PSM and AI Part One: Mapping how public service media use AI in journalism

Part 1: Mapping how public service media use AI in journalism July 2025 PSM and AI Public Media Alliance 2 3Part 1: Mapping how public service media use AI in journalism Contents 1. Introduction 4 2. Methods 5 2.1 Defining ‘AI’ and ‘Public Service Media’ 5 2.2 Data and Sample 3. Overview 5 3. Overview 7 3.1 How long have PSM used AI to inform or assist journalism production? 7 3.2. How important is organisational income for PSM in developing AI strategies? 8 3.3. How Eurocentric is the conversation about PSM and AI? 8 4. PSM and AI tools 10 4.1. What AI tools are PSM using? 10 4.2. How much did PSM depend on the ‘Big Five’ technology companies? 11 5. Speech-to-text tools 12 5.1. Transcription and captions 12 5.2. Speech-to-text tools and data privacy 12 5.3. Translation 13 5.4. Variable accuracy of speech-to-text tools and minority or Indigenous languages 13 5.5. Challenging cases: Jamaica and South Africa 15 6. Images, AI presenters, and Synthetic Voices 16 6.1. Image generation and editing 16 6.2. AI Presenters and Sign-readers 17 6.3. Synthetic voices 18 6.4. Challenges: Ethics 18 7. Emerging uses of AI 20 7.1. Data Journalism and Retrieval-Augmented Generation 21 7.2. Challenges: Journalism research, the ‘Big Five’, and LLMs 21 7.3. Open-Source Intelligence Tools 22 7.4. Challenges: The ethics of facial recognition 22 7.5. Repurposing News Content 23 7.6. Challenges: ‘Mission Creep’ 24 8. Conclusion 25 9. Appendix 28 Cover photography supplied by VRT. Funding This work was funded by BRAID, a UK-wide programme dedicated to integrating arts and humanities research more fully into the responsible AI ecosystem, as well as bridging the divides between academic, industry, policy and regulatory work on responsible AI. BRAID is funded by the Arts and Humanities Research Council (AHRC). Funding reference: Arts and Humanities Research Council grant number AH/X007146/1. Lead Author Prof Kate Wright Prof Wright is the Chair of Media and Communication at the University of Edinburgh. She’s also the Academic Lead of the university’s 80-strong Media and Communications Research Cluster. She is a former BBC journalist, who has published extensively on international news, focusing on the relationship between political economy and journalistic practice. This includes three monographs: Who’s Reporting Africa Now? (Peter Lang 2014), Humanitarian Journalists (Routledge 2023) and Capturing News, Capturing Democracy: Trump and the Voice of America (Oxford University Press, 2024). She serves on the editorial boards of two top-ranked Communication journals, Digital Journalism and the International Journal of Press/Politics and has just completed a year of service as an expert research assessor for the 50-country informational Forum for Information and Democracy. Her research has been widely covered by the press, including ABC, Associated Press, CNN, Deutsche Welle, and Vanity Fair. Co-Author Kristian Porter Kristian is the CEO of the Public Media Alliance, the largest global association of Public Service Media organisations. Kristian has extensive experience working in public media support and advocacy, with a background in the NGO sector, editorial management, events, and as a journalist. He has worked globally in his roles for PMA and works with the association’s membership and partners to develop advocacy campaigns, international relations, strategies, research and media development projects. He is a passionate advocate for media freedom, media independence and journalist safety, and sits on the steering committee for the Public Broadcasters International (PBI) conference and runs the secretariat for the Global Task Force for public media (GTF). Kristian has an MA in Media and International Development. Public Media Alliance 4 5Part 1: Mapping how public service media use AI in journalism 1. Introduction Accuracy, accountability, impartiality, universal access, and editorial independence. These are some of the core values that drive public service media (PSM) and their mandate to serve the public interest and support democratic discourse through high quality, trusted news content. But in a world where PSM face increasingly difficult financial and political conditions, and amid an increasingly threatening environment for journalists, how best can PSM newsrooms source, create and distribute content without risking the values they hold dear? Some say that Artificial Intelligence (AI) holds the key. The rapid advancement of AI has transformed the global media landscape, raising fundamental questions about the role of technology in democratic life. Nowhere are these questions more acute than PSM, where it is crucial to assess how these technologies can aid their mission without endangering public trust. AI has the potential to enhance the function of PSM. AI tools offer opportunities to streamline production processes, personalise content delivery, automate mundane tasks, and improve accessibility through tools like speech-to-text, translation, and captioning. These capabilities can help PSM reach wider and more diverse audiences, improve efficiency amid budget constraints, and respond more rapidly to breaking news or disand misinformation. However, AI also presents profound risks that go to the heart of PSM’s legitimacy. Tools powered by opaque algorithms and proprietary datasets raise serious concerns about transparency and accountability—principles essential to trust. Algorithmic bias, “hallucinations”, where a programme spits out false or nonsensical text, and the exclusionary dynamics of commercial data infrastructure can threaten the public service values PSM are expected to uphold. Financial and technological feasibility means that many smaller PSM rely on third-party vendors, with potential risks to editorial independence and technological sovereignty. These tensions have led to calls for a more collective and values-driven approach to innovation within the PSM