This paper investigates the adoption and governance of Generative AI (GenAI) in the Italian journalistic field, examining its implications for democracy and the public arena. Drawing on 13 in-depth interviews and one collective interview with journalists, union representatives, and ethics board members, the study explores how GenAI is perceived, used, and contested in Italian newsrooms. Findings show a widespread yet unstructured use of GenAI, characterized by both instrumental and critical attitudes. While GenAI is generally framed as a supporting tool for editorial work, concerns emerge around job displacement, editorial independence, opacity, and the commodification of journalistic content. By applying the framework of sociotechnical imaginaries, the study identifies four dominant risk narratives emerge -substitution, mistrusted information, machine-driven editorial logic, and surveillance-reflecting deeper anxieties about power asymmetries in the information ecosystem. The Italian case highlights the need for journalist-led governance strategies and contextualized AI adoption models, as shown by Il Manifesto. Overall, the paper argues that integrating LLMs in journalism is not only a technical matter but a profoundly political issue, demanding participatory and ethically grounded regulation to protect democratic values.
Journalism plays a key role in Western democracies and is undergoing rapid changes due to technological
advancements, profoundly transforming both journalistic processes and products, thus necessitating
innovation in traditional business models[
The potential consequences, particularly for journalism, are considered both positive and negative.
The promise of artificial intelligence (AI), and especially Generative AI, lies in its ability to rationalize
work processes, aligning with ideals of eficiency and speed, but its adoption is constrained by factors
such as professional norms, regulatory frameworks, audience preferences, and existing technological
infrastructures [
In 2023, significant events highlighted the growing impact of these tools on journalism. A notable
example is the historic agreement between Springer and OpenAI, whereby content produced by major
outlets such as Bild, Die Welt, Politico, and Business Insider would be used to train AI models. Also some
media organizations, such as the Associated Press, have signed agreements with OpenAI, recognizing
the potential of Gen AI to enhance content production and distribution, while others, e.g, The New York
Times or Getty [
In this regard, the Italian government is beginning to transpose the European AI Act regulations,
especially from a editorial point of view, focusing on a standardised certification (a marking) for
the traceability of AI-generated content, the valorisation of copyright, the defence of employment
profiles and the profession, the issue of vigilance for competition, and attention to Retrieval-augment
generation (RAG) techniques for indicating the sources of the outputs of generative models, whose
training content must be kept in a special register[
This scenario underscores the divide between diferent views on the role AI systems - notably
Generative tools such as LLMs - should play in the future of journalism and what consequences this
could have for our democratic societies. In fact, it has been recognized that AI based systems are
afecting the public arena, understood as the ensemble of ”media infrastructures” which organizes
the information production, distribution and access while mediating also relations between diferent
actors[
This study aims to advance the understanding of the adoption of generative AI in newsrooms and its
potential critical consequences for democracy, conceived in the light of the notion of communication[
First, the study investigates how journalists are efectively using AI and Generative AI in Italian
newsrooms through in-depth interviews with professionals who engage with AI in their routine
workflow or due to ethical and professional concerns (some interviewees are also members of labor
unions and the deontological order). Then, the study examines how professionals assess the risks
and broader implications of Generative models to promote their responsible integration in a complex
negotiation of their autonomy and agency. Risks and concerns are analyzed through the lens of
sociotechnical imaginaries, an emerging sociological framework that captures diverse perspectives
on AI narratives and discourses [
The adoption of artificial intelligence (AI) in journalism has sparked widespread discussion over recent
years regarding its implications for both the profession and society. AI, particularly in its earlier forms,
has been deployed across various domains, from news gathering to content production and data analysis,
from identifying relevant news stories to personalizing content distribution via recommendation systems
[
The impacts of these technologies on media and journalism indicate broader implications for
democratic societies, afecting fundamental values of the journalistic profession as outlined in articles 3, 8,
and 9 of the Global Charter of Journalism, which respectively address the reliability of journalistic
sources, respect for privacy and human dignity, and the duty to disseminate information without
promoting hatred or discrimination [32]. Moreover, by afecting the information ecosystem [ 33], and
by consequence the public arena, envisaged as the “media infrastructures that enable and constrain the
publication, distribution, reception, and contestation of information that allow people to exercise their
rights and duties” [34] AI could have a transformative efect on democracies. The study in [
These challenges underscore the importance of identifying mechanisms for properly integrating
technology into journalistic work, while also recognizing contextual and organizational factors that
shape the outcomes of AI adoption [
Thus, to understand the implications of AI adoption in Italy, it is essential to consider the broader
features of the national media system, which aligns with the ”Polarised Pluralist” model described by
Hallin and Mancini[
The study adopts a qualitative research approach drawn from the realm of interpretive sociology, which
emphasizes the subjective meanings that individuals and groups ascribe to their social worlds[44, 45].This
study is informed by the frame of social constructionism in technology, particularly through the work of
Langdon Winner [
Another cornerstone of this research stems from studies on sociotechnical imaginaries, defined by Jasanof and Kim[ 46] as the visions that societies hold about the future of technology - both positive, desirable or negative, dystopian - and how these visions influence policy and governance decisions. Starting from this intuition of the co-construal relationship between techno-scientific activity and the political order, scholars argue that imaginaries are contested, diferentiated and a good part of digital governance appear to occur around technology and its rethorics [47] entering also high-level arenas, as for the negotiation of the AI Act, contributing to the development of policy frames[48].
