AI Governance as Regulation of Collective Memory

Collective memory is increasingly mediated through AI tools. Prominent examples include initiatives that focus specifically on mediating past events—such as “virtual witnesses” of the Holocaust¹, the “library of the future”, developed as part of a collaboration between academia and Open AI², or the use of AI for restoring cultural heritage artifacts³. Additional examples are general-purpose large language models (LLMs), which are gradually becoming a major source through which people obtain information about the past.⁴

The concept of collective memory, introduced by sociologists, refers to the joint recollection of the past by societies, nations, or other communities sharing a common identity.⁵ Underlying this notion is an understanding that our memories (in the sense of ‘recollection of the past’) are not confined to our own individual experiences but are also a product of social construction—they are formed, in part, by the groups to which we belong, be they nations, minority groups, religious groups, or other communities. The emphasis in collective memory studies, then, is not on the cognitive dimensions of memory, but rather on the social dimensions and the factors that shape group memory. To quickly illustrate, we may say, in everyday parlance, that we remember the first landing of man on the moon, or the Holocaust, even though we did not personally experience these events, and may not even have been alive when they took place.⁶

Ample interdisciplinary literature from the past decades explored the significance of collective memory for the formation and flourishing of group identity. Collective memory is crucial for narrating the life stories of nations and communities and provides vital building blocks for their shared identities.⁷ It can be particularly important for minority groups, since it allows such groups to “seek a voice and a shared identity that is distinct from and may conflict with that of the nation-state”.⁸ Finally, collective memory does not merely benefit communities. Because our self-identity draws from and is entangled with the identities of the social and cultural groups to which we belong, it is also crucial for the formation of individual identity.⁹

Come AI. The rapid developments in the field of AI and its integration into almost every aspect of human life present new and significant opportunities for collective memory. For example, AI-generated memorials mediate historical narratives to the public in interactive and user-friendly ways.¹⁰ AI-based historical holograms can interact with users in a human-like way, years after the passing of the actual individuals they “represent”,¹¹ and general-purpose LLMs easily summarize vast amounts of historical information, presenting it in a friendly and accessible way.¹² The increasing integration of AI in mediating collective memory has led researchers to predict a future where AI driven systems “will effectively decide what information sources and what interpretations of the collective past gain more visibility, and thus, shape how this past is remembered.”¹³

First, collective memory has a “multifaceted nature”. At the most fundamental layer, it relies on historical events, and the accuracy and veracity of historical information constitutes a crucial building block in collective memory formation. Yet, collective memory is not synonymous with history. Rather, the same historical event can have a certain shared meaning in the collective memory of one group and an entirely different meaning in the collective memory of another community.¹⁴ Collective memory can thus encompass a multiplicity of interpretations and voices, allowing different communities to narrate their own narratives and seek a shared identity that at times may be different from, and even conflict with, that of majority groups.¹⁵

Generative AI, particularly general purpose LLMs, seriously challenges this intricate nature. The first and obvious challenge is, of course, the ease with which representations of past events and historical figures can be misrepresented or manipulated through generative AI, whether intentionally or as a result of unintentional inaccuracies and ‘hallucinations’. These, in turn, can penetrate and harm the fundamental information layer, on which collective memory is then built. Yet, AI can potentially impact collective memory in other, more subtle ways, by challenging the additional, important, layer, that allows for diverse narratives and social constructions. As I argued elsewhere, due to the technological paradigm underlying LLMs that relies on statistical frequency, their outputs are likely to be concentrated and geared toward the popular and mainstream—projecting a ‘narrow universe’ rather than a breadth of voices and narratives.¹⁶ For example, in exploring the default outputs generated by LLMs when asked about prominent historical figures, we found that their answers converged around a small number of personae (such as Lincoln, or Darwin), and were substantially less diverse than human responses.¹⁷ This, in turn, implies that when asked about the past, LLMs’ default outputs might oust niche and broader perspectives that allow smaller communities to narrate their collective memory.