sector. Efforts by the Director General 8 group (DG8), the European Broadcasting Union (EBU), the Public Media Alliance (PMA) and other associations reflects a growing consensus that shared standards, and collaborative frameworks, are needed to guide the ethical and sustainable use of AI. Yet, despite these policy-level discussions, there remains a significant empirical gap in our understanding of how PSM actually use AI in practice. Most existing research is either limited to Western Europe or is narrowly focused on recommender systems. This limits the ability of researchers, policymakers, and practitioners to assess the impact of AI on the editorial work of public service newsrooms or to identify best practices and areas of concern across different regions and media systems. To address this gap, Professor Kate Wright (University of Edinburgh) has conducted the first broadly-based international study of PSMs’ approaches to responsible AI within journalism production. The study, which was designed in collaboration with the Public Media Alliance, draws on survey data, internal documents, and interviews with senior managers and stakeholders across thirteen public media organisations from Africa, Asia, the Caribbean, Europe, and Oceania. It was funded by UK Research and Innovation (UKRI) as part of the Bridging Responsible AI Divides scheme. To disseminate the results of this research project and stimulate further discussion, Kate Wright and the CEO of Public Media Alliance (PMA), Kristian Porter, have authored two industry reports. This is the first. It maps PSMs’ current use of AI in journalism production: exploring which AI tools PSM use, how they use them, and how senior managers frame and justify their use in relation to PSM values. While the second will explore PSMs’ organisational policies, the responsibilities of PSM in relation to AI, and the dilemmas they face regarding AI procurement. It is hoped these reports will contribute to an urgently needed body of knowledge about the evolving role of AI in public service media, providing a foundation for more informed, collective discussion within the industry about how to harness technological innovation in a way that strengthens—rather than compromises— the democratic mission of PSM. 2. Methods 2.1 Defining ‘AI’ and ‘Public Service Media’ In this study, Artificial Intelligence is defined broadly as “a collection of ideas, technologies1, and techniques that relate to a computer system’s capacity to … perform tasks normally requiring human intelligence.” So, AI encompasses, but is not confined to, generative AI (GenAI): embracing various forms of automation, machine learning, and data analysis. Participants were asked about their use of AI in the production of ‘journalism’ rather than ‘news.’ This was because PSMs tend to ‘draw the line’ differently. Some distinguish between the uses of AI permitted in hard news and other content, while others differentiate between factual and other content. By asking about ‘journalism’, we were also able to address a range of other genres, including current affairs and investigative documentaries, which have not previously been analysed in other studies. PSM tend to be defined in terms of their commitment to public service values. But it is difficult to define a sample of organisations in this way, as some organisations that self-describe as PSM are state-controlled; while others move back and forth between degrees of government control over time and in response to political events. For this reason, we approached a diverse range of organisations able to exercise significant editorial independence, based on the latest version of the global State Media Monitor2, as well as PMA’s professional judgement. Full details are available in the academic article on which this report is based, which is undergoing peer review as of August 2025. 2.2 Data and Sample 13 organisations participated in this study, as listed in Table 1 (page 6). These media networks operate across five continents, work in a range of languages, and serve a mixture of domestic and international audiences. Unfortunately, we were unable to include any networks from North America as they declined, owing to the operational demands relating to general elections in the USA and Canada in 2024 and 2025 respectively. Given the importance of resourcing, organisations are listed in order of annual income: with highannual income being classed as €500M+p/a; medium-income as €100M+ p/a; and low-income as less than €100M p/a. We originally intended to have five organisations per category, but two high-income organisations were unable to proceed. Our sample therefore contains more organisations with medium and low incomes, but this is more representative of the sector. The data on which this report is based included an online survey, which all participating organisations returned, and semi-structured interviews with designated senior managers provided by 10 organisations. Data collection took place between May and December 2024, so this report can be treated as a ‘baseline study’ conducted before the release of open-source models by the Chinese firm DeepSeek, and the very rapid industry and political changes which took place in the USA at the start of the second Trump administration. This data collection period was longer than we had originally intended because AI was almost universally regarded as highly sensitive by PSM, which slowed down the process of obtaining organisational consent. But this longer data collection period had some advantages: enabling participants to attend various meetings between PSMs about AI3, and to reflect on the relationship of AI to political change, as 2024 was a ‘mega-election’ year in which voters in more than 60 countries went to the polls4. We will now go on to discuss our findings, beginning with an overview of the ‘state of play.’ 1 reutersinstitute.politics.ox.ac.uk/our-research/industry-led-debate-how-uk-media-cover-artificial-intelligence 2 statemediamonitor.com