Atelier 4: La numérisation des contenus inexploités: stratégies et outils

The moderator immediately sets the framework for the discussion: the issue is not merely digitizing content, but turning dormant content into resources that are genuinely useful for AI. He stresses the need to maintain a balance among usefulness, reliability, content governance, data quality, the public interest, and respect for the rights of authors and creators.

The moderator explicitly frames the workshop’s central question as the transformation of dormant content into resources that are « useful, reliable, and governed » for artificial intelligence, without losing sight of authors’ rights, data quality, and the public interest .

The moderator then refocuses the discussion on the French-speaking world by asking which content should be prioritized and according to what criteria. He also opens the prospect of more structured Francophone cooperation, in the form of shared corpora or a federation of commons, while recognizing the diversity of national and institutional frameworks.

The moderator explicitly asks what the criteria for prioritizing digitization should be in the French-speaking world and how to develop sovereignty through a common Francophone corpus or a « federation of commons », distinguishing, where necessary, between national and institutional corpora according to the applicable rules .

Speaker 1 explains that AI has already extensively exploited content accessible on the web, leading to a form of saturation. At the same time, a significant mass of heritage content remains offline and could enrich AI models if it were digitized and made accessible.

Speaker 1 raises the idea of a future "crash" in the AI market linked to the gradual exhaustion of data harvested from the web, then responds that, on the contrary, there is still an enormous amount of content that remains completely inaccessible on the Internet, sometimes even unknown, which justifies making it available in the digital society . He also notes that, in his own library, only 10% of the content is digitized and accessible, leaving 90% outside the digital sphere .

For Speaker 1, the challenge is not only to increase the volume of data available for AI, but above all to improve its quality and representativeness. If training data mainly reflect dominant languages and cultures, AI will mechanically reproduce biases in diversity, perspective, and historical depth.

Speaker 1 states that behind every AI model lies the question of data, both in quantity and quality, and connects this to the issue of which cultures, languages, and worldviews will be represented in future systems . He stresses that when AI responses are based on data representing only part of human knowledge, this necessarily introduces biases in diversity, perspective, and diachrony .

Speaker 1 presents heritage and cultural data as a strategic resource comparable to a raw material for AI. Accordingly, their production, availability, and control are matters of public sovereignty, especially because their digitization depends largely on public investment.

He explains that heritage and cultural data have become the raw material of artificial intelligence and that, when they are not born digital, they must be retrieved from libraries, archives, and museums, which requires investments often borne by the public sector . He concludes that making these contents available and discoverable is part of public policies also aimed at avoiding monopolistic situations and preserving alternative means of access .

Speaker 1 believes that the Francophonie can play a specific role because of its cultural and linguistic plurality. It can help prevent the erasure of knowledge, imaginaries, and languages in AI systems dominated by content that is already in the majority on the web.

He argues that issues of bias place the Francophonie before a dual challenge, while also giving it an important role precisely because of its plural dimension . He adds that, without corrective action, Francophone knowledge, cultures, and imaginaries risk becoming invisible in the AI ecosystem, even though most of the relevant collections remain inaccessible .

Speaker 1 argues that heritage digitization cannot be limited to creating visual digital reproductions. Content must become genuine heritage data, meaning contextualized, reliable, and structured so as to be truly useful to machines as well as humans.

He clearly distinguishes the simple provision of a digital facsimile from the more demanding work involved in producing “heritage data,” which must be contextualized or capable of being contextualized, and reliable . He also specifies that such data should be regarded as a strategic asset for society, which justifies investing in their quality and structuring .

For Speaker 1, effective use by AI depends on the entire technical digitization chain, not just the existence of a digital file. The quality of capture, quality control, full-text transcription, and collection structuring directly determines the reliability and usefulness of content for training models.

He describes the crucial role of digitization operators, whom he compares to “knowledge brokers,” explaining that the quality of their work and the calibration of digitization workflows determine the reliability of the final digital copy and therefore its value for training AI models . He adds that it is not enough to have attractive images and structured metadata: high-quality full text must also be produced, printed works and manuscripts alike must be processed, and collections must be transformed into structured data that machines can use directly .

