Introduction: Who will remember Ana?

Barcelona, winter 2016. At the height of the rise of the Catalan independence movement, a French correspondent makes her way to the seat of the regional government to take the pulse of the situation. But through a quirk of scheduling, she finds herself in the middle of a gathering of families of children stolen under Franco. There she meets Ana, still searching for the daughter taken from her in a maternity ward in the 1970s. Struck by her testimony, the correspondent writes to her editor on the way back to convince him to devote a piece not to the question of independence but to the memorial conflicts surrounding the survival of Francoism in Spain (Guien Reference Guien2016).

This anecdote, which would open for the co-author of this article a long series of reports on memorial questions, is in no way exceptional. Many far more decisive articles, investigations, and scoops have been born of chance encounters, of third-party errors, or of ‘accidents’ on the ground. From the Watergate break-in to the inadvertent addition of The Atlantic’s editor-in-chief to a Signal thread laying out the entire plan for an American strike on Yemen. In each case, the story exists because a human was there, judged that it mattered, and chose to make it part of the record. That human intentionality, owed sometimes to luck, sometimes to intuition, sometimes to doggedness, is precisely what the press is now contracting away.

In March 2026, the press conglomerate News Corp signed a licensing agreement with Meta granting access to the group’s American and British content and archives (eg, Wall Street Journal, The Times, and New York Post) to feed its MetaAI chatbot and train its models (Bruell Reference Bruell2026). Its value is some 50 million dollars per year over 3 years. This deal is not an isolated case: Le Monde in France, Prisa Media in Spain, among others, have recently concluded similar pacts.Footnote ¹ Nor is it the first such agreement News Corp has signed with an AI giant. In 2024, the company signed a 5-year, 250 million dollar contract with OpenAI,Footnote ² which its CEO Robert Thomson then described as a ‘historic agreement’ that ‘will set new standards for veracity, for virtue and for value in the digital age’.Footnote ³ On the most recent deal with Zuckerberg’s company, Thomson was less lyrical: ‘We’re essentially an input company’, he declared. And he made the point explicit: ‘We’re an input in the way that semiconductors are an input, in the way that datacentres are an input, in the way that energy is an input.’Footnote ⁴

This redefinition of the role of press organisations, occurring against a backdrop of acute industry crisis, deserves close attention, for it erases an entire dimension of journalistic activity, precisely the one our opening scene makes visible: the intentionality that turns an accident into a record.

What Thomson does not say is what a semiconductor lacks, and a journalistic article possesses: mnemonic intentionality. A semiconductor carries a signal. A press article carries a past. Sometimes incidentally, as the residue of editorial decisions made under deadline pressure, sometimes more consciously, as an explicit act of framing what will be remembered – like the story of Ana, gathered by accident, then deliberately transmitted in our co-author’s report (Guien Reference Guien2016).

What we argue here is at stake is the ‘third way of memory’ (Hoskins Reference Hoskins2019, Reference Hoskins2024). This is to acknowledge that memory is something that has always been situated and produced between human cognition ‘in the head’ and the mechanisms, technologies, and media of each era, external to the self, namely ‘in the wild’ (Barnier and Hoskins Reference Barnier and Hoskins2018). There has always been some form of human will and provenance embedded in this idea of the production of the past, even though journalists have not readily recognised their agency in this (Zelizer Reference Zelizer2008).

In what follows, we offer an interdisciplinary contribution at the crossroads of our fields of journalism and memory studies. We extend the third way of memory frame onto a still under-explored terrain, the mnemonic dispossession of professional journalism in the age of LLMs. We ask, what do we lose, individually and collectively, when the human work of journalism gives way to the machinic memory feed? What will the past look like and be usable for when it is no longer enriched by journalistic agency, but rather is a product of a corpus scraped, synthesised, and returned by post-human systems?