Public Media Alliance 6 7Part 1: Mapping how public service media use AI in journalism Name Acronym Country Income Category Annual Income Nippon Hōsō Kyōkai* (trans: Japan Broadcasting) NHK Japan High €4.3 billion (2023) Swiss Broadcasting Corporation SRG SSR Switzerland High €1.6 billion (2022–3) Australian Broadcasting Corporation ABC Australia High Combined government/ commercial income €741 million (2022-3) Radio-télévision belge de la Communauté française (trans: Belgian Radio-television of the French Community) RTBF Belgium (French) Medium €469 million (2023) Vlaamse RadioenTelevisieomroeporganisatie (trans: Flemish Radio and Broadcasting Organisation) VRT Belgium (Flemish) Medium €465.4 million Special Broadcasting Service SBS Australia Medium €307.1 million Sveriges Radio (trans: Swedish Radio) SR Sweden Medium €290.6 million (2022-3) South Africa Broadcasting Corporation SABC South Africa Medium €241.4 million (2023) Public Television Service PTS Taiwan Low €71.7 million (2022) Suspilne (trans: Public) Prior to 2022 known as National Public Broadcasting Company of Ukraine - Ukraine Low €40.5 million (2023) Radio New Zealand RNZ New Zealand Low €32.3 million (2022-3) Public Broadcasting Corporation of Jamaica PBCJ Jamaica Low €8.8 million (2023-4) TeleRadio Moldova TRM Moldova Low €8.3 million (2023) Table 1: List of participating organisations 3. Overview 3.1 How long have PSM used AI to inform or assist journalism production? 10 of the 13 organisations in this study said they used AI to assist or inform their journalism production. The amount of time which participants said they had used AI for is illustrated in Figure 1. To clarify, five organisations said that they had only used AI for 1-2 years. These were: ABC (Australia) PTS (Taiwan), SBS (Australia), Suspilne (Ukraine), and the Flemish-Belgian network, VRT. Four had used AI for between 2-5 years. These were: the Francophone-Belgian network, RTBF (2-3 years); RNZ (New Zealand) and SR (Sweden) (3-4 years); and SRG SSR (Switzerland) (4-5 years). NHK (Japan) appeared to be the most experienced: it stated that it had used AI tools for 8+ years. NHK has a longstanding inhouse R&D team, which publishes its own research papers online5. The three organisations that stated they did not use AI were PBCJ (Jamaica), SABC (South Africa), and TRM (Moldova). However, in their research interviews, PBCJ and TRM participants clarified this: saying that their journalists did use AI tools, but not in an ‘official’ capacity. These participants and specialised trainers said that journalists who used AI ‘unofficially’ tended to use free versions of Otter.AI (for transcription); DeepL and GoogleTranslate (for translation) and OpenAI’s ChatGPT (for research). SABC did not choose to do a follow-up interview, but previous academic research suggests that some of their journalists use AI ‘unofficially’ for translation6. Journalists’ unauthorised use of such AI tools was widely seen by senior managers and specialist trainers as posing significant risks regarding cybersecurity and data privacy. Yet the decision to adopt and regulate AI tools internally was not straightforward. For instance, TRM (Moldova) explained that the network’s income was so low that very few options were available to them, apart from using budget tools offered by or developed in China, which senior managers regarded as a step too far— even before the launch of open-source tools by DeepSeek in Jan 20257. While PBCJ (Jamaica) was concerned that formally adopting AI tools not only raised questions about affordability, but also the potential to become a target for cybercrime. In addition, adopting AI tools would require board and government approval, which risked politicisation. 3 Such as the meetings held by the Asia-Pacific Broadcasting Union, Médias Francophones Publiques, Public Television System, and the South Africa Broadcasting Corporation. 4 pewresearch.org/global/2024/12/11/global-elections-in-2024-what-we-learned-in-a-year-of-political-disruption 1–2 years 2–3 years 3–4 years 4–5 years 8+ years Figure 1: Duration of AI use as stated by survey respondents 5 nhk.or.jp/strl/english 6 https://www.tandfonline.com/doi/full/10.1080/17512786.2021.1984976?casa_ token=fFdV96BjxMQAAAAA%3AgKHrgnNB6LclPd0fUxDR4IbwsyBkXJ7m8ZX2dZN2rbao4ge6dY7Mr0PAvLz8iqf10ss03SMfTYDDoA 7 tandfonline.com/doi/full/10.1080/21670811.2025.2502129?src=exp-la#d1e451 Public Media Alliance 8 9Part 1: Mapping how public service media use AI in journalism 3.3. How Eurocentric is the conversation about PSM and AI? 3.2. How important is organisational income for PSM in developing AI strategies? The European Broadcasting Union (EBU) continues to operate as the most significant hub for collective dialogue and productisation (product development) between PSM. Suspilne (Ukraine), expressed particular appreciation for their inclusion in EBU’s European Perspective project8. However, European PSM did not seem to benefit equally. Another low-income PSM, TRM (Moldova), which is also experiencing acute tensions with Russia9, voiced frustration with other comparatively wealthy European PSM. A representative said: “We have had a very nice conversation with [A senior BBC executive] and he told me, ‘Well, even for the BBC, money is a problem.’ … The [senior executive] at Deutsche Welle also told me once that they have budget cuts every year, and they have to adapt and adjust. When I look at … the figures [in both organisations], even in the context of budget cuts, I still think that it is easier to cut a big cake than to cut nothing. I don’t really think that they get us … It is one thing to try and [hire AI experts at large PSM] which have a name and proper resources. It’s a completely different thing to hire at TRM, where the building was supposed to be demolished 20 years ago and the wind is whistling through the windows, and salaries are low.” By contrast, high-income PSM outside Europe had relatively easy access to peers via formal networks. For example ABC (Australia) and NHK-Japan are members of the Director General 8 (DG8). ABC also reported finding EBU conferences and other sessions “super useful,” as well as engaging in more informal “catch ups” online with EBU, the BBC R&D team, and Nordic PSM. But there was still a reluctance to share Intellectual Property, as one senior manager put it: “There are a lot of organisational similarities, I think, so it’s fairly easy to knock on doors and get a conversation going … Of course, it’s usually bilateral conversations or occasionally, multilateral conversations… It’s very much… a sense of ‘well, what are you up to and can we learn from you?’ or “What are your ideas?”, that sort of thing. But … none of that is productised; you can’t exactly hand code over … or [say] ‘Can we have your code please?’ because none of us are at that level, that I’m aware of.” PSM with higher organisational incomes tended to have used AI for longer, with one exception: ABC (Australia) had taken a slower, more cautious approach to adopting AI than other high-income organisations. Despite the high cost of developing AI and embedding it successfully in journalists’ routines, AI innovation was not the sole preserve of high-income PSM. In-house tools had been developed by some mediumand one low-income PSM in a comparatively short time. One medium income PSM, SR (Sweden), was consistently described as having become a leader in AI innovation by other participants. However, organisations that did have an AI strategy in place at the time of interviewing tended to have lower incomes. SABC (South Africa) is positioned at the bottom of the medium-income bracket, while PBCJ (Jamaica) and TRM (Moldova) were at the bottom of the low-income category. PSM with lower incomes said they struggled to afford AI applications that they believed to be secure and could not afford to employ AI specialists to assess or develop such tools. Low staffing levels meant that finding the ‘bandwidth’ to consider how to approach AI was also challenging. As a PBCJ executive put it, “We don’t even have the resources to think about AI … what is happening globally, how it will affect us and [how] to brace for that effect.” Questions for further discussion: • Given that most PSM in this study are relatively new to AI, how can PMA and other multilateral associations best assist and support them? • Do PSMs with more experience of AI have any normative obligations to help, work with, or simply listen to other PSM, especially those with lower incomes and those facing serious democratic threats? • What kinds of regional and crossregional networks might be appropriate, useful, and sustainable for PSM, PMA, and other supporting organisations? • How should “PSM” be defined to foster fruitful collective discussions about responsible AI? 8 europeanperspective.net/home 9 apnews.com/article/moldova-democracy-election-russia-disinformation-corruption-0a23e330da7121dbc34b085fc5d0d8ad 10 publicmediaalliance.org/how-can-ai-bolster-psm-mission Further regional conversations appear to take place between ABC (Australia) and others in Oceania, including RNZ (New Zealand). This helped offset the difficulties that other PSM in the region faced in funding the long-haul travel necessary to attend conferences outside the region. While PSM outside Europe were aware of the risk of discussions about PSM and AI becoming too Eurocentric, other regional hubs did not seem to operate as comparable hubs to the EBU. Sometimes there weren’t enough PSM to form a ‘critical mass’ and, sometimes, broadcasting unions and other regional organisations do not impose a definition of PSM on their members, which led to state-controlled media attending sessions on PSM and AI. In such contexts, PSM senior managers and trainers said they did not feel comfortable sharing much detail about their experiences or pilot projects. The lack of comparable regional hubs for dialogue between PSM seems likely to be a particular problem for PSM with lower organisational incomes, as they did not appear to participate in regular, informal, online conversations with other PSM, even though some faced serious internal and external threats. For example, a PBCJ (Jamaica) said they were aware of online conversations in Europe, as well as PMA’s international workshops, but low staffing levels meant they struggled to find time to participate or keep up with other coverage of AI. However, one organisation, PTS Taiwan, chose to deal with this issue by hosting one of their annual symposia on ‘The Challenges and Opportunities Facing PSM in the Age of AI’, in conjunction with PMA in autumn 202410. Public Media Alliance 10 11Part 1: Mapping how public service media use AI in journalism 4. PSM and AI tools 4.1. What AI tools are PSM using? 4.2. How much did PSM depend on the ‘Big Five’ technology companies? The 10 PSM which said they used AI to inform or support journalism were asked to name up to five tools , how they used them, and how they sourced them. In all, they named 42 tools, none of which dominated the sample (see Appendix for full dataset). These tools were used for a variety of purposes. The most popular was transcription, which was mentioned by all participants. The second most popular was image generation and editing, which were mentioned by half of the sample. Fewer participants mentioned AI-enabled translation, AI presenters, synthetic voice technology, and other activities relating to journalistic research, including image verification and facial recognition technology. However, the fastest growing use of AI involved repurposing news content for online services and social media platforms. Most uses did not differ significantly from those reported in the global JournalismAI project11, which included a small number of PSM alongside many other news organisations. In addition, most of the tools that PSMs used did not seem especially distinctive. Only eight of the tools used ‘officially’ by PSM (19% of the sample) were developed in-house, and only one tool (2%) was developed in collaboration with other PSM, via EBU. As Diagram 2 shows, the majority of AI tools were provided externally by commercial companies (27 tools, 64% of the sample)12. Although we can only report what participants told us, the ‘Big Five’ technology companies (Google, Apple, Facebook/Meta, Amazon and Microsoft) seemed to be less dominant than expected. Google and Microsoft supplied a total of three tools, cited by one participant each (Microsoft Github Copilot, Google Pinpoint and Google Reverse Image Search). Although Amazon and Microsoft products (e.g. AWSBedrock, AWSTranscribe, Microsoft Bing) underpinned a further five hybrid tools. So, the ‘Big Five’ either supplied or contributed to 19% of the tools ‘officially’ used by PSM, and these were deployed by 38% of study participants. However, these companies also appeared to provide other forms of infrastructure to PSM, analysis of which was outside the scope of this project. For example, Suspilne (Ukraine) decided to host its website on Google Cloud for security reasons, shortly after the Russian invasion. By contrast, other commercial companies supplied or contributed to 24 tools (57% of the sample), and 53% of participants relied upon them. The most commonly cited was Adobe (six participants referred to Firefly or Photoshop) and OpenAI (three participants cited ChatGPT, CustomGPT, Dall-E). OpenAI products (ChatGPT and Whisper) also underpinned a further two hybrid tools. PSM should be aware that OpenAI has been sued for training its tools using copyrighted material, scraped from the web13, and that its move to become a for-profit company is highly contested14. As the database of AI tools in the Appendix shows, a wide variety of commercial companies were mentioned (17 in total15, although the companies producing two other tools could not be traced16.) Common and emerging uses of AI will now be discussed in relation to specific tools. 11 lse.ac.uk/media-and-communications/polis/JournalismAI 12 We will reflect on the implications of PSMs’ use of commercial AI providers in our second report on the Responsible Procurement of AI. 13 nytimes.com/2023/12/27/business/media/new-york-times-open-ai-microsoft-lawsuit.html 14 vox.com/future-perfect/380117/openai-microsoft-sam-altman-nonprofit-for-profit-foundation-artificial-intelligence 15 These were: Adobe, Anthropic,Canva,Chartbeat, cSubtitle, ElevenLabs, EMEARobotics, GoodTape, Grammarly, GrayLark Technologies, Midjourney, OpenAI, OtterAI, Systran, Talkwalker, Trint, Wobby. 16 These were both Russian facial recognition technology tools, discussed in the final findings section: Findclone and Search4Faces. Diagram 2: AI tool provenance EBU (3%) In-house (19%) Commercial (64%) Hybrid (14%) Public Media Alliance 12 13Part 1: Mapping how public service media use AI in journalism 5. Speech-to-text tools 5.1. Transcription and captions 5.3. Translation 5.4. Variable accuracy of speech-to-text tools and minority or Indigenous languages 5.2. Speech-to-text tools and data privacy The most popular use of AI cited by PSM was speech-to-text transcription, which was mentioned by all 10 participants whose organisations used AI ‘officially.’ Transcription had multiple uses in journalism production. It aided conversations between journalists and news managers, especially within multilingual organisations, and facilitated journalistic research. AI-enabled transcription also helped to repurpose news content rapidly. For example, journalists often produced automated transcripts to help them convert broadcast interviews into shorter online articles. Speech-to-text tools were also used in publicfacing contexts to meet PSMs’ universal accessibility obligations by providing hearingimpaired audiences with full transcripts of podcasts and/or producing captions for TV and video. However, only 3 PSM said they currently used in-house models for AI transcription, all of which were high-or medium-income organisations. These were: ABC (Australia), NHK (Japan), and SR (Sweden). A fourth organisation, which chose to remain anonymous in this context, said transcription was one of the functions of its hybrid AI system. It is worth noting here that although SR (Sweden) is a medium-income organisation, it appeared to have done some of the most extensive development of in-house speech-to-text tools for universal access17. It provided in-house transcription for podcasts and other programming, and was rolling out AI-enabled captions for its app during winter 2024. SR had also developed an in-house sound optimisation tool to aid in-car listening for the hearing-impaired. Accurate transcription was widely seen as the precursor of AI-enabled translation. This interested PSM because of their commitment to diversity, but was of particular interest to multilingual PSM, as manual translation is expensive and takes so much time. Only three organisations mentioned that they currently use AI to aid translation. These were: NHK (Japan), SR (Sweden) and SRG SSR (Switzerland). The international arm of NHK (NHK World-Japan) appeared to use it most extensively, deploying an in-house tool to livestream news with English subtitles, automatically translated into nine other languages, and to provide online articles and Video-on-Demand in nineteen languages. However, the ‘unofficial’ use of AI translation tools could prove deeply problematic, especially when published without checking. A striking example was given by a senior manager at TRM (Moldova), who said that one of the network’s journalists had used an unauthorised AI tool to translate material about a sensitive court case for the Russian language section of the network’s website. The TRM journalist, who was on an overnight shift, had not properly proofread the article before publication. As a result, the network falsely reported that the court had “beheaded” the defendant, causing what the senior manager described as a “huge scandal” the following morning. The most common challenge mentioned by PSM in relation to