Modern Times 1936
What is it that software sees, hears and perceives when technologies for pattern recognition are applied to media historical sources? All historical work requires interpretation, but what kind of algorithmic interpretations of modernity does software yield from historical archives? MODERN-36 is empirically committed to everyday experiences and sets out to study how machines interpret symbols of modernity in media from 1936. By utilising photographic and audiovisual collections, MODERN-36 seeks to analyse how modern Sweden was, while also exploring how computational methods can help us understand modernity in new ways. MODERN-36 will explore how artificial intelligence and machine learning methods can foster new knowledge about the history of Swedish modernity––while at the same time critically scrutinizing algorithmic toolboxes for the study of the past. MODERN-36 will use three datasets from 1936: 15,000 digitized photographs from DigitaltMuseum, all surviving radio programs from Swedish Radio, and all weekly newsreels and short films produced by Svensk Filmindustri. The research focuses on modernity in relation to gender, urbanity and industrialization, and will: (1.) to examine how software can assist historians in discerning new historical knowledge, (2.) to construct midsize and curated datasets that increase the scholarly capacity of media historical sources, and (3.) to interrogate algorithmic detection by evaluating what machines can—or cannot—notice in historic data.
Final report
If Charlie Chaplin once struggled to understand an industrialized world with gigantic machines in his film Modern Times (1936), a common denominator in the research project (with the same name) has been to explore how digital methods can help us understand modernity in new ways. Modern Times 1936 has been a research project at the intersection of media history and digital humanities, that has partly studied how software interprets symbols of modernity in media archive material, and partly focused on how data-driven methods can help us understand the history of modernity in new ways. The project's purpose has therefore been both about how the scientific capacity of media historical source material can be expanded using digital methods. At the same time, researchers in the project have critically examined what algorithmic recognition techniques and software can (and cannot) identify in media historical datasets. The project has asked questions such as: what do contemporary algorithms for pattern recognition see and hear when applied to media historical source material? All historical work requires interpretation, but what kind of algorithmic interpretations of modernity does software yield from historical archives?
It should be noted that the research project coincided in time with the global launch of new language models, ChatGPT and not least generative AI – a technological shift that influenced research, and provided historiographic impulses. Among other things, the project became interested in the historical gaze of generative AI (Stjernholm, Eriksson & Mohammadi Norén 2025), so-called algorithmic upscaling of early film on YouTube (Stjernholm & Snickars 2024), techniques for assessing photorealism in synthetic images (Eriksson 2025), and distant reading of parliamentary records (about radio) from the 1930s (Snickars, Mohammadi Norén & Jarlbrink 2026). In digital humanities fashion, the project has also spent time compiling a number of datasets based on photographic and audiovisual collections from the Swedish interwar period: a dataset of sound files from the 1930s; a dataset of photographs from 1936; a dataset of silent films from Swedish Biograph; a larger dataset of photographic material from the 1930s (in collaboration with Nordiska museet), and an extensive dataset of the so-called SF archive with around five thousand newsreels and short films.
These datasets have been processed within the project with various software: in the case of audio, speech-to-text models have been tested and used for transcription – the project has used the Swedish National Library’s wav2vec model with good results. Regarding film, researchers within the project have been interested in algorithmic upscaling, that is, using different methods and models to increase (and change) pixel density. A model (Real-ESGRAN) was, for example, used to restore the original quality of old silent films – a kind of automated restoration – where the project also used classic film techniques such as tinting, toning and colouring to illustrate the ways in which algorithmic upscaling resembles film restoration. In terms of photography, the project has used models for object recognition, mainly visual transformation models, both on original photographs from 1936, and on the same images in upscaled form, where the model marked different objects depending on image quality, with the possibility of filtering images with modernity markers. In this way, the project has been able to compile several datasets (Aspenskog & Johansson 2025a; 2025b), and a series of applications – all openly available on the developer site Github (Johansson 2025a; 2025b; 2025c).