Imaginaries are thus interactional sense-making activities, in the broader political economic, organisational, situational and technological contexts (...) and can be analysed with a methodology that pays attention to interactions, discourses and technologies in the situations of interest[49].
Sociotechnical imaginaries emerge from narratives and discourses as focused on potential
development of technology and its contextual factors [49], and shape and organize how society interprets
technology [
The study integrates semi-structured interviews [50] to gather rich, context-sensitive data, capturing both individual reflections and collective insights. A total of 17 professionals - 13 through semistructured individual interviews and 4 in collective semi-structured interview were conducted - have been interviewed, with respondents selected on their direct experience with AI technologies, whether through their newsroom practices, ethical concerns, or governance-related roles, as members of labor unions (FNSI, Stampa Romana) or deontological order (Ordine dei Giornalisti). The medium duration of interviews is 45 minutes. The collective interview was conducted with members of the Il Manifesto’s newsroom, the duration was 1 hour, and the intervieewed discussed with the interviewer their peculiar approach to tech development in journalistic practices. The data collected have been analysed according to the thematic analysis framework [51].
Insert Tab 1.1 and Tab.1.2 for sampling of interviewees - see appendix
The interview guide was designed to explore uses in the Italian landscapes of AI and Generative AI in Italian information system; the perspectives of journalists expressed through their sociotechnical imaginaries on positive and negative implications for the profession, governance issues and implications for the public arena, by focusing in particular to the most relevant events in the italian context as described above.
All Italian newsrooms are experimenting with AI and equipping themselves with teams to understand the implications and guide the digital transition. As we will see in more detail later on, the adoption of AI, and in particular of more sophisticated technologies such as LLM, may entail a varying degree of conflict and resistance on the part of newsrooms. Before going into the specifics it is worth contextualising the use of AI in the Italian information ecosystem. This is characterised by a tendency to lag behind the newsrooms of other Western countries such as Germany, Great Britain, France or the United States and a use that is still unstructured and uncodified. As one interviewee states:
Abroad it has been used for a decade or so in all sectors [...] AP started in the financial sector, automating texts about listed companies [...] The BCC in 2016 was using AI to make its journalists speak in other languages they do not know. [...] These are structured interventions built into the organisation’s workflow. [...] There is an important delay in Italy due to the culture of innovation on the one hand, on the other, the editorial ofices are now decimated, with few journalists crushed by routine, they have no space or time to dedicate to research or study these systems. The study of the work-flow, of the problems that are time-consuming [...] is somewhat lacking in this structural and headline-driven application. (N2)
Another interviewed defined the news industry as “short-sighted”, due to the fact that “before being publishers they are building contractors, car manufacturers, fuels traders, or they are banks, insurance companies” (N4). (N6) discloses: “the bottom line is this: the newspaper is still seen as an instrument of power to interface with politics”.
This is coupled with the awareness of an ever-changing information landscape, with the emergence of players producing and distributing content on the market in direct competition to traditional newspapers, which further complicates the scenario for traditional newspapers, already battered by the advent of the web. Moreover, the distribution of articles on the web now has to deal with distribution services of the large platforms, which enhance content according to the audience in ways that are not transparent to journalists, who can only “indirectly guess the criteria by which one of our pieces is shown to the various users”, as a reporter recognizes (N9). This respondent considers this new complexity for traditional newspapers by lamenting how the importance of SEO (Search-Engine-Optimisation) technologies borders on click-baiting, with negative efects for the quality of the content and the expectations of the audience. The overall scenario seems destined to be even more complex, with the emergence of generative search engines, such as Search GpT, which challenges a part of the market that was traditionally occupied by others, and once again innovate the criteria for visibility of content, without providing clear guidance on the matter.