A second, related, tension concerns raw materials. Collective memory is facilitated by, and often depends on, access to raw materials. These materials can include iconic documentations of major events but also piecemeal recordings of more mundane occurrences, embodied in photographs, diaries, videos, witness statements, letters, drawings, and other artifacts, that together enable to draw the multi-faceted narratives comprising collective memory.¹⁸ Access to raw materials may be particularly important for the collective memory of extreme and radical events. As historian Saul Friedlander explained in the context of the Holocaust, the extremity, depravity, and scale of events, raise a “problem of representation:”¹⁹ It is difficult for people under normal circumstances to grasp the event as a reality that actually happened. In such circumstances, providing broad access to raw materials—not through paraphrase or summary—could be crucial for bridging and alleviating the problem of representation.²⁰

A third potential friction between AI and collective memory concerns trust. Traditionally, collective memory has been mediated through public or quasi-public institutions, such as museums, libraries, and archives that view the preservation and mediation of intergenerational memory as their primary mission.²¹ Even when private parties, such as Google’s Cultural Institute, are engaged in the provision of collective memory, they are still perceived, and often describe themselves, as public-oriented.²² As a result, people have a social expectation that entities active in the field of collective memory would adopt a public-oriented approach and are inclined to trust their outputs.²³

Here too, AI complicates things. Even setting aside, for a moment, intentional manipulation of authentic historical materials through generative AI, it is far from clear that AI providers, which are typically private tech-giants would prioritize the public interest when it comes to potential externalities of their activities on collective memory. Users’ intuitive trust, in other words, may be jeopardized or compromised. ²⁴

Altogether, aggregating these properties of collective memory against the realities of AI clarifies that the governance schemes applicable to AI can either exacerbate the disruption of collective memory or conversely mitigate those tensions.²⁵ The following sections elaborate and demonstrate.

Competition is not intuitively associated with collective memory. However, collective memory’s multifaceted nature clarifies the potential influence of a diverse AI arena, especially when it comes to general-purpose LLMs. The previously discussed tendency of LLMs to generate mainstream, concentrated outputs highlights the significance of ensuring users’ access to several LLMs. While each LLM is still likely to project a narrow and concentrated view of the past, our study indicates that aggregating outputs from several LLMs can increase diversity, especially when combined with other diversity-inducing methods, such as raising the models’ temperature or prompting for diversity.²⁶ To return to the previous example, when asked about personae from the 19^th century, an aggregation of models generated a longer “tail” of outputs, adding additional figures to the initial condensed list, including Albert Einstein, Florence Nightingale, Friedrich Nietzsche, Marie Curie, alongside others.²⁷

Even if each individual user were to consult only one or two LLMs when seeking information about the past, the availability of multiple models on the market means that different users will encounter somewhat different narratives.²⁸ Taken together, this diversity creates more opportunity for collective memory to flourish. Moreover, a diverse market of models would also increase users’ ability to identify errors, manipulations, and hallucinations in historical outputs generated by a single LLM.

According to reports, several countries, including Japan, Germany, France, and South Korea, have been engaged in the LLM market as active stakeholders by building, funding or otherwise supporting different “public LLMs“.²⁹ Such LLMs would presumably be trained on datasets in the local languages, which means that their outputs would give more prominence to the cultural and historical narratives of local societies.³⁰

The lens of collective memory helps to evaluate such policies. On the one hand, direct state involvement in media markets can raise free speech and other concerns. Indeed, public LLMs could allow states to influence collective memory by delineating the boundaries of their training datasets, which in turn would impact their outputs. On the other hand, the foregoing analysis indicates that collective memory is a public good, characterized by non-rivalry and non-excludability.³¹ This, according to standard public goods theory, means that leaving collective memory entirely to the market may result in a market failure: private AI companies will have no incentive to prioritize robust access to the diverse building blocks of collective memory. Cultural and historical heritage materials of local and niche communities, especially in languages other than English, may not penetrate training datasets on which AI is based. The omission of these materials from the models’ outputs will push them (and the narratives built from them) to the far margins of collective memory. “Public LLMs” may alleviate this concern and ensure the collective memory of specific societies and communities does not “drown” in an ocean of global or mainstream outputs.³² Their introduction—alongside diversity in the private LLM market—may thus be desirable.