Speaker 1 maintains that an effective digitization strategy requires shared standards, harmonized practices, and interoperable systems. The multiplication of portals and dissemination channels helps content circulate more widely, increases its visibility, and reduces dependence on single access points.

He explains that the French-speaking digital network is working to harmonize inventory practices, catalog data processing, and the consideration of copyright in order to raise the average level of digital maturity among its members . Later, he emphasizes the need to share norms and standards, promote interoperability, and multiply dissemination channels, giving the example of Belgian documents, which are more likely to be consulted if they are made available through specialized or international portals rather than a single institutional website .

Speaker 1 believes that the digitization of heritage cannot be left solely to private interests, because it serves a mission of public interest. Since these investments support access to knowledge, cultural diversity, and information sovereignty, they fall under public policy and public funding.

He recalls that digitizing heritage collections held in libraries, archives, and museums is costly and that this investment is most often made by the public sector . He adds that states have economic, legal, and moral responsibilities to protect, preserve, and enhance this shared heritage, while institutions must do more to advance these operations .

For Speaker 1, prioritization must first respond to the urgency of preservation. The most fragile media and unique content must be dealt with first, because their disappearance would be irreversible and some institutions are their sole custodians.

He explains that the first step is to redo or refine the inventory in order to know what is being preserved and what is at risk of disappearing . He cites magnetic tapes, which are highly fragile, as the top priority, followed by the press, audio sources, unique documents, and manuscript heritage, stressing that if libraries do not handle them, no one will .

Speaker 1 insists that digitization useful for AI requires institutions capable of mastering the entire technical, legal, and organizational chain. Strengthening this digital maturity helps avoid becoming dependent on service providers or more powerful economic actors.

He explains that, to get from heritage data to AI, it is necessary to master the entire chain of library and archival professions, which requires a certain degree of digital maturity; this is why the réseau francophone numérique is developing training programs and a maturity pyramid whose apex is AI . He adds that this capacity-building must make it possible to better manage public procurement, equipment, infrastructure, and partnerships, so as not to have terms dictated by economic actors due to a lack of expertise .

Speaker 1 sees the current rise of AI as an opportunity to put heritage digitization back on the political agenda. But he emphasizes that this window of opportunity will have an effect only if it is translated into explicit public priorities and funding programs suited to the entire digitization chain.

He states that AI is bringing digitization issues back to the forefront and offers an opportunity to once again raise awareness among public authorities of the need to fund the preservation and digitization of heritage . He notes, however, that for the moment, the policies observed are aimed mainly at regulating the use of AI in the public sector, rather than refinancing all the upstream chains that make this AI possible, which he describes as "the next battle" .

Speaker 1 acknowledges that one could choose to keep a distance from AI, but believes that this option has already been largely overtaken by actual use. Since generative AI is increasingly being used instead of search engines, action must be taken on the quality and completeness of the data that feeds it.

In response to Speaker 4's intervention, he explains that many people already use generative AI applications to get quick answers, which makes the question of whether such AI should exist partly outdated . He also invokes the idea of “information overload” to point out that the Internet already contains so much information that people rely on opaque ranking mechanisms; in this context, his goal is to make training data as complete and reliable as possible .

Speaker 2 complements the heritage-based approach by recalling that AI also needs recent and culturally relevant content. He therefore emphasizes that twentieth-century works and contemporary content, often still protected by copyright, are also essential to high-quality AI.

He states at the outset that AI systems need high-quality, culturally relevant data drawn not only from the public domain and older works, but also from contemporary and recent data generally covered by copyright .

Speaker 2 insists that the digitization of oral traditions and community knowledge cannot be conceived without consent and safeguards. To build trust, the intended uses must be clarified and the communities concerned must be allowed to retain some form of control over the future exploitation of their content.

He warns that questions of copyright and the protection of folklore must be addressed from the outset when dealing with unpublished heritage of local communities or Indigenous peoples, in order to avoid mistakes already made in certain experiences . He adds that, without control over the conditions of use and subsequent uses by AI systems, these communities will have very little desire to share their knowledge .

