Media Self-Regulation in the Use of AI: Limitation of Multimodal Generative Content and Ethical Commitments to Transparency and Verification

Abstract

The expansion of the use of artificial intelligence (AI) across different stages of production and distribution in journalism is opening a debate on its applications within newsrooms and in business models. This research studies how different media outlets, media groups and institutions are beginning to create internal regulations for the use of AI, both from a technical and an ethical perspective. To do so, an international sample (N = 45) of editorial stylebooks and internal self-regulatory guidelines published between 2023 and early 2025 have been compiled—all links are openly available here—and put through a process of content analysis. The results indicate that the self-regulatory guidelines emerge from an individual initiative of the media themselves, with a focus on limiting the use of generative AI, particularly in text creation. The guidelines emphasize ethical commitments such as transparency, content verification, and respect for data and copyright while underlining the importance of human oversight. Key objectives include avoiding bias, ensuring information quality, and strengthening audience trust. Despite progress, regulation remains in its early stages and requires continuous adaptation to keep pace with technological advancements.

1. Introduction

While there are studies exploring the adoption of AI in journalism, few address, in detail, how news organizations have developed internal guidelines for the self-regulation of AI use. This study seeks to fill this gap by investigating emerging self-regulatory practices with a focus on transparency, human oversight, and ethical management of AI. This gap is addressed through the analysis of 45 recently published guidelines and style manuals, allowing for a comprehensive understanding of current self-regulation practices.

After seven decades of development, artificial intelligence (AI) has recently entered the mainstream with accessible tools like Chat-GPT (Pavlik, 2023). AI’s ability to process algorithms, learn from data, and perform tasks with or without human oversight (Rouhiainen, 2018) has given rise to two models: analytical and generative. Although the two are related, their differences lie in the capabilities and applications required of each. Generative AI is capable of creating multimodal content (Feuerriegel et al., 2023) and is especially used in the creation of text, images, videos, music, etc., whereas analytical AI focuses on data analysis and decision making based on that data (Lim et al., 2023; Jiménez Martínez et al., 2025) and is used in the identification of patterns and prediction of results (Codina, 2024).

AI’s applications span various sectors, including journalism, where it supports processes from data collection to automated content creation and personalized dissemination (Túñez-López et al., 2021; Sánchez-García et al., 2023). Described as ‘automated journalism’ or ‘robojournalism’ (Caswell & Dörr, 2018; Latar, 2018), AI addresses growing user demands (Canavilhas & Giacomelli, 2023).

In the professional field, the implementation of AI in newsrooms has barely been in development for a decade. It began in an experimental way with examples such as the use of Quakebot for The Los Angeles Times with an earthquake alert in 2014, and the automation of content with AI baptized as Heliograf in The Washington Post in 2016 (Flores-Vivar, 2020). Since then, different national and local media have been introducing experimentation in a progressive way, although with limitations and even some distrust in companies (Sánchez-García et al., 2023), especially because of the challenges that still represent the generation of automated content in terms of errors of the so-called ’hallucinations’ of AI, copyright, the need for verification and issues of lack of transparency. However, despite the fact that some of its risks must be constantly assessed, the effectiveness of generative AI has already been demonstrated in newsrooms, such as in some documentation work; in the creation of economic content, sports content (Segarra-Saavedra et al., 2019), and electoral content, such as the case of RTVE in Spain; in the organization of data (Biswal & Gouda, 2020) or even in the creation of avatars and synthetic products. Its usefulness is being routinely demonstrated in specific automatable tasks such as audio-to-text transcription, summary generation, automatic translation, audiovisual editing, metadata or documentation (Lopezosa et al., 2024), besides audiovisual editing, social media management, trend analysis and content distribution, among other repetitive tasks that can leave more room for creativity to the journalist. It is confirmed that the use of AI in an integrated way throughout the journalistic process is possible (Sánchez-García et al., 2023).

From an academic research point of view, a debate has arisen, highlighting both the benefits of AI for journalism and its associated risks. Different academic studies highlight three main benefits of the use of AI in newsrooms: the ability to increase the accuracy and verification of content (Sandoval Martín et al., 2019), reduce work time and costs (Hansen et al., 2017; Hassan & Albayari, 2022) and give space to journalists to explore their creativity and grow in other areas (Ufarte Ruiz & Manfredi Sánchez, 2019). Hassan and Albayari (2022, p. 175) consider that the survival of the press “lies in its adoption of the concepts of flexibility, sustainability, coexistence with uncertain conditions, and the transition to the new normal (...) including the acceleration of the adoption of advanced technologies such as artificial intelligence”.

From a more pessimistic view, different threats stand out, specifically ethical debates (Noain Sánchez, 2022) and the perceptions of journalists themselves about being able to retain their jobs in a debate that is divided between integration or replacement (Saad & Issa, 2020), resistance to change (Sánchez-García et al., 2023), lack of budget or technical abilities and the influence of the institutional landscape (de-Lima-Santos & Ceron, 2021).