One intense debate centers around the question of whether the training process itself constitutes permissible use or whether it requires the consent of the owners of copyright in the training materials (this question exceeds our scope and I do not explore it here).³³ Yet, from the perspective of collective memory, the influence of copyright starts even earlier, at the stage of digitization. Digitizing historical materials and archival collections—typically held by remembrance institutions, such as museums and archives—not only makes vast amounts of items accessible to broader audiences online, but also plays a critical enabling role in AI. Because a digital format is a precondition for inclusion in the training data on which AI is based, such digitization can also allow these heritage materials to enter the universe of algorithmic collective memory.

Copyright may hinder such digitization. Due to its typically long duration, it can pose an obstacle to the digitization of historical materials even if they were created a few generations ago.³⁴ This, in turn, may result in a ‘file drawer’ effect: what is not included in the digital space will only be seen in the (limited) space of the physical museum, or worse, will end up in a file-drawer. In the long term, AI’s disregard for such materials could lead to their marginalization—and the marginalization of the narratives built upon them—to the outermost periphery of collective memory.

 Additionally, when it comes to materials documenting historical events—especially extreme events—there is an inherent tension between copyright law’s “fair use” exception and digitization efforts. The doctrine of fair use, adopted by various jurisdictions, allows the use of copyrighted materials without the consent of the copyright owners, under certain circumstances. However, the exception prioritizes partial and transformative uses, that alter the original work and do not copy it in its entirety.³⁵ This may discourage AI developers from providing users with access to complete, verbatim, copies of cultural and historical heritage materials. Yet, as the previous discussion suggests, providing full access to raw materials—rather than paraphrasing or summarizing—is crucial to mitigate the problem of representation. ³⁶

These misalignments can be mitigated by calibrating copyright law in different ways: for example, by crafting specific exceptions that would enable digitization of entire collections held by remembrance institutions, as recently adopted by some jurisdictions,³⁷ or by a purposeful interpretation of the fair use doctrine that would facilitate the digitization and digital-display of historical materials in their entirety.³⁸ A detailed review of these schemes exceeds our scope. My point here is more general: When determining the different ways in which copyright may or may not apply to AI, legal policy makers should take into consideration the possible impact of any copyright governance scheme not only on current and future stakeholders, but also on our collective past.

Finally, regulation applicable directly to the AI field could clearly ease the frictions between AI and collective memory. To illustrate, obligations imposed on AI developers to abide by data security standards could reduce the vulnerability of those systems to hacking and technical failures that could harm the authenticity and integrity of the historical contents they mediate, and mitigate the risks of collective memory manipulation.³⁹ Duties of explainability and transparency⁴⁰ can alert users to the shortcomings and limitations of AI as a memory agent—for example by detailing the datasets on which ‘virtual witnesses’ or ‘generative archives’ were trained. Additionally, incorporating “multiplicity” in AI regulation as a high-level governance principle, as I suggested elsewhere, could benefit collective memory by obligating–or incentivizing–AI providers to expose users or alert them to the existence of multiple contents and narratives beyond the generated outputs, and encourage them to seek additional information.⁴¹

More specifically, Guy Pessach and I previously proposed conceptualizing entities that mediate collective memory through AI as “memory fiduciaries”, who are subject to fiduciary obligations toward their users. Such conceptualization, we argued, would provide a flexible governance scheme, rooted in the common law, that could adapt to new developments in the interface between AI and collective memory and respond to novel tensions as they arise.⁴²

As part of those fiduciary obligations, we proposed recognizing a duty of integrity. Integrity means that entities that mediate collective memory through AI need to conform with “standards of accuracy and aspire [..] to be consistent with historical facts”.⁴³ The analysis above suggests that this duty is needed for mitigating concerns of distortion and manipulation of collective memory. Notably, the reminiscent “accuracy” obligations, recently introduced in the EU AI Act, could similarly promote this objective, and the foregoing analysis suggests that there are compelling reasons and for applying them to the mediation of collective memory.⁴⁴

As Derrida wrote three decades ago, “[t]here is no political power without control of the archive, if not of memory”.⁴⁵ AI will increasingly function as our modern “Archive”, the major source for obtaining information about our shared past. Our analysis suggests that regulation affecting AI, both directly and indirectly, could substantially influence this Archive, and with it, collective memory. To maintain a robust space in which collective memory can flourish, regulators in the AI field must adopt a dual vantage point—directing their gaze simultaneously both toward the future and toward our future’s past. Collective memory deserves a seat at the table of AI regulation.