Speaker 2 shows that the risks are not theoretical: content collected for non-profit purposes can later be reused in commercial channels without recognition of the source communities. This type of example justifies putting a framework of trust in place from the digitization strategy stage.

In the discussion on oral content, he cites the example of a Kanak religious song recorded in New Caledonia by musicologists in the 1980s for non-profit purposes, then commercially reused twenty years later by a successful musician without crediting the source community . He concludes that future uses must be considered from the very design stage of the strategy in order to regulate such reuses .

Speaker 2 presents copyright not as an absolute obstacle, but as a framework to be balanced with preservation and access. He highlights limitations and exceptions as a tool enabling States to support the public interest while respecting authors' rights.

He explains that his remarks concern how to digitize 20th-century analogue works while respecting authors' rights, particularly through the issue of limitations and exceptions to copyright for preservation and public access . He sets out the “three-step test”: exceptions must apply to certain special cases, must not conflict with the normal exploitation of the work, and must not unreasonably prejudice the legitimate interests of authors .

Speaker 2 calls for avoiding confusion between several operations that are often conflated in public debate. Authorizing access to content, making it visible online, and using it to train AI fall under different legal regimes and have different legal implications.

He explicitly distinguishes several scenarios: giving researchers access to collections is not the same as making a document accessible on the Internet . He adds that digitization intended for text research with limited previews, such as Google Books, differs from using digitized works as training data for AI .

Speaker 2 emphasizes that making digital content accessible does not necessarily mean it is free. He points to the growing existence of licensing contracts and structured content markets for AI, which requires thinking through the economics of these uses.

He states clearly that “digitized” and “accessible” do not mean free of charge, then mentions the proliferation of licensing contracts between content holders and AI systems . He adds that companies are presenting to WIPO specialized marketplaces for structuring multilingual datasets, demonstrating the existence of real economic demand for this content .

For Speaker 2, the issue is not only access, but also the distribution of the value created from digitized content. This means considering licensing, remuneration mechanisms, collective management, and the effects of public-private partnerships on rights and benefits.

He identifies as a third message the need to reflect on value sharing between content creators and AI companies, taking into account the possibilities of licensing agreements . He also calls for examining the consequences of public-private partnerships for the digitization of content, particularly regarding the transfer or retention of rights, emphasizing that everything depends on the policy objectives being pursued .

Speaker 2 stresses that a strong information infrastructure is essential for making digitization work in the age of AI. Metadata, identifiers and semantic layers make it possible to identify authors, ensure attribution and effectively manage the associated rights.

He states that a robust infrastructure for attribution and rights management must be put in place, based on a semantic layer describing works and making it possible to identify their authors . He illustrates this requirement by citing several standardized identifiers - ISBN, ISSN, DOI, ISAN, ISRC, ISWC - and mentions standardization work carried out notably by the national libraries of Finland and Latvia .

Speaker 2 explains that metadata are not used only to describe works, but also to govern their circulation. They can incorporate the purposes of digitization, restrictions on use, and levels of accessibility, so that these rules are readable and enforceable by machines.

He answers a technical question by saying that it is entirely possible to include, in the semantic layer describing digitized objects, the purpose of digitization and the conditions of use, whether the access is broad, restricted, or sensitive . He adds that extremely precise metadata make these objects visible to machines and usable in accordance with the policy set by the State .

Speaker 2 responds to the idea that some unused content would remain worthless by showing that its value depends heavily on its uses. Elements that may appear banal or commercially insignificant can become very valuable for training specialized AI systems or meeting niche documentation needs.

He gives the example of a generative AI company needing images of helicopters in flight: to correct a weakness in its system, it may request the creation of a dataset from television news broadcasts, reports, documentaries, or films, and then have those images annotated, which is "worth a great deal" . He adds that a work with no commercial success is not necessarily without value, citing the French ReLIRE database, which contains niche works and self-references that are useful for specialists .

Speaker 3 recalls that not all societies preserve their memory in the form of documents already organized as they are in libraries. A digitization strategy focused solely on structured content would therefore risk excluding oral knowledge, particularly African knowledge, even though it is essential and relevant.