Both perspectives highlight the need to deepen the debate and, especially, to analyze how the media propose to regulate or establish internal guidelines for the uses of AI by their journalists. This is not only because of how these technologies fit into the newsrooms but also because of journalistic work processes, which introduce new challenges to AI usage before—such as care in training AI with heterogeneous data and instructions to avoid bias—and after its application—such as verification and data processing (Rojas & Ramírez, 2023). The following is a set of “issues to regulate or legislate and assess from an ethical-philosophical point of view” (Diego Olite et al., 2023, p. 4) that address transparency and accountability of journalists and media outlets: “To what extent can news media and journalism be held to account for the use of AI-driven tools? And to what extent can the implementation of AI tools in news media be effectively regulated?” (Porlezza, 2023, p. 372). These are questions that have not yet been investigated academically from a perspective that addresses the decisions made by media outlets on the internal use of AI, which this research attempts to address.

This study offers a unique contribution to the field, mapping the first concrete initiatives for AI self-regulation in journalism and highlighting critical areas for future academic reflections and journalistic practices.

Self-Regulation and Ethics of Artificial Intelligence Usage in the News Media

Researching the impacts of AI in society requires an understanding of its ethical implications. AI ethics, also known as “machine ethics”, appears as an emerging interdisciplinary field that studies theories, guidelines, policies, principles, rules, and regulations related to artificial intelligence, with the goal of programming it according to values and principles of right and wrong (Huang et al., 2022).

Given the risks and challenges posed by the advance of AI and the uncertainty of its future, ethical issues and limitations are among the priorities of supranational organizations. In the last five years, key recommendations have been issued in this regard. The report on the ethical use of AI agreed by UNESCO in 2021 focuses on preserving human rights and equality in the face of technological progress and establishes concrete recommendations on its use that can be summarized in three principles: transparency, responsibility and equity (UNESCO, 2021). Along the same lines, the OECD, in its first recommendations (2019, updated in 2024), prioritizes respect for human rights and democratic values, recommending an innovative and, at the same time, reliable use of AI-based concepts such as transparency, explainability and accountability (OECD, 2019–2024). From the professional field specializing in technology, the recommendations made by the Institute of Electrical and Electronics Engineers (IEEE), the main body that brings together the engineering sector, recommends safety, transparency and accountability in its development through specific principles (IEEE, 2019) that has been progressively updated through specific standards and frameworks are also worth noting.

Based on these non-binding ethical frameworks, more specific legislative frameworks have begun to develop. For the time being, the most successful approach has been the European Union’s AI Regulation, popularly known as the AI Act (European Parliament, 2024). This is the first international AI regulatory law that limits the uses of artificial intelligence by setting different levels of risk, and among its aims is the effort to protect human rights and to support a human-centric, ethical and responsible use of AI (Helberger & Diakopoulos, 2022; European Parliament, 2024). The AI Act project started in 2018 when the European Commission published a communication on AI governance, which was followed by the announcement of a European AI Strategy and the creation of a High-Level Group on Artificial Intelligence.

The AI Act, which will require subsequent development in each EU Member State, does not include regulation in specific sectors, which, as is the case for media, depend on their own internal criteria of use. According to Porlezza’s study (2023), existing regulations on AI in the European Union and the Council of Europe rarely refer to media or journalism. When the media is mentioned, it is generic, and the focus is on four main axes: risks posed by AI usage, literacy, social responsibility and press freedom and the need to include the media as part of the AI policy and governance debate. However, the guidelines and recommendations on its use are generic, leaving ‘many open questions’ and ‘uncertainty’ about how to delimit it in the Communication sector (Porlezza, 2023).

Some researchers and scholars delimit journalistic ethics in the use of AI in specific aspects that are considered key to the sector. Reinecke et al. (2021) point out aspects to regulate, such as security, privacy and safety but also principles, standards, protocols and guidelines. The authors highlight, among other aspects, the importance of copyright regulations (Trapova & Mezei, 2022) for AI-generated content used by the media or the relevance of transparency ‘in order to nurture a trusting relationship between journalistic media and their audience and to provide the latter with an instrument they can use to report suspected error or unethical conduct in news coverage’ (Haapanen, 2022, p. 81). The need to address the identified gap in delimiting the use of AI is seen as global, affecting both countries in the developed North and the so-called Global South, where the expansion of the use of newsroom automation is confirmed, such as in sub-Saharan Africa (Gondwe, 2023), with codes of conduct and training (Radcliffe, 2025) that are becoming increasingly necessary in these geographic areas as well.

Like other types of journalistic guidelines, these regulations are developed and made explicit in two types of documents: style books and manuals and codes of ethics. The former refers to purely journalistic norms and criteria, including the correct and standardized use of language and linguistic formulas—abbreviations, acronyms, etc. (Font, 2008). The latter is a compilation of moral criteria and forms of conduct established and assumed by professionals in a given sector, such as journalists’ professional organizations and associations. These include, for example, ethical codes and recommendations on the coverage of sensitive issues, such as humanitarian catastrophes and tragedies, communication for development or violence against women, among other ethical aspects related to the corresponding editorial line. Thus, it is a matter of delimiting questions of journalistic quality and ethical limits that, in some way, commit them to their readers.

This type of manual and guideline constitutes a fundamental tool in the history of journalism, implying a public recognition of the ethical conscience of a given profession or activity (Aznar, 1997). In the media sector, these regulations and ’self-regulation’ manuals have evolved depending on the social, cultural and technological context. Thus, the guidelines of traditional media have been adapted to the digital environment, regulating, for example, broadcasting in digital environments and the journalistic use of social networks as a source or as a distribution channel, among other aspects (Confino, 2010). And in this evolution, it is again necessary to adapt to the new context of the expansion of artificial intelligence as a statement of intent to the risks and ethical dilemmas that shape issues such as automated content creation, authorship rights, automated content, data protection or cyberattacks, among other risks (Rojas & Ramírez, 2023).

