Propaganda images reveal how terrorists think

Researchers have used artificial intelligence to analyse over 30,000 propaganda images produced by the Islamic State (IS) between 2014 and 2022. 

A group of IS soldiers with weapon and an IS flag.
From Diyala, Iraq, in the spring of 2022: a group of IS soldiers swears Bay’ah – an oath of allegiance to the new IS caliph. This is one of the 30,000 IS images that have now been analysed. 

Propaganda is an important part of the operation of a terrorist organisation. They use text, video, audio, and images to not only spread fear, but also to recruit and build their image internationally. By using the internet and social media, even terrorist groups with limited resources can reach a global audience.

For researchers studying terrorist groups, propaganda material is an important source in understanding who the actors are, what they want, and how they operate. The challenge for researchers is the sheer amount of propaganda that is being produced, which makes it almost impossible to get an overview through manual means. 

Because IS controlled a large area in Syria and Iraq for several years, they gained more resources and more members who could work as ‘journalists’ or content producers. Propaganda was clearly a prioritised area in the organisation, and the battles they were involved in provided them with plenty of raw material. 

‘IS stood out because their propaganda was better than the content of earlier terrorist groups. They managed to create a “brand” and a linguistic and visual expression that appealed to the youth, combined with extremely brutal content where close-ups of killings were almost a given,’ explains Skretting. 

Although the group is now marginalised in its original core areas in Syria and Iraq, its branches in other parts of the world have made significant progress – especially in Africa. IS still maintains a unified and substantial propaganda apparatus. 

IS – a pioneer in terrorist propaganda 

The researchers have based their work on the extensive image material published on IS’s official channels over the years. The dataset, consisting of 30,000 images, was fetched from an IS bot on Telegram. 

‘IS has revolutionised how non-state actors and terrorist groups conduct propaganda. This is considered one of the keys to the group’s success,’ says researcher Vidar Skretting. He co-authored the report with colleagues and AI researchers Mathias Bynke and Bernt Ivar Nødland.

The rise and fall of IS 

In 2013, a group called the Islamic State in Iraq and the Levant (ISIL), managed to occupy large parts of Syria and Iraq. In 2014, they seized Mosul, Iraq’s second-largest city. Shortly thereafter, ISIL declared that they had established a new caliphate, to which all Muslims were obliged to submit. At the same time, they removed the geographical reference from their name and adopted ‘the Islamic State’ as their designation, or IS for short. 

From 2014 to 2018, IS gained international notoriety for its brutal conduct. They implemented a reactionary form of Islamic law and carried out executions and amputations of prisoners, genocides against minorities, public slave trading, and numerous terrorist attacks on civilian targets in the Middle East, Asia, Africa, and Europe. 

At the same time, they wanted to show that they had established a true Islamic ‘state’. The group emphasised creating a civil state apparatus and established institutions for law and order, tax collection, education, and healthcare. In addition to violence and brutality, the ‘civil’ side of IS became a recurring theme in the group’s propaganda production. 

IS reached its territorial peak in 2015, when the group controlled about one-third of both Iraq and Syria. After this, IS was gradually pushed back by an international coalition. The last IS-controlled area was captured by Kurdish forces in early 2019. Thus, IS went from being a state-like entity back to an underground organisation.

A dedicated IS social media 

Initially, IS used established social media platforms like Twitter (X), Facebook, YouTube, and, later, the Telegram app to spread their content. 

‘The spread of propaganda on social media is one of the reasons so many foreign fighters from around the world joined IS,’ says Skretting. 

‘Today, it’s much more difficult for IS to reach out. Major social media platforms remove such content almost immediately after it’s posted. However, IS still manages to disseminate its content effectively through other channels, and it’s relatively easy for sympathisers to find them,’ Skretting adds. 

IS spreads its propaganda through three main channels: ‘private’ social media platforms that they run on their own servers, bots on Telegram, and regular indexed websites. 

‘The private social media platforms IS runs have been active for years. Telegram bots and the indexed websites are regularly taken down by administrators and authorities but are usually reopened by IS under different names shortly thereafter,’ says Skretting. 

Two researchers present at an FFI event.
Mathias Bynke (to the left) and Vidar Skretting (to the right) present their report during an FFI event on 12th of December 2023. Photo: FFI / Anders Halvorsen Fehn.

Sorting 30,000 images 

In their work, the researchers used a machine learning model from OpenAI called Contrastive Language-Image Pre-training (CLIP). The model is trained to compare images with text. 

CLIP has been trained by collecting enormous amounts of images and corresponding captions from the web. By comparing millions of images and captions, the model has gradually learned how images can be described in words and, conversely, how an image might look based on the text describing it. 

By running the 30,000 images through CLIP, the researchers could, for example, ask the model to find all images showing a ‘person praying’ or ‘combat action.’ However, instead of manually defining the categories in which the images should be sorted, the researchers used a clustering algorithm to group the images by theme. 

‘We didn’t know in advance which descriptions would suit various themes. When you provide the CLIP model with a limited number of categories, an image of a man diving into a river could easily end up in the category “image of someone praying”,’ explains Mathias Bynke.