We develop our argument in five parts. The first establishes the journalist as an agent of memory, crossing media and memory studies with a historical reading of the successive regimes through which this function has been exercised. The second retraces the trajectory leading from post-journalism to the input economy evoked by Thomson, and marks the qualitative break ushered in by generative AI. The third describes the post-human memoryscape that results. The fourth maps the resistances now underway and proposes to read these as forms of institutional glitches. The fifth, finally, sketches a response to the mnemonic dispossession this commentary diagnoses: that of an archive journalism, carried by the figure of the journalist as memorial self.

The journalist as an agent of memory

To grasp what is at stake, we need to recall the trajectory of journalism’s memorial function. In the era of mass broadcasting, the journalist operated in the service of a collective public whose experience, opinion, and shared memory were shaped by media institutions. Nonetheless, as Zelizer (Reference Zelizer2008, Reference Zelizer2022) argues, the timebound enterprise of journalism tended to resist recognition of the industry’s memorial functions and contributions, particularly the idea of the individual journalist as an ‘agent of memory’. This included journalists themselves, despite being a key agent of mnemonic work, who rarely foregrounded this function in their production of a shared past. In this way, memory and journalism ‘resemble two distant cousins’ […], they ‘proceed apace as autonomous phenomena without seeming to depend on the other. And yet neither reaches optimum functioning without the other occupying a backdrop’ (Zelizer Reference Zelizer2008: 79).

Journalism needs mnemonic work to situate its narrative; collective memory, in turn, needs journalism to have one of its most public drafts of the past at its disposal. This interdependence is founded on journalism’s ontological relationship to time (Zelizer Reference Zelizer2022). Time is not a tool for the journalist, it is the very condition of journalism’s existence. And contrary to what the professional rhetoric of immediacy suggests, this time is never linear. Journalism continually weaves several temporalities together; it goes back, anticipates, compares, and makes sense between disjointed strata of events. Anniversaries, obituaries, media templates, and other types of historical analogy all contribute to a significant register of journalistic practice embedded in its relation to the past.

Perhaps it is this relatively implicit (Erll Reference Erll2022) rather than explicit journalistic memorial consciousness that renders it so fragile, so much at risk from the mnemonic concession enacted in the licensing contracts signed with AI giants. An industry cannot readily defend even its own assets if it has remained largely unacknowledged by its own people.

The paradox revealed by Robert Thomson’s declaration is, first of all, a paradox internal to the profession. It is at this level that the journalistic act meets the act of memory properly speaking. As Zelizer (Reference Zelizer2008: 83) puts it, the moment a journalist decides which subject to cover, in which medium and with which tools of transmission, ‘they find themselves squarely in the realm of memory’s work’. Carolyn Kitch (Reference Kitch2002) has also shown how certain formats – commemorative editorials, special issues, and retrospectives – function as fully assumed mnemonic rituals.

What is at stake reaches beyond the past alone. There is a body of important work on ‘collective future thinking’ that highlights the interdependencies of imagining the future and the (re)construction of the past (Szpunar and Szpunar Reference Szpunar and Szpunar2015). For example, Conway et al. (Reference Conway, Loveday and Cole2016) identify a ‘remembering-imagining system’ as the deeply entangled human capacity to reconstruct the past and to project the future. In journalism, the past is similarly seen as a reference for reading the events to come (Tenenboim-Weinblatt Reference Tenenboim-Weinblatt2013). News media (images, templates, and narratives) are used to ‘premediate’ (Grusin Reference Grusin2004) what is to come, to help make the future plausible and thus more graspable, manageable, and controllable. Premediation, in essence, protects individuals and societies from future shocks through the development of a greater preparedness for what might come based on some kind of persistence of previous experience.

The anecdote that opens this commentary illustrates this on a small scale. Choosing to devote a report not to the planned independence demonstration, but to a gathering of Franco-era victims that fortuitously took its place, is not only a journalistic arbitration. Rather, it is a mnemonic act that brings into existence a memory that had been marginalised and configures the grid through which the readership will read the news that follows. This individual gesture, however, is also inscribed within historical regimes that determine its scope.