In this context of technological and journalistic change, it is pertinent to ask how the media and media organizations propose to regulate their use within their newsrooms and business models and whether they are already doing so. This research is relevant given the limited academic studies on the subject, limited to specific topics, such as the implications of copyright laws adapted to the use of AI in countries like China, Europe, and the United States (Kuai et al., 2022; Kuai, 2024); AI regulatory frameworks at institutional and European level (Helberger & Diakopoulos, 2022; Porlezza, 2023); and specific self-regulation in European media councils, the latter ‘waiting and watching’ (Haapanen, 2020, p. 5). It is, therefore, a matter of providing solutions to the knowledge gap on the ‘self-regulation’ of the media in the face of AI.

2. Objectives and Background Assumptions

The main objective of this research is to analyze self-regulation proposals on the usage of artificial intelligence tools in the news media, which have been published by companies, collectives and journalistic entities (O1). The study unfolds into three specific objectives:

o1. Collect and review different editorial style manuals and/or internal guidelines that include recommendations and regulations for the use of AI in international media (local, regional and/or national) and gather their metadata in open access.

o2. Analyze the permitted, limited and/or prohibited uses of AI in journalistic production, especially around the generation of multimodal content.

o3. Identify the ethical stance and role of human oversight in the application of AI, transparency of usage and data verification.

The design of these research objectives is based on a starting assumption, based on an initial field observation, that the media are particularly concerned about the use of AI and have recently begun to publish recommendations in their newsrooms focused on delimiting the use of generative AI and establishing ethical commitments to transparency, also seeking audience trust.

3. Materials and Methods

This research uses a content analysis methodology, one of the most widespread methods in the Social Sciences, as it allows the researcher to systematically and objectively study the characteristics of messages, identify and quantify patterns and infer meanings in order to draw valid and replicable conclusions (Aigneren, 1999) through the establishment of categories and variables (Cauas, 2015).

One of the difficulties posed by the object of study lies in the complexity of gathering a limited and systematic international sample of AI media self-regulation guidelines due to the dispersion of their open publications and the absence of previous work on the subject due to the recent nature of the topic. For this reason, different systems were used to locate the sample, starting with the ‘snowball’ sampling technique recognized in academic literature (Van Meter, 1990; Voicu & Babonea, 2011; Alloatti, 2014), typically recommended for experimental studies and especially advisable when the sample is small, limited and difficult to find.

Thus, the final refined sample (N = 45) was obtained with two collection patterns. On the one hand, non-systematized random searches were conducted across the academic and grey literature, among which the international lists provided by Liuzzi (2024), Mayoral Sánchez et al. (2023) and Repilado (2024) stand out. These contributions are complemented, on the other hand, by manual searches with the keywords ‘artificial intelligence guidelines’ or ‘manual of style’ combined with the names of relevant media outlets, international news agencies and media groups and other entities of various kinds. Both findings feed back into the aforementioned ‘snowball’ process, resulting in a final sample of N = 45, which includes 38 guidelines, 5 stylebooks, 1 standard and 1 usage policy published between 2023 and 2024 and beginning of 2025 across 26 media outlets and 18 international groups and entities.

The snowball sampling method was selected due to the limited availability of publicly accessible self-regulation guidelines on AI in journalism. As this field is nascent and evolving, relevant documents are often scattered, unpublished or inconsistently indexed. Snowball sampling effectively identified resources by leveraging connections between organizations and researchers, a technique suited for studies with small or hidden populations (Van Meter, 1990; Voicu & Babonea, 2011).

This method, however, carries limitations such as selection bias and the over-representation of proactive or interconnected entities. To address these issues, random searches in the academic and grey literature, along with manual keyword-based queries, complemented the sampling process. Despite these efforts, the findings are exploratory and not generalizable, highlighting the need for future research with larger, systematically sourced datasets.

The limited sample size also reflects the scarcity of published guidelines during the 2023–2024 period. This exploratory study aims to provide an initial overview of practices rather than a comprehensive inventory. Nonetheless, the sample includes a diverse range of media outlets, organizations, and geographies, offering a representative cross-section of available resources and can serve as a reference for the media’s first steps in ’self-regulating the use of AI’ and can serve as models for other similar entities.

Each unit of analysis is subjected to a coding sheet based on five categories (

Table 1. Coding sheet for the analysis of self-regulation and ethical dilemmas of AI in international media.

C1. Identification	-Media -URL -Date -Country
C2. Type of AI and use	Generative AI	1. Prohibition 2. Limitation 3. Wide use 4. Not mentioned
	Analytic AI
C3. Multimodal use	Regulation on text generation	1. Yes  a. No limits  b. Limited 2. Prohibited 3. Not mentioned
	Regulation of photo and/or video generation
C4. Ethical regulation	Allusion to the importance of respecting copyright	1. Yes 2. Not mentioned
	Transparency measures
	Allusion to verification and ‘fact-checking’
	Regulation of data processing and protection
C5. Human supervision	Reminder of the importance of human supervision	1. Yes 2. Not mentioned

) that allow for the delimitation of the object of study—AI self-regulation in the media—and analysis across 13 variables.

