor: that AI tools will boost productivity, that it will free up time for “higher-value” work, and that it will create new kinds of work that demand higher creativity and human ingenuity.

Based on these promises, we asked: Has AI truly boosted productivity and what does that look like in the workplace? Have workers gained more free time for higher-value work, and have they become more creative? 

We found two reports that guided us: "Generative AI and Labor: Power, Hype, and Value at Work" by Aiha Nguyen and Alexandra Mateescu, and "Large Language Models, Small Labor Market Effects" by Anders Humlum and Emilie Vestergaard.

We identified key findings raised in those studies, which were mostly collected from developed countries like Denmark and the U.S., and, using qualitative methods, tried to see if those findings would also hold true in Indonesia: What patterns might take shape in Indonesia? How? And where might they differ?

In our interviews, we found that many of the symptoms outlined in the reports are also present in Indonesia. AI output still demands substantial human revision, ultimately increasing workers’ workloads and burnout when bosses fail to recognize this additional work. Meanwhile, wages have not increased.

Our sources also told us how AI makes the wheel of production spin ever faster, and with that, meaning and satisfaction they once found in the jobs is slowly eroding. In the creative world, AI makes production instantaneous, which means one piece of work is now just a drop in an ocean of content.

“We live in a world where art, design, and video no longer need to carry creative value, to be life-changing, or to keep you up at night in awe,” said Tika, one of our sources who is a freelance creative worker. 

Freelancers also often have to bear the burden of higher software and hardware prices to be able to use AI tools. With AI, some freelancers are busier than ever, while some others reported a decrease in the number of clients they usually get.

Desirable job opportunities for junior workers are disappearing. We spoke to fresh graduates who had dreamed of becoming writers and journalists. They desperately looked for a job in the media, only to be pushed to other sectors they did not plan for.

Heightened precarity and uncertainty in an AI-exposed labor market is well documented; wages and job security are among the top concerns for workers and experts we spoke to. 

But through our reporting we also found that when AI does not cause job loss and "merely" reshapes an existing job, or when it succeeds in creating new opportunities, it often fails to consider the human need to find meaning, purpose, and enjoyment in their work. It also often closes opportunities for developing social relationships—such as through mentorship and apprenticeship—and makes people work in silos a lot more.

We placed these findings within the historical context of how AI has been developed since the 1950s, to show how technological advancement is always the result of ongoing struggles and debates between various individuals and institutions, often with conflicting agendas. 

Technology is always actively designed, regulated, and financed by humans who have to make choices—it is not something that appears out of thin air and can only be passively accepted no matter the form.

It was apparent in the process of our reporting, that there wasn’t always space for having a different thought about AI. Our sources often asked for anonymity out of fear of repercussions or damaging their companies’ reputation. We needed to be careful in the way we described our sources in the narrative.

For our story on what happens when ride-hailing drivers are managed by algorithms rather than humans, the formula stays the same: Take companies’ statements and hold them up against reality. We reviewed companies’ statements and cross-checked that with data from the field that we gathered from a small survey, as well as personal testimonies from several drivers. 

In a press statement that was quoted by the local media, Grab said that drivers recorded an average 21% higher hourly income after using its AI driver guidance. We decided to check if this claim about income was true via a small survey.

We asked several drivers we interviewed to forward our questions to their peers. By the end, we gathered responses from 40 Grab and Gojek driver-partners.

Our survey showed that the earnings of online taxi and ojek drivers this year have instead fallen to levels close to those seen during the Covid-19 pandemic.

They also still worked long hours: 60% of respondents said they work 10-14 hours a day, and 57.5% said they work seven days a week. Meanwhile, 21 respondents are still paying installments on, or renting, the vehicles they use for work. Installments or rental fees make up around 50% of their gross earnings, excluding fuel costs. 