Some may counter and say that a journalist, as a kind of individual agent of memory, has always been more of an ideal, always subject to the overbearing machinery of the media of the day. For instance, even the mid-twentieth century’s ‘high modernism’ of American journalism (Hallin Reference Hallin1994, 6) was an era in which the memorial function of journalism was exercised in a largely unilateral way. The mass media of the twentieth century produced nothing less than reality itself, in a one-way relationship from medium to public – centrifugal, mythic, and largely uncontested (Merrin and Hoskins Reference Merrin and Hoskins2025).

The twenty-first-century digital turn redistributed this power, partially and unevenly, towards individuals as participants and co-producers of memory, rather than only as spectators, in fragmented news streams and personalised but also algorithmically governed flows. Yet in both mass (or ‘mainstream’) and digital media regimes, the human intentionality of the journalist remained the pivot. Is there not something human in the recognition of the individual journalist as carrying trust, authenticity, reputation, and ultimately authority over their memorial trajectory of making the past present? The memorial function of journalism is an embodied practice, carried by singular human agents whose intentionality – made of attention, intuition, obstinacy, sometimes accident – constitutes the very condition of possibility of the mnemonic operation.

However, generative AI reassigns memorial value from its human entanglements, from the journalist as memory maker (no matter the state of the news business), from once helping make sense of the present and guiding the future, to the input economy’s not so much remaking of journalistic memory, but rather ingesting and reinventing it.

From post-journalism to the input economy

The shift symptomatised by Thomson’s declaration did not begin with generative AI. It has a prehistory, which the concept of ‘postjournalism’ (Mir Reference Mir2020) allows us to reconstruct.

According to Andrey Mir, journalism has undergone a structural reversal that he reads as the onset of a ‘historical extinction’ of the profession. Faced with the collapse of advertising revenue and the need to court paying digital subscribers, mainstream newsrooms inverted their function: ‘The media […] switched from news supply to news validation. This is how, sometime between 2016 and 2018, the reversal to post-journalism was completed’ (Mir Reference Mir2026; see Mir Reference Mir2020 for the initial development of the concept). This shift had a precursor. After the early years of the social web, when alternative content had briefly contested the centrality of major newsrooms, the mainstream media had reasserted control: ‘The mainstream has enveloped the extreme. It has regained and renewed its powers of gatekeeping, of verification, of defining agendas’ (Hoskins and O’Loughlin Reference Hoskins and O’Loughlin2015: 1321). The media war had been arrested through absorption rather than disappearance. But this reassertion came at a symbolic cost: the value of the mainstream now lay less in what it produced but rather more in what it branded and legitimised.

Mir’s post-journalism is the logical continuation of this turn. What initially appeared as an ecological adaptation became a business model, the mainstream had learned to verify rather than produce, and it discovered it could make this directly its source of revenue. Generative AI does not merely accelerate this process: it represents something qualitatively new. Mir, Hoskins, and O’Loughlin draw attention to the displacement of the mnemonic function of the journalist in the shift from the production of news to its validation. However, today, generative AI extracts a different kind of value from the news corpus itself.

When LLMs are trained on journalistic archives, they absorb the text but dissolve its intentionality. The framing decisions, the editorial choices, the acts of contextualisation that made the production of news content a mnemonic act – are obscured in its ingestion. What remains is pattern, not agency. The result is a memory synthesised by a machine from a corpus whose human intentionality has been evacuated, generating a past that never actually existed. The historical extinction diagnosed by Mir then changes register; it no longer designates only the end of an economic model, but the erasure of a mnemonic function.

Shuwei Fang has proposed a structural name for this process, namely the B2A2C model (Business to Agent to Consumer), in which content is produced to be consumed and synthesised by AI agents before reaching human readers. In this model, journalistic production becomes what Fang calls infrastructure food. Value no longer resides in the article but in the pipeline that synthesises it. Information itself becomes ‘liquid’ – constantly reformed depending on who queries the model, and with what intent. AI systems acquire a functional agency: ‘the question is no longer who controls what we read, but who controls what we think to ask. This is the intention economy: where the most valuable real estate is not human attention but human curiosity itself’ (Fang Reference Fang2025).