He points out that libraries work with structured data, whereas in some African societies information is primarily oral, which means digitization strategies must be adapted to avoid exclusion . He also stresses that oral information can be extremely important and relevant even if it is not based on any already formalized medium .

Speaker 3 warns of a risk of reproducing the economic imbalances already seen in other extractive sectors. The regions that produce knowledge could become mere suppliers of cultural raw material, while others would capture the value through commercialization and AI applications.

He reframes the issue by saying that we must avoid having producers of cultural content become merely consumers of products derived from their own heritage . He explicitly draws a parallel with raw materials in Africa, stating that AI could reproduce the situation in which one ‘produces, but remains poor,’ this time with knowledge .

Speaker 3 argues that a realistic strategy must begin upstream of digitization itself. It is necessary to consider the collection and temporary archiving of oral material, preservation methods, classification trade-offs, as well as the material and energy costs this entails.

He asks how orally transmitted knowledge should be archived, for example when an elder is recorded with a voice recorder, and stresses the need to think through a transitional phase between collection, archiving, and digitization . He adds that the strategy must also take into account preservation methods, trade-offs over what is sovereign or shareable, as well as the costs linked to servers, storage, data centers, and even the lack of electricity .

Speaker 3 recalls that digitization is not just a cultural or technical project: it depends on substantial physical infrastructure. In contexts where energy and basic equipment are lacking, a large-scale shift to digital poses major financial and environmental constraints.

He states that the major problem is funding, because digitization is expensive and requires servers, data centers, and storage capacity . He adds that, in some contexts, the very absence of electricity makes these strategies difficult and calls for reflection on what is sustainable for the planet if everyone digitizes everything .

Speaker 5 warns of future falsification or distortion of historical narratives if local heritage, particularly that transmitted orally, is not digitized. Without this work, AI tools may reproduce incomplete, translated, or distorted versions of important cultural references.

She gives the example of Senegal and suggests that a national hero could one day be reproduced by AI in a degraded or Frenchified form simply because oral transmission had not been digitized and made usable . She adds that if systems are queried about information that has not been digitized, there is a risk of ending up with a falsification of history .

Speaker 5 believes that simply collecting oral knowledge is not enough: it must be structured and made reliable in order to be transmitted properly in the digital age. In her view, this structuring is a prerequisite for preserving cultural belonging and ensuring reliable uses in AI.

She stresses that heritage must be digitized and structured, particularly in Africa where much information remains oral, so that everyone can know what culturally belongs to them within diversity . She explicitly raises the question of when information is sufficiently reliable to be made available to everyone, making reliability a central element of the strategy .

Speaker 5 presents intellectual property as a tool for ensuring the reliability of digitized content and framing its use. Integrating it early into the strategy would help better secure the circulation of knowledge and strengthen its credibility in digital environments.

She emphasizes the importance of also being able to rely on intellectual property, which she associates with a form of information reliability . She states that if this aspect is integrated very early into the strategy, it could help better regulate and secure content .

Speaker 6 asks how concerns about information bias can be translated into operational digitization priorities. The intervention emphasizes the need for an action plan, recommendations, standards, and political and financial commitment to make the objectives effective.

Speaker 6 asks whether priorities have already been identified in libraries and in data classification, along with recommendations for digitizing certain categories of information . She also emphasizes that without an action plan, funding, national political priorities, and strong mobilization, it will be difficult to achieve the goal of reducing bias in online content .

Speaker 4 takes a more skeptical position and points out that there is an irreducible gap between the encyclopedic ideal of total access to content and the constraints of the real world. In his view, a great deal of content will remain unused, either for legitimate reasons or because no institution can materially process or make everything accessible.

He offers the fiction of a “Martian” arriving in a thousand years to write the history of Earth and wanting to consult a virtually infinite mass of archives, conversations, tickets, private correspondence, unsold collections, photos, and videos, in order to illustrate the scale of potentially unused content . He concludes that such content remains so in part for legitimate reasons today, and that libraries and legislators cannot be expected to work according to such a totalizing ideal .