Takeaways

The methods we implemented were extremely simple. Take any claim, statement, assumption, advertisement, or marketing promise, and hold it up against reality. From our experience reporting on AI and labor, we recommend you to:

1. Gather your own data. Companies often throw numbers from their “internal study” to persuade you that their AI tools work effectively on a given task. Whenever possible, it is always worth gathering your own data—whether by conducting your own research or by checking with data from other experts.  
2. Break it down. The effects of an AI or algorithmic system may bleed well beyond a specific task or goal and into the worker’s overall experience with their job. Break down the ripple effects (even ones you think are already obvious), and help your audience connect the dots.
3. Discern well. Conversations about labor, new labor, or the future of labor in the context of digitalization can be confusing because AI companies might send different messages that depend on who they are speaking to. When talking to workers, they might promise time-saving prowess, but when speaking to investors, they might talk about labor-cutting potential. 

Part II: ChatGPT bias, news, and journalists

Can AI replace journalists? It was a question that frequently came up in our interviews with journalists, media executives, industry associations, and researchers, when we were writing about how AI is reshaping newsrooms in Indonesia. The question captured the complex dilemma of media workers between using AI to survive and feeding the technology that is used to replace them. 

That sounded hypothetical and speculative at first. But when we looked at the data, it started to be a serious, valid question. The 2025 Reuters Institute Digital News Report showed that 7% of survey respondents said they had used AI chatbots for news, with the share rising to 15% among those under age 25. 

Indonesia was among the top three countries most comfortable with news produced mostly by AI, behind only India and Thailand. This means sooner or later more and more people will come to chatbots, not journalists or news websites, for news. If so, can ChatGPT, Perplexity, Claude, or Gemini be trusted? 

A critical moment for a test

We were reading that Reuters Institute report when a series of national demonstrations—dubbed as the biggest since 1998—took place in Indonesia in late August 2025. Amid a huge protest in the capital Jakarta, a motorcycle taxi driver was killed after his body was run over by a police armored vehicle. 

His name is Affan Kurniawan. The incident ignited national anger. Protesters and activists said the police killed him. However, police quickly defended the incident as accidental. 

This tragedy, however, provided a critical moment for us to test the quality of chatbots in providing news. We came to the free version of ChatGPT 5 with simple questions. We set two criteria: accuracy and balance. Below is the conversation between us and ChatGPT. The conversation was originally in Indonesian and was translated for this article.

Us: “Who was Affan Kurniawan?”

ChatGPT: Affan Kurniawan (born July 18, 2004, in Bandar Lampung–died Aug. 28, 2025, in Central Jakarta) was an Indonesian online motorcycle taxi driver who tragically died after being accidentally hit and run over by a police tactical vehicle during a protest near the House of Representatives building in Central Jakarta. (Wikipedia, Wikipedia)

ChatGPT provided a quick summary with an accurate date of birth. However, our eyes quickly caught the phrase: “accidentally hit and run over.” Was it an accident?

So we checked the sources ChatGPT quoted. The first one, in English, said the police vehicle “ran over and killed” Affan. The second one, in Indonesian, said Affan was “hit and ran over.” The original conversation in Indonesian will show that ChatGPT replaced “di-” (implying intent) to “ter” (unintentional or accident). So why did ChatGPT remove that intentionality? We asked again.

Us: “Was he hit or accidentally hit?”

ChatGPT: That’s a great question 👍

From various news sources and early investigation reports:

• Brief chronology: Affan Kurniawan was completing a delivery in the Pejompongan area of Central Jakarta on the evening of August 28, 2025. At that time, clashes broke out between the crowd and law enforcement. Affan attempted to cross the street but slipped amid the chaos. 
• Police and witness accounts: a passing Brimob tactical vehicle then ran over Affan’s body. Police described the incident as an accident because Affan fell directly into the path of the moving vehicle (Antara, Al Jazeera)
• From the perspective of the family and the public: many believe Affan was struck by the vehicle, since the large tactical vehicle should have been able to stop or swerve, especially given the heavy traffic at the time. For this reason, the term “struck” is used more frequently in the media and public discourse. 