C1. Identification. This includes data such as the type of document (style guide or guidelines), the responsible medium, group or entity, the URL, the date of publication and the country of origin.

C2. Types of AI and their usage. This includes the tools and applications mentioned (generative or analytical) and whether the document prohibits, totally permits or does not mention their use.

C3. Multimodal use. This includes indications on the use of AI to generate text, photographs and/or videos and graphics and whether it is allowed (with or without limits), prohibited or not mentioned.

C4. Ethical regulation. Code whether the document includes references to ethical issues such as copyright, transparency, verification and fact-checking or the processing of personal data.

C5. Human oversight. This determines whether or not the documents include a reminder about the importance of human oversight in the AI journalistic process.

4. Results

The main results of the analysis of media self-regulation documents on the use and application of AI tools are presented below, following the order of the proposed study categories through a comparative analysis. Due to the size of the sample (N = 45), the data are given in nominal numbers.

Firstly, we present, as a main result of the study, the open compilation of the guidelines and stylebooks found in this experimental sample (

Table 2. Sample of editorial style books and guidelines by news media, groups and entities on AI use (2023–2024). Click on each medium to consult the original document ‎¹.

(A) News Media Outlets and Agencies
Type of Document	Media/Agency	Date	Country
Guidelines	Aftonbladet	5 April 2023	Sweden
Editorial style book	Agencia EFE	30 January 2023	Spain
Guidelines	Al Jazeera	10 July 2023	Qatar
Guidelines	Algemeen Nederlands Persbureau (ANP)	April 2023	Netherlands
Guidelines	Associated Press	16 August 2023	USA
Guidelines	BBC	February 2024	United Kingdom
Guidelines	Business Insider	April 2023	USA
Guidelines	CBC	12 June 2023	Canada
Editorial style book	El País	19 January 2025	Spain
Guidelines	Expresso	14 June 2024	Portugal
Guidelines	Heidi News	11 April 2023	Switzerland
Editorial style book	Hong Kong Free Press (HKFP)	No date available	China
Guidelines	Jot Down	April 2024	Spain
Guidelines	Mensagem de Lisboa	10 May 2024	Portugal
Guidelines	New York Times	9 May 2024	USA
Guidelines	Norwegian Tabloid	12 April 2023	Norway
Guidelines	The Guardian	16 June 2023	United Kingdom
Guidelines	Thomson Reuters	No date available	Canada
Guidelines	Todo Jujuy	No date available	Argentina
Guidelines	Wired	No date available	USA
Guidelines	NPR	2024	USA
Guidelines	Núcleo	No date available	Brazil
Guidelines	The Globe and Mail	15 June 2023	Canada
Guidelines	USA Today	4 December 2023	USA
Guidelines	Verdens Gang (VG)	12 April 2023	Norway
Guidelines	Volkskrant	No date available	The Netherlands
(B) Groups and Media Organizations
Type of Document	Group/Entity	Date	Country
Guidelines	Atresmedia	8 February 2024	Spain
Editorial style book	CDMSI	30 November 2023	European Union
Guidelines	DJV	No date available	Germany
Guidelines	DPA	3 April 2023	Germany
Usage policy	Estadão (Group Estado)	23 November 2023	Brazil
Guidelines	Flemish Organization for Journalism	No date available	Belgium
Guidelines	Groupe Les Echos Le Parisien	24 May 2023	France
Guidelines	Mediahuis	No date available	Belgium
Guidelines	News Media Alliance	24 April 2024	USA
Editorial style book	Nordic AI Journalism	2024	Sweden
Rules of use	Radiotelevisión española (RTVE)	5 September 2024	Spain
Guidelines	Ringier	30 May 2024	Switzerland
Guidelines	RSF	10 November 2023	France
Guidelines	RTDNA	No date available	USA
Guidelines	RVDJ	No date available	Belgium
Guidelines	Sanoma	16 February 2024	Finland
Guidelines	Sociedad Interamericana de Prensa	6 September 2023	Latin America
Guidelines	STT	No date available	Finland
Guidelines	TT Nyhetsbryan	No date available	Sweden

¹ Online access to the guides and manuals has been provided from April 2024 to January 2025.

) with access to their content and with the first metadata of their identification (C1): the responsible medium, agency, group or entity, the date of publication and the country to which it belongs as an expandable list. The titles of each of them are linked to the corresponding document analyzed, providing open access to the entire sample.

4.1. Levels and Types of AI Regulation in International Media

The first category of analysis (C1) allows us to identify the general characteristics of the sample (N = 45). The documents on AI usage produced by media outlets and news agencies are grouped into two main types: guidelines or ‘guides’, in the form of a ‘declaration of principles’, to help guide journalists’ work (39), and so-called ‘style books’ (both in PDF and web format), which provide guidance on standards and recommendations to ensure consistency and quality in the writing and presentation of media content (5). All of them are published openly and are accessible to the general public on the organizations’ own websites, which indicates a first exercise in transparency on the part of the media outlets analyzed, both towards their employees and towards the audience. Exceptionally, the sample also includes the first internal regulation on the use of AI, approved by the public media group Radio Televisión Española (1), and a usage policy of the Brazilian Estado Group, found in the media Estadão (1).