speech-to-text tools was their accuracy could vary dramatically. English and Chinese Mandarin transcriptions were consistently identified as the most accurate because AI tools were trained using large corpora scraped from online sources. However, interviewees said AI tools still struggled to transcribe some English accents and dialects, such as New Zealand English. PTS (Taiwan) had also opted to use a Hong-Kong based specialist, cSubtitle21 (despite considerable tensions with China), because other AI transcription services could not differentiate sufficiently between Chinese Mandarin, Taiwanese Mandarin, and Hong Kong Cantonese. Data privacy, as well as universal access, also informed SR’s approach to the development of in-house speech-to-text tools. As a network executive said, “If we [used] external AI services, we can’t guarantee that they’re not training on our content—and we don’t want them to do that.” Other organisations used transcription tools provided by various commercial companies, including Sonix18, and a paid-for version of OtterAI. But some were becoming uncomfortable with this and were looking for alternatives. For example, one senior manager who preferred to remain anonymous in this context, explained: “Everyone at [our network] is concerned about where our information is going because we just don’t know. We did register with Otter.AI for transcription… [But] people who had their own accounts on their laptops [found that] every time they entered a meeting, it would say, ‘John’s AI’ and then it would randomly record stuff. It made people really nervous that they had signed up for a transcription tool, but all of a sudden their meetings were being weirdly recorded ... it was quite alarming.” In a follow-up conversation with another PSM in (northern) spring 2025, a senior manager at ABC (Australia) confirmed that the network had banned its staff from using Otter.AI, as it had failed all of ABC’s data privacy and security assessments. Some PSM chose to address concerns about data privacy by using GDPR-compliant providers (e.g., GoodTape19 and Trint20), even when they were based outside Europe. 17 thefix.media/2024/9/23/what-can-newsrooms-learn-from-swedish-radio-ai-strategy 18 sonix.ai 19 goodtape.io 20 trint.com 21 csubtitle.com Public Media Alliance 14 15Part 1: Mapping how public service media use AI in journalism Romansch Hakka Māori PSM working in minority languages experienced even lower rates of transcription accuracy because of the lack of readily available training data online. But building training corpora is an expensive, highly-skilled, and time-consuming task. SRG SSR (Switzerland) addressed this problem by commissioning a commercial start-up launched by a university to construct a corpus of Romansch22, a language spoken by fewer than 10,000 people in the Grisons canton, with five different idioms. Developing Indigenous languages corpora was widely regarded as the most difficult task of all: compounding the challenges of minority languages with concerns about Indigenous data sovereignty26. For this reason, Māori non-profit, Te Hiku Media27, secured funding from the New Zealand government to support its crowdsourcing of a Māori corpus and use it to train speech-totext and text-to-speech tools in a project known as Papa Reo28. Beginning in 2020, by 2024 Te Hiku Media had produced remarkably accurate tools29. This is particularly impressive since New Zealand speakers often “code-switch”: incorporating words from Te Reo Māori (the Māori language) and New Zealand English within a single utterance, creating additional technical challenges. Like PTS in Taiwan, RNZ (New Zealand) contributed decades of archival material to the Māori corpus. In return, the network will be permitted to use the tools produced by Te Hiku Media for transcription and on-screen captioning. THe RNZ interviewee also said they hoped that RNZ’s use of tools provided by Te Hiku Media would avoid perpetuating harmful historical biases within the network’s archival metadata. However, RNZ’s approach means that any other future translation tools the network uses (for example, to translate Mandarin Chinese), will have to be designed to be compatible with those provided by Te Hiku Media. However, this was not an option for low-income PSM, which tended to feed into governmentfunded projects run by non-commercial actors instead. For example, PTS (Taiwan) contributed archival material to an academic corpus-construction project supported by the government’s Hakka Affairs Council. Hakka, with its five dialects, is spoken by 15-20% of Taiwan’s population and is considered to be endangered23. The corpus, which was published in March 2024, took two years to build24. At the time of writing in (northern) spring 2025, PTS was collaborating with Taiwanese academics to use the corpus to develop live captioning for Hakka TV25. 22 slator.com/automated-romansh-translation-now-available-for-the-first-time/#:~:text=Now%2C%20for%20the%20first%20time%2C%20it%20 is%20possible%20to%20use,translation%20for%20the%20Romansh%20language 23 ieeexplore.ieee.org/abstract/document/10482979 24 hakkatv.org.tw/news-detail/1709288985235265 25 hakkatv.org.tw/news-detail/1679632232246710 26 gida-global.org 27 tehiku.nz 28 papareo.nz 29 rnz.co.nz/news/te-manu-korihi/527405/maori-media-company-head-peter-lucas-jones-named-on-time-magazine-list-for-preserving-tereo-through-ai 30 nytimes.com/2023/10/11/world/americas/jamaica-official-language-patois.html 31 web.sabc.co.za/digital/stage/editorialpolicies/Policies/SABC-Editorial-Policy-LANGUAGE.pdf 32 These are: Afrikaans, English, Ndebele, Northern Sotho, Southern Sotho, Tsonga, Tswana, Swati, Venda, Xhosa and Zulu. Sign language was added in 2023. 