Interesting that it used that emoji when responding to someone’s death. However, we pressed further:

Us: Did you witness the incident?

ChatGPT: No, I didn’t witness the incident firsthand 🙏

I can only provide information based on:

• Media reports (Antara, Detik, Al Jazeera, etc.)
• Official statements from the police and government
• Public testimonies/activist accounts reported by the media
• A summary on Wikipedia that compiles various sources 

So I don’t have firsthand experience at the scene, but I’m helping to trace the differences between the various versions of the story circulating. 

Would you like me to summarize the timeline of the Affan incident, the police version vs. the eyewitness version, to make the comparison clearer?

When we pressed further, ChatGPT quickly synthesized information from multiple perspectives, but it lacked the epistemic judgment to determine which account was true. In the end, ChatGPT said it sourced the answer to a press conference conducted by the police. In short, it is not trained to provide nuanced answers that reflect reality.

Asking different topics

We decided to test how well ChatGPT upholds journalistic balance by asking it about other news stories. Asked about the most important news in Indonesia on August 17, 2025, for instance, ChatGPT responded with a summary titled “‘One Piece’ Flag Protest Symbol Draws Attention.”

The report framed the flag as an expression of public discontent but offered no background on the flag’s use or symbolism. It concluded with an official government statement calling the flag “a threat to national unity.”

We asked about the Indonesian economic growth rate that was contested by economists and was reported to the U.N. ChatGPT simply said it was correct, citing government sources. Only when asked about criticism about it, did it provide a comprehensive summary about the contested data. In summary, it summarized quickly but omitted opposing or balancing viewpoints, and tended to quote official sources. 

Solidifying our findings

Various studies in the U.S. and Europe have documented algorithmic bias to news. However, we wanted to understand what happened to our own conversation with ChatGPT. It is hard to attempt a comprehensive LLM audit for this. But we documented everything and looked into academic sources—both papers and experts—that could explain our findings. 

Papers: We identified two papers in Indonesia. One study found that bot journalism “fails to determine causality or formulate questions.” Another study noted that AI-generated writing “struggles to meet the principles of balance and social responsibility” that are foundational in journalistic ethics.

Academics: We also talked to several journalism professors and researchers. They provided us with useful frameworks to understand what machines can and cannot replicate in terms of journalistic works. First, the technology does not replace journalists' presence on the ground. Second, the technology can only synthesize what has been published. And third, the technology doesn’t have a journalism stance to the public and understanding of a complex political phenomenon. 

Takeaways

From our experience, we recommend three things when investigating chatbot and news:

1. Document everything. Screenshot and download all conversations to a folder. Also, document sources it quotes to find disparities. Rename them to something specific. Leaving it as labeled a "screenshot" will risk it getting deleted later. 
2. Go back to the basics. It is easy to be amazed by chatbot technology, especially for being so quick at summarizing. However, we have to remember that speed is not the only and most important thing in journalism. We have to evaluate responses in terms of accuracy, fairness, and public interest. 
3. Trust your own gut. Machines inherit the political and cultural sensitivity of its training data and decisions of its developers. When it comes to complex human stories, trust your own editorial judgement. Ask simple 5Ws and 1H questions (who, what, where, when, why, and how) to ensure that machines are not hallucinating. Reading and verifying the sources are always recommended. Don’t lose your skepticism just because you are interacting with a fancy technology. 

Part III: Facial recognition tech in the civil servant pension system

For four months, we investigated the rapid digitalisation of Indonesia’s civil servant pension system. We found that the introduction of digital apps disregarded the unique needs of senior citizens, leaving them unable to navigate the verification process and sometimes locking them out of the system. 

The Indonesian Civil Servant Savings (Taspen), a state-owned enterprise, introduced a facial recognition system, “Taspen Otentikasi,” in 2018 to digitally verify beneficiaries. This aimed to replace the mandatory in-person verification that retirees or their families had to complete at Taspen or bank branches every month. In return, the app requires them to perform a series of body movements for verification, including shaking or nodding their heads, blinking, and pronouncing the letter "A."