What everyone seems to agree on is the idea of the role of tools such as ChatGpt, Copilot, Claude or Gemini have in the journalistic workflow: it is a helper, but feared, because one doubts its future capabilities. In a scarcely structured and codified framework, with editorial ofices reduced to the bone, there are those who use it to make up for the scarcity of editorial resources, for the “inauspicious work, that of filling the gaps” allowing for cuts of working time (N9); those who use it to gather ideas and materials for a piece, to “analyse databases or extract information from them” but always with “the human in the loop” (N12). Its use is supportive also as a tool of translation or summarization of contents, as well as optimizing SEO, with recognized possibilities in fostering the distribution across diferent countries (N6). The promise of eficiency is described by N2: “the reduction of working time could open possibilities for the qualitative part of the work, such as the writing, the research of sources, analysing, investigating”. But still, the impact is disparate and “sector dependent”: “we who do local reporting are more on the street. I think a greater impact is for those working on foreign afairs, because of language issues and the constant research you have to do” (N9). The names journalists use to characterise LLMs indicate a hierarchy: “a valet” (N4), an “assistant” (N12), or a “colleague” who does not have the same skills (N9). There is also someone who just refuses to employ those systems: “I’m ideological in this, I don’t want to use it, but that’s just how I am. I know my colleagues use it as a facilitator of their work” (N13). Contributing to connote the subordinate relationship of LLMs is the recognition of various limits that, as well as imposed by deontology, always imposes verification and control on the journalist of the output provided by the Generative AI. Emblematic (N7):
I have dealt with the Mafia for some time. I had written an article on Matteo Messina Denaro. Out of curiosity, I gave strings to GPT to see how he would do such an article. It is full of gaps. In specific fields, expectation is never the same as final realisation.
Others complain about the inability to properly work with inverted commas from an interview, or the analysis of themes within a body of text: “you have to review everything it gives you” (N9). In addition, hallucinations, bias and explicitness are considered problematic dimensions both in terms of the reliability of the content produced and the understanding of how it is produced. (N1) makes it clear that there is a risk that the algorithm expresses value judgements that are “totally arbitrary” with the consequence that “even on a seemingly non-dangerous article, considerable damage can be done”. The fallout, having established that for the Order of Journalists it is always the journalist who is responsible, is the risk of lawsuits for carelessly generated titles and subtitles. Overall, in this unstructured context, the use of generative AI afects, even if just as a “assistant”, all the stages of workflow.
Journalists were stimulated to elaborate their visions about the adoption of LLMs and AI in their workflow by focusing especially on critical aspects, risks, and concerns. Four diferent imaginaries emerged about the potential broader efects of this technology into the information ecosystem and society.
LLMs: augmenting or substituting?
If, on one hand everyone seemed to agree on the auxiliary, subordinate role of Generative AI, which feeds the idea of irreplaceability of the journalist, mainly due to the limits of the state of the art of the current technology, on the other hand, the fear of loss of control of the tool, of its ungovernability and the impossibility of understanding it, caused by its power, emerges strongly. As (N4) argues, “news gathering will continue to be done by the reporter, the correspondent”. “[LLMs] don’t go and talk to the policeman, the prosecutor, the deputies” (N9) adds. As we have seen in the previous sub-section, LLMs are seen as helpers, within a hierarchical perspective which poses the humans at the top. This idea of the uniqueness of the human being in news gathering is reinforced by (N7), vice-director of a national news agency, which states that the automation of the information flow from the internal desk to the external editorial ofices - so, as a primary source - will never happen, because “it would compromise the transparency of information and its quality upstream”. Nevertheless, job cuts are strongly feared, mostly “for the desk workers, the graphics, the designers” (N9). The reduction of workforce could be facilitated by “a cutting cost approach” (N1). What raises concern is not the tool in itself, but the approach of publishing companies in the economic fragile context they are experiencing (N10), as clearly (N4) considers:
Publishers have in their hands a technology with low cost but huge operational impact [...] and it is a technology that has entered into daily use. The risk that some publishers, even the most emblazoned ones, might want to use AI [instead of journalists] one day is real. [...] This is the industrial revolution of our time and it may produce job cuts. (N4)
This idea of substitution, recognized as a broader phenomenon - “ICT is replacing professions that seemed unassailable” (N3) - is accompanied by the fear of the loss of control of the systems themselves. It recalls the theme of explicability, but declined in relation to the power of the tool: “This type of technology is so powerful and potentially powerful that caution on the part of those who know it, test it and develop it is inevitable, the risk is that it will get out of control” (N8).