The breakdown of the data makes it possible to distinguish regulations across several levels: on the one hand, self-regulation of the media as individual entities (26) and, on the other, recommendations established either by media groups, such as Atresmedia or RTVE; by media organizations and alliances, such as Nordic AI Journalism or the Inter-American Press Association or by state press associations and institutions operating at the EU level, such as the Council of Europe (specifically, the Steering Committee on Media and Information Society, CDMSI) (19). Regarding the dates of publication, it is worth noting that these have only recently been published, which confirms that the process is only beginning. Most of the documents included in the analysis were published between mid-2023 (20), 2024 (11), 2025 (1), and others are undated (13).

The coding of the country of origin allows us to determine that there are more regulations at a national level (20) than at the regional or community level (with only one document in the European Union and Latin America, respectively) or local level (with only one case found in the Portuguese Mensagem de Lisboa).

The United States leads the sample with 8 manuals, followed by Spain (5), and, with 3 manuals or guidelines each, Belgium, Sweden and Canada.

The geographic diversity of the materials found, as well as the variety in size of the media entities that have published it—local, regional, national—reflect that the drive for AI ’self-regulation’ in the media is an international phenomenon that affects large and small entities in all countries of the world.

4.2. Types of AI and Regulation of Their Use (C2)

The analysis of the types of AI mentioned across media guidelines and editorial style manuals indicates a greater concern for the regulation of generative AI—AI that creates content—which is specified in 42 of the 45 manuals analyzed. Analytical AI—AI that reads datasets to identify patterns and make predictions—is mentioned in 15 manuals, with more generic references to data use and personalization.

The concerns around generative AI are expressed in the limitation of its use on the majority of the documents (39), as opposed to a broader or ambiguous permissiveness (3) that corresponds to organizations such as Nordic AI Journalism, the journalists’ association RVDJ or the European Commission, which, operating outside of the media outlets themselves, seek to establish generic recommendations which will then be subjected to the final responsibility of editors and publishers.

The limitations include the use of generative AI only as an aid to the journalist, never as a substitute: ‘We will never use AIs to generate the entire content of a publication, as an editor or final producer’ (Núcleo); “AI will be at the service of professionals to improve the quality of journalistic work and not to replace it” (El País); ‘We do not intend to publish editorial content generated in whole or even in part by an AI without editorial and human supervision’ (Les Echos-Le Parisien). The documents also associate this limitation with their commitment to the supervision of the AI-generated content and its verification: ‘There is no room for unreliability in our journalism, nor our marketing, creative and engineering work. At a simple level, this means that the use of genAI requires human oversight’ (The Guardian) or ‘It is essential to review, check and verify every input we take from AI tools’ (EFE).

4.3. Internal Recommendations on the Multimodal Use of AI (C3)

The analysis of the mentions of multimodal use of AI in the media shows it is typically applied to three main types of content: text generation, photographs and/or videos and graphics (Figure 1). When the use of these resources appears in media guides and guidelines, it is with reservations, i.e., with limitations in scope of usage or limited to specific use cases. The only exception is the generation of graphics, which is allowed without limits in Norwegian Tabloid (Norway) and Núcleo (Brazil).

The use of generative AI is clearly limited in text creation (32), which is not expressly prohibited in any of the analyzed documents but is also not allowed without limitations. Its application is limited to specific tasks, such as translation, proofreading, synonym search and headline writing.

The use of generative AI to create multimedia content such as photographs or videos reflects a greater disparity of criteria: nine media outlets, such as Hong Kong Free Press (China), Mensagem de Lisboa (Portugal) or Business Insider (USA), explicitly prohibit it. Some of them make very specific exceptions, such as when AI is the news itself (Associated Press, TT Nyhetsbryan) or when it is used to create illustrations (STT, Verdens Gang). Most of the books and guidelines analyzed, however, do allow a reserved use of this technology (24) for image editing, cropping or processing, although they are stricter regarding the creation of photorealistic content (banned in media outlets such as The Globe and Mail, Norweigan Tabloid, USA Today and Verdens Gang).

The generation of AI graphics is the least mentioned multimodal use. Even so, an explicit ban is included in media outlets’ documentation (The Volkskrant, TT Nyhetsbryan, Mensagem de Lisboa, Estadão and STT), and limitations are found in 17 of them. The use cases for this type of content are less explicit, with general allusions such as ‘requiring the consent of editors-in-chief’ (ANP).

Each media outlet’s self-regulatory guidelines reflect a spectrum of measures, from the strictest and most explicit to the laxest. The Associated Press, for example, stipulates that text generation will only be allowed internally and in experimentation, while the CBC allows it ‘provided it is identified and approved’. This lack of specificity is also identified in media organizations and groups, such as the Flemish Organization for Journalism, which makes each media outlet responsible for the decision of whether to use these tools or not without providing specific guidelines or example cases.

On the other hand, RTVE’s regulations stand out as some of the most specific in the prohibition of practices, listing four forbidden usage cases (deployment of subliminal techniques, exploitation of the vulnerabilities of a specific group of people, profiling of people according to their behavior, or facial recognition systems), although they do not mention any permitted uses of generative or analytical AI.

Table 3. Analysis of media style books and guidelines regulating the use of AI in journalistic work: multimodal use allowed (✓) or limited (x) in alphabetical order.