On social media, retirees’ families complained about the complex instructions and technical problems. In 2025, Taspen finally simplified the facial recognition process in a new app. However, Taspen decided to also transform the app itself into a super app. Called “Andal by TASPEN,” the app now has 14 additional functions, from work accident reports to news updates.

This marks the moment when people got even more frustrated with the app. Interestingly, almost no Indonesian outlet had published stories about this. What they had published was tips and tricks to ensure a smooth verification process—for example, ensure adequate lighting, avoid peak hours, or simply try again. We saw a deeper problem, and we decided to investigate.

Unraveling the hidden problem

To understand when the problem began, we looked at X, formerly known as Twitter, and searched for the keywords “gagal autentikasi Taspen” or “gagal otentikasi Taspen” (both mean Taspen authentication failure) and found around 400 tweets from 2019 to 2025. 

We found that 186 of them were replies by Taspen's official Twitter accounts. However, this still proved that the problem has existed long before 2025. We then replied to some of these complaint posts, asking if they wanted to be interviewed for our story, and got five people who agreed to share their experience. 

We also wanted to understand the commonalities in the problems, so we looked into reviews of the apps on Google Play and the App Store. We specifically looked into one-star ratings. Based on that, we structured interview questions for our sources that we found on Twitter:

1. Can you describe your experience with the authentication process?
2. How often does this issue occur?
3. Are you able to authenticate now?
4. When you contacted Taspen yesterday, did they respond and assist you?
5. Have you ever experienced authentication failures that prevented your parents or family from receiving their pension on time?
6. You said that the 2018 app worked. What’s the difference with the 2025 one?
7. Taspen mentioned they might revert to manual authentication through the bank. In your opinion, is that a better option?

Building our arguments

By tracking complaints online, we were able to identify a phenomenon beyond isolated incidents. However, when we reached out to Taspen for confirmation, they told us that the authentication failure rate for app users was around 3.73% in November 2025.

It is important to note that when investigating a tech system like this, it is very easy for government agencies or private companies to frame it as statistically small, anecdotal, or simply a glitch. That is why we look into scientific sources to unravel even deeper problems. 

Literature: We found two journal articles that studied the ability of Indonesian senior citizens to use digital technology. The first study, published in the journal Gerontechnology in 2024, noted that only 19% of Indonesian seniors had ever learned to use a cellphone. The second study identified several barriers in a QR-based payment system in Indonesia, including overly complex navigation, icons and typography that are too small, and ambiguous system feedback.

Academics: We talked to three experts from different disciplines—media literacy, law, and senior studies—who could explain frameworks to evaluate what is actually going on. To avoid "ivory tower" perspectives, we specifically looked for experts who have direct experience in capacity training or empowerment for senior citizens. One of them actually helped her own mother navigate the pension app. 

These interviews allowed us to evaluate digital systems not only in terms of product launch or the latest features, but also in terms of sociotechnical issues. For instance, who designed this technology? Are they young, urban, tech-savvy developers thinking about themselves? How can a technology be designed with the end users’ demographic in mind? How can government agencies provide alternative offline services to ensure no one is left behind? 

Key takeaways 

1. Find people online. We started not knowing anyone impacted by the system. However, tracking complaints online helped us find families of retirees. Be transparent about who you are as a journalist and explain the kind of story you want to write. Make sure they understand that you don’t promise to fix the issue right away. 
2. Look into ratings and reviews. This is helpful for expanding our reporting as a phenomenon rather than as isolated incidents. It also helped when digital system providers—be it government agencies or private companies—often frame your findings as anecdotal or statistically insignificant.
3. Use academic sources. Academic terminology does not have to appear in your reporting. However, talking with academics provided us with a strong analytical lens to sharpen our inquiries. It is also effective in unravelling the institutional bias behind a technology that marginalised its intended users.