Mistrusted Information Ecosystem
The relationship with readers is also a hot topic. “Trust [in journalism] is at stake, but also the reputation of journalists” (N1). Reactions from readers to machine-generated texts, granted it must be declared, is perceived as strongly uncertain: journalists address it as an issue of trust. As (N12) complains, “If I say this piece was 20 percent done by AI, what does it look like? There is uncertainty about readers” reactions to this phenomenon, given how many socio-economic and cultural factors contribute to technological acceptance. Related to this is the risk of fuelling phenomena such as filter bubbles or echo chambers, with the tendency exacerbated by SEO and generative implementations of it “to profile users” and that “unconfirmed prejudices, ideas, beliefs are reinforced” (N1). Of course, the issue of trust deals also with disinformation, with the idea that the job of the journalist increasingly becomes that of intervening in the information ecosystem to cleanse it of falsehood, distinguishing himself by “the ability to certify the trustworthiness of a source and content” (N8).
Information ecosystem driven by machine interests
Related to the centrality of AI systems in dictating news visibility and editorial authority is the gatekeeping function that the holders of digital distributions, such as Google Discover, with a rising role of OpenAI, come to play. As one head of digital at an Italian newspaper made clear, the problem of visibility on these platforms, as well as on systems such as OpenAI’s Chat GpT, following commercial agreements with publishing groups, implies others:
How are those sources put in order? Who ranks first? Who pays the most, as is already the case for some products on Google? Or on the basis of authoritativeness? And how do we define authoritativeness? These are business questions but above all social implications because they concern access to information. (N12).
The access to information is more and more mediated and decided by algorithms and AI systems, with the need of feeding LLMs models with news information by Big Tech companies. As one respondent stated “The Gedi OpenAI agreements are the extreme consequence of following machine’s interests instead of those of readers” (N10). In this scenario journalistic independence is challenged by prioritizing profit over public interest (conceived as the “citizens’ right to be informed” N4), ranking sources based on payments or biased measures of ”authoritativeness,” and an amplification efect of disintermediation.
Your archive, your data are the newsroom’s oil well, and you are handing them over to a third party [...] If GPT provides me with the answers I seek, I will lose interest in subscribing.
This becomes an existential risk for journalism (N2).
Surveillance and control
Confronted by gigantic private enterprises, journalists are aware of the growing asymmetries of power and the risk of loss of independence of publishers, and that of disintermediation. These elements create a scenario of control and surveillance in which democracy is challenged. In this dystopia, information flows are strictly controlled and true pluralism eliminated, creating an environment where dissent is suppressed, and democratic debate is stifled. As an interviewed clearly express: Information without debate, made up of controlled flows, controlled sources of news, does eliminate dissent, true pluralism, any form of true antagonism of constituted power. This is not a political discourse but an institutional one, about the functioning of democracy. This is the real point. The information on which we have been moulded over the last twenty years is a highly controlled information (N3).
But there is a strong awareness on the part of professionals not to fall into the temptation of giving in to easy technological determinism. As one editor-in-chief of a national newspaper well explains: I wouldn’t give technology the ability to change these principles [the democratic ones], it is a tool, it helps, it can be used well or badly but it doesn’t have that kind of function in itself. You have to be aware that misuse can produce damage, and that there can be some malicious people [...]. In the power of technology as a thaumaturgic power, as a power to change democracy I do not believe in it (N8).
Instead, there is a strong awareness that social, economic and cultural dynamics will determine these possible outcomes. In the midst of this complexity, journalists also reflect on how to address these challenges integrating technologies in a responsible way.
In order to face the uncertainty surrounding the integration of AI and GenAI in the information ecosystem, journalists claim a major role in this technological shift and support at national and international level: because, “technological innovation, if not properly governed, exacerbates the problems of journalism” (N4). Therefore there is both the necessity of regulating AI’s learning processes and ensuring fair compensation for the knowledge it absorbs, and of considering how AI-generated content should be controlled and verified. As two important members of national union states, “journalists must be part of strategic development of newsrooms [...] and not having to undergo these choices”(N5). A central issue emerging from the interviews is the economic exploitation of journalistic content by AI systems. As one journalist, member of a regional union, highlights:
AI is feeded on content that is not its own. Inside AI there is our work. If the alienation of value once occurred through hours spent working, today intellectual labor is absorbed through agreements, ending up in an AI that appropriates it forever for uses we do not even understand (N3).