33 isca-archive.org/sigul_2023/markl23_sigul.html 5.5. Challenging cases: Jamaica and South Africa Two of the PSM most interested in developing AI translation in future were PBCJ (Jamaica) and SABC (South Africa). But both are likely to face pronounced technical challenges. Jamaican Patois or Patwa has a distinct grammar, syntax and vocabulary—drawing words from African, European and Asian languages30. Like most oral languages, Patois is spoken differently across the country, so there is no “gold standard” to measure accurate transcription against, let alone translation. SABC is committed to providing programming in all of South Africa’s official languages31, several of which are Indigenous32. “Code-switching” is common in South Africa: that is, when a speaker uses words from multiple languages in the same utterance. In addition, some languages like Xhosa are primarily oral, so are spoken differently across the country33. These compounded by the imposition of writing by colonial-era missionaries in ways that often bore little resemblance to how languages were actually spoken. Accurate AI transcription and translation continues to be a significant challenge for many PSM. Some of those facing the greatest technical challenges also have low organisational incomes. Questions for further discussion: • Might PSMs, which have developed their own transcription and translation tools, be willing to share some of what they have learned with others thinking of moving in this direction? • What are the pros and cons of using transcription tools from large companies without strong regulatory mechanisms versus using smaller, GDPR-compliant providers? • What are the pros and cons of contributing to government-funded corpus-construction projects in partnership with universities or nonprofits? Might those with experience of these projects be willing to share their experiences? • Is there scope for a working group or informal network of PSM working on virtual sign-readers? • How could PMA and other associations best support PSM coping with challenges regarding the transcription and translation of minority and Indigenous languages, especially organisations with lower incomes? For example, might it be useful to collate a list of academic experts skilled in developing minority language corpora, or facilitate collaboration between PSM with overlapping minority populations? Public Media Alliance 16 17Part 1: Mapping how public service media use AI in journalism 6. Images, AI presenters, and Synthetic Voices 6.1. Image generation and editing 6.2. AI Presenters and Sign-readers The second most cited area of current AI use was image generation and editing. This was mentioned by seven participants: ABC (Australia), PTS (Taiwan), RNZ (New Zealand), RTBF (Francophone-Belgium) and VRT (Flemish-Belgium), SRG SSR (Switzerland), and Suspilne (Ukraine). AI-enabled tools rapidly improved the efficiency of editing images required in high volumes for online publishing. However, AI-generated images were generally graphics and non-realistic illustrations used to “promote” other journalistic content on social media. So, AI-enabled image generation and editing both seem to have emerged in response to the longstanding pressure to repurpose content and compete for audiences’ attention on multiple online and social media platforms. Most PSMs used Adobe’s Firefly and Photoshop tools for image generation and editing. PSMs’ choice of these tools was influenced by familiarity and ease of access because, as one interviewee said, these products were already “on their desktop” via organisational subscriptions to Adobe’s Creative Cloud. Organisations also mentioned commitments made by Adobe and another commercial provider, Canva, to avoid training their tools using customer-created content. However, two PSMs said they used Midjourney, which has been sued in the United States for breach of copyright34. The case is complex, ongoing and no final judgement had been reached at the time of writing. The most serious ethical concerns raised by PSM senior managers involved the use of AI tools to create human-seeming synthetic ‘presenters’, which are becoming common in parts of Asia and the Middle East35, following the introduction of AI news anchors by China’s wire agency, Xinhua, in 2018. Most participants argued that AI presenters would not be accepted by PSM audiences, who might view them as deceptive, which risked damaging public trust in their journalism. In addition, participants were profoundly concerned about playing into the hands of malicious actors who already create ‘deepfakes’ of their content to spread misand disinformation. As an ABC interviewee (Australia) explained, if “…someone else makes up an avatar and fakes ABC News, you [would be] in much more difficult territory saying, ‘Oh no, no, that fake person that has been created is the wrong ABC avatar. This is the real ABC avatar’. AI anchors seem to be more acceptable to PSM in some parts of Asia. For example, Korean Broadcasting System (KBS) has introduced them to “read” local news bulletins36. PTS (Taiwan) also has a 2-D AI presenter called “Handsome P” which is used to host on special occasions or to present quizzes after the news37. While NHK (Japan) has previously used an animé character named “Yomiko” to read the news38, but withdrew her due to programming changes. However, NHK (Japan), PTS (Taiwan) and SRG SSR (Switzerland) are all working on developing virtual sign-readers because of their commitments to universal access. NHK hopes to use a virtual sign-reader called “Kiki” as an “ambassador” during the DeafOlympics in November 202539. But at the time of inquiring in January 2025, NHK researchers said that “Kiki” was not yet operational because of several “practical issues,” including translational accuracy. This is not surprising because signlanguage generation is a very challenging task, which academic researchers have been trying to solve for over two decades40. 35 theguardian.com/tv-and-radio/2023/oct/20/here-is-the-news-you-cant-stop-us-ai-anchor-zae-in-grants-us-an-interview 36 Seo, S. (2025) ‘Performing Publicness: AI and PSMs in South Korea.’ International Communication Association conference, Denver, CO, 12-16 June. 37 youtube.com/watch?v=lJC4uMekmX0 38 nhk.or.jp/strl/english/publica/bt/74/8.html 39 nhk-ep.co.jp/signlanguage/en 40 link.springer.com/article/10.1007/s10209-021-00823-1 34 reuters.com/legal/litigation/ai-companies-lose-bid-dismiss-parts-visual-artists-copyright-case-2024-08-13 Public Media Alliance 30 31Part 1: Mapping how public service media use AI in journalism Multifunction Copy editing/Proofreading Summarisation Automated generation of WhatsApp messages *In development: Title suggestion Translation Generate TikTok scripts  In-house N/a 1 Published atticles Currently journalistic research, reversioning, pilot extending functionality to aid ertciency Makes errors with larger atticles Scalability Workftow integration Multifunction Transcription Suggests video clips Hybriddetails not given Based on AWSBedrock, Anthropic Claude LLM AWSTranscribe 1 Video + transcription or image searches Transcribes video and proposes clips from longer pieces using transcription or image search Struggles with some contexts (e.g., spotts, press conferences) Multifunction Analysing data, time fflter, person detection, locations, organisations, plus transcription, audio search, text recognition in images External Google PinPoint 1 Text Investigative journalism Sometimes too eager to link people and locations. Can only transcribe one language per ffle. Research/Content generation External OpenAI ChatGPT 1 Prompt Online and broadcast coverage of ChatGPT only Privacy and Security concerns Research External OpenAI ChatGPT and Microso� Bing (integrated tool) 1 Prompt Assisted searches Research (Data journalism) External Wobby 1 Government Open Datasets Enables conversations with big datasets for investigative journalism Background in data journalism still necessary to interpret results. Some hallucinations. Research (Data journalism) External OpenAI Custom GPT Code CoPilot 1 Prompt Helps journalist write code for data journalism Requires expettise in coding Research (Investigative) External Findclone 1 Image Assists person-identiffcation Russian tool requires VPN, combined with other methods No subscription for staffi, so used at own risk/expense Research (Investigative) External EMEARobotics PimEyes 1 Image Assists person-identittcation Used in combination with other methods. No subscription for start, so used at own risk/expense Research (Investigative) External Search4Faces 1 Image Assists person-identittcation Russian tool, requires VPN, combined with other methods. No subscription for start, so used at own risk/expense Research (Fact/claim checking) External Google Reverse image search 1 Image Find image published on internet based on image uploaded Research (Fact/claim checking) External Graylark Technologies GeospyPro 1 Image Describes location where picture taken Accuracy limited, esp. if indoors Sound optimisation In-house 1 Audio Increase audibility, esp. in-car listeners Synthetic voices External ElevenLabs 2 Audio/Video Text -Clones voice for use in news, spoft, enteftainment -Breaking news when not enough start on shiff; when don’t have enough voices to translate speeches; voices online afticles for accessibility -Variable quality, e.g., poor expression/too fast/accent/pronunciation -Paftially free, only have one subscription. Initially piloted using presenters’ voices but have decided to use wholly synthetic Synthetic voices In-house N/a 1 Audio of presenter reading + Text Clones voice for use in TV news, political broadcasts, weather forecast and more. Limited to use in script-reading style speech Transcription In-house N/a 1 Accessibility, metadata generation Transcription External cSubtitle Pro 1 Text Used internally to produce video captions. Distinguishes three languages/dialects: Taiwanese Mandarin; Chinese Mandarin; and Hong Kong Cantonese Transcription is not su�ciently accurate to fully automate captions. Manual checks and corrections need to be carried out by journalists ttrst Transcription External OffierAI 1 Audio Transcribes radio interviews/repofts Accuracy poor, di�culty coping with dialect Transcription External Trint 1 Audio/video Summarisation for research purposes Scalabilityuse a dirterent tool for other purposes Copilot Public Media Alliance 32 33Part 1: Mapping how public service media use AI in journalism Transcription External Sonix 2 Audio/video -Transcribes Chinese and English -General transcription -Real-time accuracy needs improving -SAAS product so not used with highly conttdential content Transcription External Various -not named 1 Audio Radio journalism Accuracy too poor for publication Not all tools used by journalists satisfy organisational requirements re: contracts/data security. Transcription Hybridsome customisation AWSTranscribe but switching to Google (DeepMind) 1 Audio/video Generic journalism Varying quality of transcription per language Transcription Hybridcustomised internal intertace Systran’s FasterWhisper Uses OpenAI Whisper +a fast inference engine, CTranslate2 1 Audio, video, YouTube links Generic news, Scalability, limited use. Transcription Hybridcustomised internal intertace - 1 Audio AllRecognises local dialects and voices speaking in German, French, Italian, Romansh (more for international service) Quality/Accuracy - must be reviewed. Can only cope with monolingual transcription, cannot switch between languages Translation Hybrid – industry collaboration Microsoft 1 Published news Radio journalism. Uses own transcription tool in combination with Microsoft Copilot to translate, for internal use only. Translation Collaboration EBU EuroVOX 1 Published news Translation Translation External - 1 Any Integrated into journalistic workffows Accuracy/Quality, needs to be reviewed Video editing In house N/a 1 News video Auto-edits longer video repo�s into summaries for online news. Limited to structured news segments, as it relies on the structure of news item . publicmediaalliance.org Published by the Public Media Alliance © Public Media Alliance 2025 July 2025 Image: Public media values. Produced using Microsoft Copilot. Prof Kate Wright and Kristian Porter