With his ‘input company’ formulation, Thomson describes the completion of the B2A2C transition: journalism as infrastructure food, its memorial function evacuated by the system that absorbs it. In this model, journalists become the involuntary supplier of the mnemonic health of a system that dispossesses them. Thus, they are ensnared in a paradox, at once victims of the system and the condition of its survival.

The post-human memoryscape

But let us return to our opening scene. A journalist, having stumbled by chance on a gathering of families of stolen babies, collects Ana’s testimony. She turns it into a report on Francoism’s persistence in the Spain of 2016. Now imagine the same subject 10 years later, reported not by a journalist in a press article, but by a chatbot. ‘What do we know about the stolen babies of Francoism?’ The answer would probably be coherent in its broad outlines – the dictatorship, the role of the Church, the clinics, the absence of any reparations policy… But there is little chance Ana would feature in it. The testimony, the clinic, the daughter whose trace was lost: all this still exists, archived in an article – but who will still think to go and look for it? It is in this displacement – from archive to synthesis – that the post-human memoryscape takes hold.

In the B2A2C world, memory ceases to be constituted by a relatively stable body of artefacts – articles, archives, and works – produced by human agents anchored in a present with specific intentions. The past becomes liquid as information has become liquid: constantly reformed, intention-specific, with no fixed form to which one can return and that one can examine critically. Each query reconstructs an ad hoc past, stabilised for the duration of an answer, then dissolved.

However, recent advances in Generative AI usher in new memorial operations, if not a stability of sorts, in its capacity to take decisions as to what to ‘modify, forget or save’ and in adapting to changing environments, such as xAI’s Grok 3 (Hoskins Reference Hoskins2026: 8). This is machinic agency in its production of post-human memory.

The problem, then, is not the destruction of the archive. The article on Ana still exists, indexed somewhere. Newspaper archives are, for the most part, still accessible; journalistic records have not disappeared. What shifts is the mnemonic point of access, the first contact with the past. The LLM synthesis becomes that first interface, relegating the original journalistic artefact to the rank of second-order source or deeper still – consulted only if one already knows what to look for and/or if one is willing to cross the access barriers: paywalls, the obligation to leave the interface, the detour through a search engine. The journalistic archive is not quite erased as such, but rather functionally neutralised.

This regime adds to the proliferation of ‘grey memory’ (Hoskins and Halstead Reference Hoskins and Halstead2021) namely, ‘how contemporary technologies push out of individual human reach a conscious, active, willed memory, through obscuring the risks of the ownership, use, access, costs, and finitude of digital data’ (Hoskins Reference Hoskins2024: 13). The user retains the sensation of agency – the interface is fluid, the answer arrives – while the actual conditions under which that past is constituted escape them. Memory is available in appearance, opaque in its conditions of production. This impoverishment is layered with further risks. Post-human memory is not just a matter of machinic replacing or displacing human memory, but rather it is the result of the transformation of an entire ecology of remembering and forgetting. For instance, LLM models trained on their own outputs end up ‘forgetting improbable events’ – a model collapse (Shumailov et al. Reference Shumailov, Shumaylov, Zhao, Gal, Papernot and Anderson2023: 2) that threatens precisely the singular cases, exactly that which Ana embodies in the opening scene.

This all begs the question: who owns the archive? If the intentional human record becomes, to borrow Shumailov et al.’s (Reference Shumailov, Shumaylov, Zhao, Gal, Papernot and Anderson2023) formulation, ‘increasingly valuable in the presence of content generated by LLMs’, the question of its access, preservation, and control ceases to be technical and becomes political. Journalistic archives may today be held by press groups whose business model, as we have seen, encourages them to sign them away by contract. Mnemonic sovereignty passes into the hands of those very actors who monetise it.

This dispossession does not go unanswered. Within this very post-human memoryscape, frictions are organising. It is these we now turn to examine.