This raises concerns about fair compensation and demands a reevaluation of remuneration mechanisms to ensure that the value extracted from journalistic work benefits newsrooms, not just technology companies. N4 underscores the imbalance by pointing out that AI models not only retrieve journalistic content but also train on it, meaning that “the mechanisms of value must be redesigned”. Copyright is a central battleground in this sense, with N10 noting that ”copyright is currently the only tool available even if it is not ideal” after years and years of “looting” by OpenAI which has left rooms open for privacy concerns alswell.
The issue of editorial control and accountability emerges as another major challenge. The inability to oversee or amend AI-generated content can create legal and ethical problems, which could severely undermine trust in newsrooms. N14 notes that ”what OpenAI does with that data is dificult to understand. Modifying already published information is crucial for newsrooms due to legal implications, AI errors, and factual updates”. This underscores the necessity of mechanisms that allow journalists to track, amend, and verify AI-generated content, with amendability raising as a fundamental value in the puzzle. Furthermore, this urge of transparency regarding how AI systems process, store, and utilize data is considered critical to protect both journalists and audiences.
Thus, even if in Italy newsrooms and institutional bodies have still not published organic guidelines, the preservation of journalistic ethics remains paramount, as N1, president of the deontological order, states: ”The principles of professional ethics remain the same. Journalism has evolved over decades and centuries with these principles, which must now adapt to new production methods and tools”. However, he notes that AI-specific ethical guidelines remain underdeveloped and that tracing the origins of AI-generated content is crucial, as he argues that ”journalistic content must show readers how it was constructed. From there, you touch all sector regulations.” Legal and regulatory frameworks play a fundamental role, and Italy is working with an ad hoc commission and a law in discussion in senate, with the proposal of a standardised certification for the traceability of AI-generated content, the defence of employment profiles and the profession, the issue of vigilance for competition, and techniques to properly understand the outputs of GenAI, whose training content must be kept in a special register [52]. N12 is concerned with the “professionalistic approach” of the proposal, which could make it dificult to adapt it to courts, for example. Others pointed out the dificulty of efectively tracking all contents produced with AI, and once again, the uncertain efects on readers’ trust.
A core strategy for a responsible integration of Generative AI in journalistic workflow is considered investing in research and development. N2 highlights the need for ”experimentation and transparency, identifying problems to be solved in order to integrate AI structurally within news organizations rather than relying on individual initiatives”. He also underscores the necessity of ”new professional competencies to mitigate risks”. N10 also emphasizes the importance of algorithmic accountability, advocating for ”reverse engineering techniques - a mix of computational journalism, coding, and investigative reporting - to uncover algorithmic biases and resist opaque AI systems”.
Training and education are also trivial. In Il Manifesto, as we shall see later on, the editorial board is engaged with technologists to build an in-house model of AI which has the ambition of being an AI open to the community of readers. A sort of an “AI Task” has been formed at La Repubblica, with some journalists attending courses at Oxford, with Reuters and international colleagues, to come back equipped with the necessary knowledge to navigate this technological transition and transfer in-house skills. The need for education is not limited to journalists, but, according to N12, it extends to users because these tools “determine their access to a whole range of services, information, products”. Overall, governance of AI in journalism requires a multidimensional approach that balances technological innovation with editorial autonomy, fair remuneration, and accountability. From legal measures and professional training to ethical codes and transparency mechanisms, structured interventions are necessary. Ultimately, a responsible approach in integrating AI and LLMs in journalism is not just a technical issue but a political one, demanding active engagement from journalists, policymakers, and media organizations.
In September 2024, during the Italian Tech Week, John Elkann, president of the Gedi Group, announced an agreement between Italy’s largest media company and OpenAI. The deal was framed as part of Gedi’s digital transformation strategy, leveraging AI for content translation and improving search results for Italian users of ChatGPT, granting OpenAI access to Gedi’s archives 3. Journalists were informed about this decision hastily on the same day as its public announcement, during a strike organized by the newsroom. As one journalist of La Repubblica recalls, the contents of the agreements have been announced “at a general level without going into specifics” (N13). One an editor-in chief of La Repubblica reconstructs:
Mindful of past experience [...] I am talking about social media, about the changes in the lfow of information... With the advent of AI they wanted to play in advance because there is a reciprocal need on the part of the platforms to educate their intelligence and provide verified content, while on the part of the publishers the need to no longer be passive, to come out of the game so as no longer to have someone scraping your content without your control. 3https://www.milanofinanza.it/news/editoria-accordo-tra-john-elkann-e-sam-altman-sull-ai-i-contenuti-di-gedi-in-italianovanno-su-202409261434075592
Hence the agreements with major publishing companies, agreements about which little is known but which involve integrated services between search and artificial intelligence” (N8).