Mutimodal Use (C3)
	Text Generation	Multimedia Generation
Aftonbladet	✓ Suggest ideas or headlines	✓ No photo manipulation
Agencia EFE	✓	x Only specific cases
Al Jazeera	Not mentioned	Not mentioned
Algemeen Nederlands Persbureau (ANP)	✓	✓
Associated Press	x Only internal experimentation, not publishable	x Only if an AI-generated image is the subject of the news story
Atresmedia	Not mentioned	Not mentioned
BBC	x Not informative content	x Can be used in graphics and deepfakes for anonymity, no deception
Business Insider	✓	x Will check again in the future
CBC	✓ As long as it is identified and approved	✓ As long as it is identified and approved
CDMSI	✓ With supervision	✓ With supervision
DJV	✓ With participation of the journalist	Not mentioned
DPA	✓ Transcriptions, translations	Not mentioned
El País	✓ Limited use with oversight and transparency	✓ Do not create content that confuses with reality. No full illustrations
Estadão	✓ Transcriptions, translations, data analysis, text edition	x Generating photos, illustrations, videos or audio is prohibited
Expresso	✓ As a support tool	✓ As a support tool
Flemish Organization for Journalism	✓ Dependent on editorial decisions	✓ Dependent on editorial decisions
Groupe Les Echos Le Parisien	✓ With supervision	✓ With supervision
Heidi News	✓ Synonyms search and proofreading	✓ Certain cases and illustrative purposes
Hong Kong Free Press (HKFP)	x Except for summarizing or in-house tools	x
Jot Down	✓ With reservations	✓ With reservations
Mediahuis	Not mentioned	Not mentioned
Mesagem de Lisboa	x Only translations	x
New York Times	Not mentioned	Not mentioned
News Media Alliance	✓ As long as it is a transparent use and respects copyright	✓ As long as it is a transparent use and respects copyright
Nordic AI Journalism	✓ Text review, transcriptions	✓ Document states that any limitations do not fall within its purview
Norwegian Tabloid	Not mentioned	✓ Image processing and cropping ✓ Illustrations, graphics x Photorealist images or editing
NPR	Not mentioned	Not mentioned
Núcleo	✓ To summarize, transcription improvements	✓ Only urgent cases
Radiotelevisión española (RTVE)	Not mentioned	Not mentioned
Ringier	Not mentioned	Not mentioned
RSF	✓ With reservations	✓ With reservations
RTDNA	Not mentioned	Not mentioned
RVDJ	✓ Dependent on editors criteria	✓ Dependent on editors criteria
Sanoma	Not mentioned	Not mentioned
Sociedad Interamericana de Prensa	Not mentioned	Not mentioned
STT	✓ Automated translation, suggestions of headlines x No content writing	x Only illustrations
The Globe and Mail	✓ Headline ideas x Not for summarizing or writing	x No photorealistic images, only basic editing
The Guardian	✓ Ideas, proofreading and suggestions	Not mentioned
Thomson Reuters	Not mentioned	Not mentioned
Todo Jujuy	✓ Summaries, improving or cutting out wording	✓ Illustrations in urgent and exceptional cases
TT Nyhetsbryan	✓ As long as it has been started and reviewed by a journalist	x Only when the multimedia content is the news itself
USA Today	✓	x No photorealistic images, always with supervision
Verdens Gang (VG)	Not mentioned	x No photorealistic images or complex editing ✓ Illustrations and graphics
Volkskrant	x Only for transcriptions or text-audio translations	x Except to show how AI works
Wired	x Only exceptional, like text suggestions or cutting out wording	x Only in certain conditions and without replacing files

shows the individually analyzed measures by the media and entities, after the application of the categories and variables of the study.

The analysis of the main results in Table 3 shows a majority of the recommendations made by the media on the use of AI for text creation. Thus, 30 of the guidelines make specific mention of its use in text format, and of these, 100% limit its use to specific tasks that are permitted, such as text translation, audio transcription, summaries, or the contribution of ideas, mainly under the supervision of the journalist. Some expressly prohibit the creation of news content with AI (BBC), and most express it in other ways, such as allowing its use only internally and not for publication. As for the multimodal use of images and other types of content, it is expressly delimited in just over half of the references (27), of which two cases prohibit its use and the rest limit its use with three main ideas: do not use it to create realistic images that confuse with the normative reality; use it limited to AI topics; it is allowed for infographics, with identification of the authorship of the AI.

4.4. Regulations and Ethical Considerations on AI Usage (C4)

Regulations on ethics and good practice are set out in detail in

Table 4. Analysis of media style books and guidelines that regulate the use of AI in journalistic work: ethical dilemmas in copyright, transparency and data. (✓ means: concept mentioned).