Journalism as a necessary ‘glitch’

While major groups, such as News Corp, Le Monde, or Prisa Media, negotiate licensing deals worth tens of millions of dollars, independent outlets are caught in an asymmetric trap. They are too small to negotiate a contract that carries weight against the AI giants, too visible for their corpus of articles to escape the scraping of their archives, without compensation. At the same time, their ecosystem is saturated by press slop, that flow of AI-generated synthetic content that drowns out the visibility of human journalistic work. For instance, over 14,000 French sites catalogued in 2026 by the investigative journalist Jean-Marc Manach (Reference Manach2026) revealed how output multiplied tenfold in a single year. Two forms of asymmetric aggression combine: scraping upstream, saturation downstream.

Yet resistance does not come only from independents. Some major groups, too, refuse the input logic. In 2023, the New York Times sued OpenAI for using, without authorisation, millions of the paper’s articles to train its models and produce answers reproducing or substituting themselves for its journalistic work. Widely read as a simple copyright dispute, this lawsuit in fact carries stakes that go beyond intellectual property law. By refusing to have its articles treated as raw material – as input, in Thomson’s sense – the Times formulates, by legal means, a claim of mnemonic sovereignty. Copyright here becomes an available means to assert a stake that does not yet have its own categories, that of control over what will become of the past these articles carry.

The second form of resistance is infrastructural: a growing number of publishers choose deliberate exit from the algorithmic pipeline. The outlet Newpress builds its model on direct, community-driven relationships between journalists and readers, short-circuiting both recommendation systems and AI synthesis. The French ecological fanzine Climax,Footnote ⁵ distinctive for its events organised around newsstands repurposed as social spaces, follows a kindred logic, as does a broader return to print observable in certain segments of the independent press. In this context, the return to paper is not necessarily nostalgia for an earlier medium; it is rather part of a structural refusal to become infrastructure food. By withdrawing – or controlling what they release of their production to the indexable web – these outlets significantly complicate the scraping that feeds the models.

In the Synthetic Memories project by the Domestic Data Streamers studio in Barcelona, images generated from survivors’ descriptions signal their own syntheticity through their imperfections – blur, imprecise features, and visible grain – obtained through the deliberate choice of less performant AI models (DALL-E 2, Stable Diffusion from 2022).Footnote ⁶ The concept of glitch memories names precisely this gesture, these are productions ‘entangling the human and the technology in new relations to produce a negotiated and humanly recognisable vision of the past’ (Hoskins Reference Hoskins2024: 16). Unlike the smooth AI images that erase their conditions of production, glitch memories render visible the hybrid human and machinic conditions of their making.

We can extend this principle to journalistic practices that, within the B2A2C flow, keep visible the human conditions of mnemonic production. The Times lawsuit, the infrastructural exit of Newpress, the return to print, and the events organised by Climax contribute many institutional glitches that, through their very friction, signal that an intentional human act is taking place – that someone frames, chooses, publishes, and takes a stand. From this perspective, the journalist who refuses the input framing functions as a necessary glitch, not as the nostalgic defence of a declining institution, but as a friction that preserves human agency in the making of the past.

What is at stake, ultimately, is the third way of memory. The B2A2C transition throws the in the head/in the wild model (above) of remembering and forgetting off balance. It ushers in a new regime of the ‘memory feed’ (Hoskins et al., Reference Hoskins, Čimová and Pilkingtonforthcoming), a flow in which the memory of the self is no longer something firmly cognitively oriented or understood, but hostage to the machine and its inputs, pipelines, and outputs. Whereas the human and the non-human were seen at least in some kind of consort in contemporary visions of the workings of ‘information infrastructures’ (Bowker and Star Reference Bowker and Star2000), today AI warps the entire ecology. What Fang (Reference Fang2025) describes for her part as a ‘great inversion’ – the tipping from an information ecosystem shaped by human curiosity into an ecosystem that in turn shapes the curiosities it wants us to formulate – is therefore also, in the terms of this commentary, a mnemonic great inversion.

If the post-human memoryscape described in the preceding section remains contestable, it is because some journalists – through lawsuits, infrastructural exits, and institutional glitches – insist on remaining agents rather than becoming absorbed by today’s AI-dominated infrastructure. If the third way of memory is to keep its human balance, it needs them – not as inputs, but as agents. This is the subject of our final part.