The agreement with OpenAI is seen as detrimental to their profession and autonomy and it’s considered an expression of a “self-destructive approach” which leaves newsrooms “subordinate regarding Big Tech companies” (N4). Moreover, it intensified existing power asymmetries in the information ecosystem, highlighting the broader labor-capital conflict in AI adoption. The labor union emphasized two main demands: AI should support, not replace, journalistic work, and fair compensation should be ensured for intellectual labor. One union representative cited Article 14 of the national contract, arguing that publishers should compensate journalists when their work is transferred externally: Normally the employed journalist when hired gives up the content of his journalistic, photographic content because it is used by the various platforms of a publisher: crossmediality, which is already provided for in the contract. Where does this end and the content of Article 14 begin? Which requires that a fee be paid for the use? It begins when the content is transferred externally (N4).
Despite these eforts, journalists faced significant challenges in negotiating with entities that possessed ifnancial resources exceeding the GDP of some sovereign states. The EU copyright regulations, the AI Act, and recent decisions from the copyright national authority provided some legal safeguards, but structural issues persisted. A crucial demand is the involvement of journalists in AI-related decisionmaking processes which involves technical improvement in production, as stipulated in Worker’s Statute:
The game is for journalists to be part of the strategic development directorates [...]. One challenge is to understand how to acquire the most useful technologies for journalistic work, which must be part of the process of deciding what is needed, and not having to be subjected to these choice (N5).
How regulate Generative AI, particularly following the agreements between major publishing groups such as RCS and Gedi with OpenAI, has become a central issue in the renegotiation of national contracts for journalists. The contents of these agreements, protected by trade secrecy, have not been yet disclosed in detail to journalists, who are demanding greater transparency and adherence to regulations safeguarding journalistic work and intellectual property rights. A labor dispute is currently underway, involving all national publishers, newsroom representatives, and policymakers. LLMS have intensified the debate on AI Italian politics, underscoring the need to address key issues such as value creation through machine learning, labor protections, control over content and the training processes of foundation models, and ensuring that technology serves the interests of newsrooms.
It is worth focusing on a diferent approach in the Italian context which shows how to deal with these problems in a less conflictual way. Il Manifesto, an editorial social cooperative, pursued a distinctive path by developing its own AI tool, Memoria Manifesta (MeMa), in collaboration with the start-up Isagog. Unlike La Repubblica’s agreement with OpenAI, Il Manifesto prioritized data ownership and transparency. MeMa mainly uses data from its own archive and enhances the newsroom’s ability to process and retrieve archival content through knowledge graphs, semantic searches, and summarization tools. Besides the certitude of data employed by the model, developing a local in-house AI can bring other advantages, such as greater environmental sustainability, due to the solicitation of far fewer parameters when compared to the OpenAI model, more transparency and explicability, and the reduction of the peril of disintermediation. As the technical development managers make clear: If you do a search on 7 October, it returns a series of summaries generated from articles cited as sources. There is a strong element of transparency compared to generic chatbots, which also begin by providing sources, but over these there is no control: you cannot tell GPT which to consider and which to exclude, nor is there any way to control how it takes information from these sources (N17).
Here from MeMa we extract concepts, ontologies, people, places, entities that allow you to do completely diferent consistency checking operations compared to AI systems based on neural networks (N16).
This gain in explainability and transparency, has a consequence for journalistic work and is a solution to the need conceptualized by N14 as amendability, the faculty of correcting an output, whenever it is generated by a human or a machine.
Moreover, part of the project is to extend this process of controlling the knowledge produced by the AI with a wider public, realizing a “community AI” in which members would be able to correct some entities retrieved by MeMa, or validate some specific content generated by it, fostering once again transparency. As they show me in the videoconference:
There is this important role, for example, this is the archivist, who has the power to say this content is good, this correction is good, and then, how you see, the display of the entities changes and shows the reader that it is a content blessed by a human who has interpreted it, and judged it, to be correct, whereas those other entities were only interpreted by Mema.
This is the journalists who taught me (N16).