Ethical Regulations (C4)
	Respect for Copyright	Transparency for the Audience	Importance of Fact-Checking	Personal Data Protection
Aftonbladet	Not mentioned	✓ If the content has impact	Not mentioned	✓
Agencia EFE	Not mentioned	✓	✓	Not mentioned
Al Jazeera	Not mentioned	✓	✓	✓
Algemeen Nederlands Persbureau (ANP)	Not mentioned	✓	✓	Not mentioned
Associated Press	Not mentioned	✓	✓	✓
Atresmedia	✓	✓	✓	✓
BBC	✓	✓	✓	✓
Business Insider	✓	Not mentioned	✓	✓
CBC	✓	✓	✓	✓
CDMSI	✓	✓	✓	✓
DJV	✓	✓	✓	✓
DPA	Not mentioned	✓	✓	Not mentioned
El País	Not mentioned	✓	Not mentioned	Not mentioned
Estadão	✓	✓	✓	✓ Includes media business data protection/management
Expresso	✓	✓	✓	✓
Flemish Organization for Journalism	Not mentioned	✓	Not mentioned
Groupe Les Echos Le Parisien	✓	✓	Not mentioned	Not mentioned
Heidi News	Not mentioned	✓	✓	Not mentioned
Hong Kong Free Press (HKFP)	✓	Not mentioned	✓	Not mentioned
Jot Down	✓	✓	✓	Not mentioned
Mediahuis	✓	✓	✓	✓
Mesagem de Lisboa	Not mentioned	✓	Not mentioned	✓
New York Times	Not mentioned	✓	✓	Not mentioned
News Media Alliance	✓	✓	Not mentioned	✓
Nordic AI Journalism	Not mentioned	✓ In what is published not mentioned Internal use tools	Not mentioned	Not mentioned
Norwegian Tabloid	✓	✓	Not mentioned	✓
NPR	✓	✓ When its use has a relevant role	✓	✓
Núcleo	✓	✓	✓	Not mentioned
Radiotelevisión española (RTVE)	✓	✓	✓	✓
Ringier	✓	✓	✓	✓
RSF	✓	✓	✓	✓
RTDNA	✓	✓	✓	✓
Sanoma	Not mentioned	✓	Not mentioned	✓
Sociedad Interamericana de Prensa	✓	✓	Not mentioned	✓
STT	Not mentioned	✓	✓	✓
The Globe and Mail	✓	✓	✓	✓
The Guardian	✓	✓	Not mentioned	Not mentioned
Thomson Reuters	Not mentioned	✓	✓	✓
Todo Jujuy	Not mentioned	✓	✓	Not mentioned
TT Nyhetsbryan	✓	✓	Not mentioned	✓
USA Today	✓	✓	✓	✓
Verdens Gang (VG)	Not mentioned	✓	✓	✓
Volkskrant	✓	✓	✓	Not mentioned
Wired	✓	✓	✓	Not mentioned

, which reflects how they are present in most of the guidelines and style guides under analysis (summarized in Figure 2). Most of the sample includes references to transparency measures (43), followed by references to the verification and fact-checking of AI-generated content (34), while references to copyright (29) and data processing regulations (28) come in third and fourth place.

The transparency measures proposed by these institutions and organizations (43) focus on the general requirement to label and identify content generated by artificial intelligence with expressions similar to the following: ‘[Atresmedia] will be transparent in the use of AI tools and will inform its users of all the necessary aspects of their operation so that they have a sufficient understanding of their technological capabilities and expectations of their results’ (Atresmedia). Only media outlets such as NPR create further limitations in the name of transparency, limiting this labeling when the use of AI plays a ‘significant role’ in the creation of the journalistic piece. Nordic AI Journalism references the importance of warning readers about the use of AI in published content, although it does not note any need to inform the public about AI tools used internally. With the exception of this case, it should be noted that the importance that the media attach to transparency is consistent with the publication of the guidelines and measures in the form of press releases aimed at the audience, with the intention of ‘manifesting’ the principles and values regarding the use of AI that are self-imposed internally in the generation of content and the use of technological tools of this type in an exercise of explicit commitment, giving value to trust.

The justification for these transparency measures is often linked to respect for copyright (29) and original human content: ‘As with any editorial content, we want to protect ourselves against any risk of error, copyright infringement and also support the work of artists, photographers and illustrators’ (Les Echos-Le Parisien).

That same justification also highlights a commitment to content verification (34), focusing on internal reminders to journalists to constantly monitor and check AI-generated material, as it may be biased or contain ‘hallucinations’ (errors in which the AI produces incorrect or unfounded information); as well as security and privacy risks, alluding to special caution when dealing with sensitive or personal data (28) and including it in their prompts and AI queries.

4.5. Relevance of Human Supervision (C5)

The importance of human supervision in journalistic work with AI is reemphasized in 44 of the 45 manuals analyzed, except for News Media Alliance (USA) and RVDJ (Belgium), which make no references to the subject. The general discourse alludes to the use of AI as a complement to human work, never as a substitute or replacement: “The newsroom may use AI to facilitate or improve its work, but human intelligence remains at the core of all our editorial production. No content will be published without prior human oversight” (Heidi News); ‘If we wish to include significant elements generated by AI in a piece of work, we will only do so with clear evidence of a specific benefit, human oversight, and the explicit permission of a senior editor’ (The Guardian).

Professionals are seen as bringing critical judgment and cultural context that AI cannot replicate, ensuring journalistic quality and respect for ethics that technology alone would not be able to achieve: “Automation of routine tasks to AI means that journalists must monitor and intervene to ensure that stories are treated appropriately” (Al Jazeera).

Media outlets such as DJV connect human oversight and transparency in the use of AI to the quality of the final product, credibility and audience trust: “Credibility is and remains the most important prerequisite for journalism, and can only be maintained if people are responsible for selection and distribution” (DJV.).