The project has been exploratory in nature. Method development and preparation of datasets has taken a lot of time – sometimes at the expense of research results. Nevertheless, a number of results have been reported on the project blog (modernatider1936.se), as well as at several conferences such as NECS (2025), DHNB (2024) and ECREA (2024). In a concluding joint article, the project researchers use several of the methods and models we have previously worked with, and apply this to the SF archive as a dataset: text mining of intertitles from silent films (with the application stum), signal archaeology (of all sound in the SF archive), as well as NER models and geocoding of five thousand films using the application SweScribe (automatic speech recognition), which can transcribe and timestamp speech from the collection's audio track (Aspenskog et al. 2026).
The three most important results of the project are: (1.) new applications offer significant opportunities for analyzing older media material – from the ability to convert narration in newsreel films into searchable text, algorithmic upscaling of film, to models for object recognition in photographs; (2.) the relationship between historical content (in different modalities) and generative AI is very complex; historical text can today be synthesized into adequate statements, but with regard to images, films and sound, even moderate results are still lacking; (3.) However, technological developments in generative AI are very rapid, with visual transformation models such as Google's Veo 3.1 now being able to produce historical newsreels that look authentic. Synthetic media, in both a contemporary and historical sense, will all likely lead to new research questions that challenge the relationship between digital technology, media and the past.
It should be noted that the research project coincided in time with the global launch of new language models, ChatGPT and not least generative AI – a technological shift that influenced research, and provided historiographic impulses. Among other things, the project became interested in the historical gaze of generative AI (Stjernholm, Eriksson & Mohammadi Norén 2025), so-called algorithmic upscaling of early film on YouTube (Stjernholm & Snickars 2024), techniques for assessing photorealism in synthetic images (Eriksson 2025), and distant reading of parliamentary records (about radio) from the 1930s (Snickars, Mohammadi Norén & Jarlbrink 2026). In digital humanities fashion, the project has also spent time compiling a number of datasets based on photographic and audiovisual collections from the Swedish interwar period: a dataset of sound files from the 1930s; a dataset of photographs from 1936; a dataset of silent films from Swedish Biograph; a larger dataset of photographic material from the 1930s (in collaboration with Nordiska museet), and an extensive dataset of the so-called SF archive with around five thousand newsreels and short films.
These datasets have been processed within the project with various software: in the case of audio, speech-to-text models have been tested and used for transcription – the project has used the Swedish National Library’s wav2vec model with good results. Regarding film, researchers within the project have been interested in algorithmic upscaling, that is, using different methods and models to increase (and change) pixel density. A model (Real-ESGRAN) was, for example, used to restore the original quality of old silent films – a kind of automated restoration – where the project also used classic film techniques such as tinting, toning and colouring to illustrate the ways in which algorithmic upscaling resembles film restoration. In terms of photography, the project has used models for object recognition, mainly visual transformation models, both on original photographs from 1936, and on the same images in upscaled form, where the model marked different objects depending on image quality, with the possibility of filtering images with modernity markers. In this way, the project has been able to compile several datasets (Aspenskog & Johansson 2025a; 2025b), and a series of applications – all openly available on the developer site Github (Johansson 2025a; 2025b; 2025c).
The project has been exploratory in nature. Method development and preparation of datasets has taken a lot of time – sometimes at the expense of research results. Nevertheless, a number of results have been reported on the project blog (modernatider1936.se), as well as at several conferences such as NECS (2025), DHNB (2024) and ECREA (2024). In a concluding joint article, the project researchers use several of the methods and models we have previously worked with, and apply this to the SF archive as a dataset: text mining of intertitles from silent films (with the application stum), signal archaeology (of all sound in the SF archive), as well as NER models and geocoding of five thousand films using the application SweScribe (automatic speech recognition), which can transcribe and timestamp speech from the collection's audio track (Aspenskog et al. 2026).
The three most important results of the project are: (1.) new applications offer significant opportunities for analyzing older media material – from the ability to convert narration in newsreel films into searchable text, algorithmic upscaling of film, to models for object recognition in photographs; (2.) the relationship between historical content (in different modalities) and generative AI is very complex; historical text can today be synthesized into adequate statements, but with regard to images, films and sound, even moderate results are still lacking; (3.) However, technological developments in generative AI are very rapid, with visual transformation models such as Google's Veo 3.1 now being able to produce historical newsreels that look authentic. Synthetic media, in both a contemporary and historical sense, will all likely lead to new research questions that challenge the relationship between digital technology, media and the past.