The engagement of the public or specific members of the community is expected to foster trust in readers, by engaging them in fact-checking procedures while enhancing their critical thinking on the use of AI. Furthermore, the approach of cooperation and development has put journalists and other professions - such as the archivist - in a position not to sufer innovation as an extractive imposition, but as an opportunity to broaden and renew skills.
This study expands existing research on how the Italian media landscape is dealing with AI
systems[43, 41] and notably with Generative AI, by focusing on professional practices and governance
strategies, through the lens of sociotechnical-imaginaries. The analysis, grounded in a set of in-depth
interviews, provides empirical insight into how Italian journalists are responding to and helping shape
the introduction of generative AI within a pluralist-polarised media ecosystem. Unlike other contexts
where AI implementation appears more structured and codified - such as Denmark and the Netherlands
[
In response to the second research question, our findings show that journalists conceptualise LLMs in
instrumental terms, often as “assistants” or “colleagues” with limited capability. However, some critical
aspects are particularly worth noting and necessitate human oversight to verify and control content
before publication - aligning with findings from other studies [
These concerns are compounded by opaque agreements, such as those between OpenAI and major
Italian publishers, which exclude journalists from strategic decision-making processes, fueling internal
tensions and labor disputes, fostering power issues and conflict for information control, as recognized
also in [31]. Furthermore, echoing concerns raised in [
Overall, journalists claim structured interventions and an integrated approach that provides
multilevel support to news organizations in order to truly exploit the potential of AI. Regarding the third
question, journalists articulate sociotechnical imaginaries of AI that highlight existential threats to
the public arena. This is particularly relevant for democracy, understood not only in deliberative
or procedural terms, but above all as a deep communicative process that can also be conflictual, as
suggested by Sustain[53] and Coeckelbergh[
This work is subject to several limitations. First, it focuses exclusively on Italian journalism and does not ofer a comparative perspective. Second, although the interviewees were selected for their significant insights—due to their institutional roles or first-hand experience with the technologies in question—the
Head of digital and AI projects, journalist Member of the Board, journalist AI architect and data scientist, MD
Lead Researcher in language, semantics and knowledge representation sample remains limited. Third, the sample is unbalanced wotards male respondents, a fact which could be expression of the fact that male journalists are more dedicated to innovation - fact who is due to complex cultural and social factors intersecting gender and power in labour - or just due to the limited and non representative nature of the sample. Fourth, the study could be expanded by widening the range of interviewees, including freelance journalists, early-career professionals, and staf from local newsrooms. This would allow for a more comprehensive understanding of the integration of generative systems in journalism, incorporating perspectives from more precarious segments of the profession and from roles that, while editorially relevant, are not strictly journalistic. In addition, integrating more structured methods of data collection—such as questionnaires—could provide a valuable complement to qualitative interviews, enabling a more comprehensive and systematic account of journalists’ perspectives on how technology is impacting their work.
Gender
Age M M M F M M M M M M M M M
During the preparation of this work, the author used X-GPT-4 in order to: Grammar and spelling check. After using these tool(s)/service(s), the author reviewed and edited the content as needed and take full responsibility for the publication’s content.
University Press, 2019. [22] T. F. Waddell, A robot wrote this?, Digital Journalism 6 (2018) 236–255. doi:10.1080/21670811.
2017.1384319. [23] K. N. Dörr, Mapping the field of algorithmic journalism, Digital Journalism 4 (2016) 700–722.
doi:10.1080/21670811.2015.1096748. [24] M. Kotenidis, A. Veglis, Algorithmic journalism: Current applications and future perspectives,
Journalism and Media 2 (2021) 244–257. doi:10.3390/journalmedia2020014. [25] N. Helberger, On the democratic role of news recommenders, Digital Journalism 7 (2019) 993–1012.