In this line, media outlets such as the Portuguese A Mensagem de Lisboa state their commitment “to identify and mitigate biases in all AI applications. Special attention will be given to ensuring that the voices of marginalized communities are not misrepresented or overlooked”. These values are also shared by the Spanish RTVE group, with its express commitment that “AI systems should be fairly designed, avoiding unfair and discretionary biases”, and they should be accessible and subjected to systems of accountability, governance and continuous evaluation to monitor compliance with the specified standards.

5. Discussion and Conclusions

This research on the regulation and self-regulation of AI across various international media organizations, groups and entities (O1) has compiled a first catalog of 45 guidelines, stylebooks and regulations in place in 18 countries and regions (o1). The analysis of this catalog confirms the starting assumption of this research about the existence of a recent and growing trend of media self-regulation in the use of AI. While awaiting more detailed regulation by the different governments, the journalistic sector itself or professional associations, it is the media that respond individually to the challenges of AI with an express commitment in line with their professional and ethical positioning. Their proposals, mainly based on verification, accountability and transparency, are aligned with the general recommendations of supranational organizations (Unesco, OECD and the Council of Europe). And, by publishing them in open digital formats, they reflect an intention to bring reliability to the audience in the ’self-policing’ journalistic use of AI.

The study confirms the starting assumption that the media feel challenged about the ethical risks of AI and have recently begun to publish recommendations and codes that reflect two purposes: on the one hand, to establish internal indications to their editorial staff to delimit the use of generative AI to journalists; and, on the other hand, to publish these indications openly to express their ethical commitments to transparency, also seeking the trust of the audience.

This study on the internal allowances for AI usage and its limitations (o2) yields two main conclusions. Firstly, there is a greater concern on the part of the media about journalistic use of generative AI—that which creates ‘new’ textual and multimedia content (Feuerriegel et al., 2023)—than for analytical AI—that which analyses large amounts of data and detects patterns This result can be explained as a possible reaction to the democratization of the use of this type of AI with the announcement of Chat-GPT, a generative AI tool made available to the general public (Pavlik, 2023) at the end of 2022, and subsequent similar technological developments.

Secondly, we are able to confirm that, although the generation of multimodal content is limited in most cases, this limitation is established differently across the self-regulatory documents analyzed, from specific guidelines and use cases—such as using AI to suggest headlines, but not to write information (The Globe and Mail)—to generic recommendations that allude to the supervision and prior approval of content (CBC, CDMSI, Groupe Les, El País, Echos Le Parisien). Only one internal regulation with a legal framework has been identified (RTVE), which serves to exemplify the primitive and still developing state of media regulation in the face of AI in the context of rapid technological evolution and an accelerated need to respond with internal guidelines that require constant adaptations.

The analysis also allows us to detect the disparity in how media outlets and organizations express their ethical stances regarding transparency and the importance of human supervision in the application of AI (o3). There is consensus in the discourse: a majority notes that AI should not replace the journalist but act as an auxiliary tool (Canavilhas & Giacomelli, 2023) and as a support in specific tasks. In addition to their active participation in the process, the professional adds the supervision, verification and identification of the content that is generated, ensuring accuracy, reliability and mitigation of algorithmic biases. The ethical use of AI is linked to its relationship with the audience since most of the regulations declare the importance of publicly identifying any content generated with AI, linking this disclosure to upholding the quality of their information and the trust of their readers.

Overall, the reviewed guidelines reflect a concern on the part of the media with a technical perspective—by delimiting what content can be created—and an ethical one—by determining under what principles this content is created. However, the scarce mentions of analytical AI, the interest focused on resources such as the generation of text or images, and the uncertain generic recommendations found in some documents (Porlezza, 2023) indicate a still developing state for both the different uses of AI and the commitments to self-regulation. At the same time, it is opening a path of study and reflection for researchers and journalists in a continuous process of necessary internal updating. There is greater media concern about the internal limitation of the use of generative AI and a strong internal ethical positioning in favor of transparency, content verification and the correct treatment of data and copyright. Self-regulation in the use of AI is mostly directed towards a triple objective: avoiding bias, preserving informative quality and audience trust. These axes can serve as a reference for media and entities in future discussions about the development of specific regulations for the sector.

The geographic diversity of the materials and the variety in the size of media entities reflect that AI self-regulation in the media is an international phenomenon affecting both large and small entities worldwide.

This research has limitations that should not be overlooked: firstly, the difficulty of locating a suitable sample for the object of study, as it is dispersed, difficult to locate and cannot be considered representative in its entirety. The use of the snowball technique is the most appropriate for this type of exploratory study in order to overcome the absence of previous studies. However, at the same time, it implies a certain bias in the search, so it is recommended that future research be completed by expanding the sample, adding internal manuals that are not published openly and complementing with other techniques, such as interviews with media managers and editors regarding the self-regulation of the use of AI or establishing focus groups with experts to establish a guide that can be extrapolated to the sector as a whole.

This research serves as an approach to an issue that will continue to be subject to technological advances and ethical dilemmas and that, therefore, does not require that the media are “waiting and watching” (Haapanen, 2020, p. 5) but that it takes initiative toward the adoption of a framework for action that is applicable to all phases of the journalistic process in which AI already acts, as well as the implications that its use may have on business models, sustainability and the relationship with audiences.