The memorial self of journalists

In a recent interview with the independent environmental magazine Vert, the French investigative journalist Fabrice Arfi (Mediapart) speaks of his interest in what he calls ‘archive journalism’ – a form of investigation that excavates the past not to preserve it but to shed light on the present.Footnote ⁷ In some ways, this is an established function of human memory, as we set out above, the past offers individuals, groups, and societies guidance in the present based upon prior experience, and a preparedness in terms of anticipating the future. Is this not the same for journalists, too? We could ask what the function of an individual journalist’s remembering-imagining system is, namely, how their own deeply entangled human capacity to reconstruct the past and to project the future gives unique value to their production of what we call ‘news’. Embedded in this system are, of course, issues of deontology, trust, recognition, and reputation, but often these are not couched as memorial functions.

Put differently, we can say that each journalist has a memorial self, and this might be personally, if not professionally, increasingly implicated in and responsible for a past produced by a life of digital entanglements. This seems a good starting point to leverage some kind of resistance to the B2A2C logic of AI systems, anticipating the curiosities it wants us to formulate.

When Arfi, the Mediapart journalist, revealed in 2012 that the French Budget Minister Jérôme Cahuzac was concealing a Swiss bank account, the trail came from his own journalistic past.Footnote ⁸ Two years earlier, in an unrelated investigation, a small detail had caught his attention, an unexplained leniency shown by Cahuzac – then chair of the parliamentary finance committee – towards a minister of the opposing camp implicated in precisely the same kind of offence. That fact, kept in the journalist’s notebooks and in his head, would in due course open the investigation that brought the minister down.

In 2017, a second variant of this memorial self was at work when Jodi Kantor began her New York Times investigation into Harvey Weinstein. Her first move was to phone her colleague Megan Twohey, who, the previous year, had taken down the harassment allegations levelled at Donald Trump during his presidential campaign.Footnote ⁹ What Kantor came looking for was not a source but a way in, the approach Twohey had patiently honed with Trump’s accusers. The one that lived only in her practitioner’s memory. It worked here, too, and helped tip what became the #MeToo era.

Such gestures of the journalistic memorial self, however far-reaching in their consequences, remain individual. Yet imagine what such a journalistic memorial self could become, sustained by a wider memorial ecosystem enboldened rather than extracted by the archives of today. To counter the memorial dispossession of journalism, individual persistence is not sufficient. Rather, there is a need for a broader alliance with other guardians of memory – archivists, historians, heritage institutions, and oral history projects – who face, from different positions, the same scraping and extraction of their archives and the same algorithmic pressures. What is needed now is a shared memorial front. A coalition that refuses the impoverishment of the memoryscape by insisting, collectively, on the irreducible value of the dated, situated, contestable human trace, a past with a kind of living provenance.

Overall, journalists have long played a vital role in sustaining the balance between the head and the wild production of the past, a third way of memory. Yet they are increasingly mnemonically dispossessed as they are caught up in producing content which is sold and used to feed AI models and systems. This endangers the value of the work of journalists as human agents of memory in creating a past that they have a stake in. Instead, the value of content in feeding AI models and systems is being rendered to help anticipate or predict what users will want to look for. This is the creation of a prompted past that never existed before, rather than that experienced, reported on, and interpreted through human senses.

Archive journalism might then make it possible to invert the framing Thomson posed at the opening of this commentary. Where the CEO of News Corp ranked journalism among the inputs – alongside semiconductors, datacentres, and energy – the journalistic memorial self opens an entirely different regime. The archive is not infrastructure food. It is present in both senses of the word – a human present from which the past can be questioned, and a gift whose value grows as the memoryscape impoverishes around it. A wealth that the B2A2C logic cannot manufacture and that it has every interest in pre-empting. But a present has value only for those who know how to recognise and engage it. There lies the wager, and the responsibility, of the journalistic memorial self: to keep alive and open, against machinic extraction and synthesis, a space in which the past can continue to be illuminated yet also interrogated for human-designated ends. The battle for journalism’s soul, then, is not quite yet lost to AI’s seizure of human agency in the making of memory.