doi:10.1080/21670811.2019.1623700. [26] L. Dierickx, C.-G. Linden, A. L. Opdahl, Automated fact-checking to support professional practices: Systematic literature review and meta-analysis, International Journal of Communication 17 (2023) 5170–5190. [27] K. Dörr, L. Hollnbuchner, Ethical challenges of algorithmic journalism, Digital Journalism 5 (2017) 404–419. doi:10.1080/21670811.2016.1167612. [28] A. K. Schapals, C. Porlezza, Assistance or resistance? evaluating the intersection of automated journalism and journalistic role conceptions, Media and Communication 8 (2020) 16–26. doi:10. 17645/mac.v8i3.3054. [29] B. Garcìa-Orosa, J. Canavilhas, J. Vázquez-Herrero, Algorithms and communication: a systematized literature review, Media Education Research Journal (2023). [30] B. Jones, E. Luger, R. Jones, Generative ai and journalism: A rapid risk-based review, 2023. Preprint or internal report. [31] F. M. Simon, Rationalisation of the news: How ai reshapes and retools the gatekeeping processes of news organisations in the united kingdom, united states and germany, New Media & Society 0 (2025). URL: https://doi.org/10.1177/14614448251336423. doi:10.1177/14614448251336423. [32] G. Romeo, E. Griglié, Ai ethics and policies: Why european journalism needs more of both, in:
J. Mokander, M. Ziosi (Eds.), The 2021 Yearbook of the Digital Ethics Lab, Springer, 2022. [33] C. W. Anderson, A. Valeriani, Re-considering journalism as an ecosystem… introduction, Problemi dell’informazione 1 (2023) 3–12. doi:10.1445/106767. [34] A. Jungherr, R. Schroeder, Digital Transformation of the Public Arena, Cambridge University Press, 2022. [35] A. Jungherr, Artificial intelligence and democracy: a conceptual framework, Social Media +
Society (2023). doi:10.1177/20563051231186353. [36] M. Castells, The network society revisited, American Behavioral Scientist 67 (2022). doi:10.1177/ 00027642221092803. [37] S. Papathanassopoulos, I. Giannouli, I. Archontaki, The media in southern europe: Continuities, changes and challenges, in: S. Papathanassopoulos, A. Miconi (Eds.), The Media Systems in Europe: Continuities and Discontinuities, Springer International Publishing, Cham, 2023, pp. 133–162. [38] A. Baroni, G. Rigoni, Afective polarisation: The use of emotional language by italian news outlets on twitter, Mediascapes Journal 23 (2024) 46–66. [39] P. Mancini, The italian public sphere: a case of dramatized polarization, Journal of Modern Italian
Studies 18 (2013) 335–347. [40] E. De Blasio, R. Rega, M. Valente, Polarization and platformization of news in italian journalism: The coverage of migrant worker regularization, in: D. Palau-Sampio, G. López García, L. Iannelli (Eds.), Contemporary Politics, Communication, and the Impact on Democracy, IGI Global, 2022, pp. 74–92. [41] M. F. Murru, S. Carlo, Ai e newsmaking: Un’indagine esplorativa nelle redazioni nazionali e locali italiane, Mediascapes Journal 23 (2024) 184–198. [42] S. Spillare, M. Bonazzi, P. D. Esposti, Ai imaginaries and narratives in the italian public discourse:
The impact of chatgpt, Im@go (2024). [43] P. D. Esposti, L. Tirabassi, The human-algorithmic entanglement in the news realm, Problemi dell’informazione 1 (2024) 41–64. doi:10.1445/113228. [44] A. Schütz, The Phenomenology of the Social World, Northwestern University Press, 1967. [45] M. Weber, Economy and Society: An Outline of Interpretive Sociology, University of California
Press, 1978. [46] S. Jasanof, S. H. Kim, Sociotechnical imaginaries: Insights from the history of science and technology, Science as Culture 22 (2013) 233–256. [47] A. Mager, C. Katzenbach, Future imaginaries in the making and governing of digital technology: Multiple, contested, commodified, New Media & Society 23 (2021) 223–236. doi: 10.1177/ 1461444820929321. [48] R. Corsi, E. d’Albergo, La politica dell’intelligenza artificiale general purpose: immaginari sociotecnici, democrazia e policy frame nel processo decisionale della regolazione europea (“eu ai act” 2022–2024), Im@go 23 (2024) 109–130. [49] C. B. Sanders, J. Chan, Methodological reflections on researching the sociotechnical imaginaries of ai in policing, in: S. Lindgren (Ed.), Handbook of Critical Studies of Artificial Intelligence, 2023, pp. 773–782. [50] S. Kvale, Interviews: An Introduction to Qualitative Research Interviewing, Sage, 1996. [51] V. Braun, V. Clarke, Using thematic analysis in psychology, Qualitative Research in Psychology 3 (2006) 77–101. [52] D. Bianchi, Giornalismo e intelligenza artificiale: Aspetti giuridici e normativi, in: Report Osservatorio Giornalismo Digitale 2024, Ordine dei Giornalisti, 2024. URL: https://www.odg.it/ osservatorio-sul-giornalismo-digitale, accessed January 30, 2025. [53] C. R. Sunstein, #Republic: Divided Democracy in the Age of Social Media, Princeton University Press, Princeton, NJ, 2017.