Illustration of the Clip model.
The Clip model.

The CLIP model translates an image into an embedding vector, that is, a sequence of numbers. It also translates the image captions into a sequence of numbers. If you have an image of apples, the number sequence for the image should be approximately the same as the number sequence for the text ‘image of apples.’ 

The number sequence that CLIP generates for each image can be converted into coordinates in a coordinate system. This gives each image a position on a map. Then, an algorithm can cluster these points into groups. You can choose how many clusters you want, and the algorithm organises the images for you. 

When the researchers asked the algorithm to create two clusters, one group was dominated by military images, and the other by civilian images. 

In the end, they settled on 14 different named clusters: combat scenes, soldiers outside combat, enemy bodies, executions and killings, close-ups of individual fighters, weapons, civilian crowds, peaceful scenes, collapsed buildings, civilian casualties, food, public works, crafts and industrial production, and burning cigarette packets.

What have we learnt about IS from the images? 

Once the images were sorted, the researchers wanted to answer the following question: How has IS’s image propaganda evolved from 2014 to 2022, and what does this development tell us about IS as a group? 

‘The images IS publishes, the number of images published, and where they are published give us an indication of what IS is doing, how active they are, and in which areas they’re active,’ says Skretting. 

He believes that insight into the development of terrorist organisations can be important in predicting their direction and strategies: By looking at the images from the African provinces where IS is on the rise, we might gain insight into where IS is heading in the future. 

The analysis shows that in the early years (2014–2018), IS presented itself as both a military organisation and a civil state apparatus. This is linked to their attempt to build a ‘real’ Islamic state. 

Since 2019, the focus has shifted. IS now presents itself almost exclusively as a military movement and insurgent group. 

Two men working.
An example of one of the ‘civil’ images of IS’s Iraq and Syria branch, in the ‘public works’ cluster.

2015 was the year IS was most active in terms of propaganda. Nearly half of all the images in the dataset are from this year, which coincides with the time when IS was at the height of its power in Syria and Iraq. Thus, most of the images in the material are from Syria and Iraq. 

‘The group has been defeated in their previous primary areas but is rebuilding in Africa. We also see this trend in the image material,’ says Skretting. 

Activity was at a low point in 2020 but has since increased. The majority of images are now published in West and Central Africa. These differ from those produced in the Middle East in several ways. 

‘There is much less focus on individual jihadists and martyrs. At the same time, the propaganda images are generally more brutal. They’re dominated by military content, reflecting the fact that they are most active as a military organisation,’ says Skretting. 

Further development of the methods is needed 

The researchers conclude that the combination of CLIP and clustering algorithms is an effective method for quickly analysing large amounts of propaganda images. 

‘The division into clusters was not perfect. A certain proportion of the images appeared misclassified, and we had to implement mechanisms to filter these out. But the method helps to speed up the analysis,’ says Skretting. 

He emphasises that the combination of CLIP and clustering algorithms can be used for more than analysing terrorist images. 

‘The method is relatively simple and scalable. You can easily use it to sort image collections far larger than 30,000 images into thematic clusters. The method is particularly useful if you don’t know a lot about the content of the image material in advance.’ 

The researchers believe we should further develop machine learning methods to map propaganda from state and non-state actors. 

‘This type of method is not only relevant for image analysis but also for getting an overview of large text, audio, and video materials,’ stresses Mathias Bynke. 

FFI-Report 2026

Det militærteknologiske kappløpet – utviklingstrekk under stormaktsrivaliseringen

There is an ongoing international competition underway to leverage emerging technologies in military operations. This military-technological arms race is shaping the operational environment of the Norwegian Armed Forces. Therefore, it is important to understand the underlying trends and the broader security context associated with this competition. This report examines several issues related to the relationship between the United States and China, with a focus on China as the challenger to the technological edge that the United States has traditionally enjoyed. The race is related to several global trends and is driven by the interplay between intensifying international rivalry and revolutionary technological advancements. The military-technological race, in turn, affects the distribution of power in the international system and is, consequently, relevant for Norwegian security and defence planning. Emerging and disruptive technologies play a central role in industrial development in general. The integration of such technologies into military capabilities requires new types of civil-military technology cooperation. Concepts such as civil-military integration and military-civil fusion are strategies for addressing this challenge. In addition, an important objective with such strategies is to optimize the use of resources to ensure that the economy remains sufficiently sustainable to compete in the long term. Another important trend is China’s rise in international science and technology. Scientific progress provides an essential base for integrating technology into military capabilities. Technology alone, however, is not sufficient. Conceptual innovation and an appropriate organizational framework are also necessary for armed forces to realize its full potential. Still, considerable attention is generally devoted to the USA’s need to maintain, and China’s desire to close, the technology gap. Our report is based on a review of existing literature and thus provides a synthesis of mainly English-language knowledge on this topic. By compiling secondary sources across a broad range of subtopics, we intend to describe how single topics relate to one another and thereby contribute to a more comprehensive understanding of the subject.
FFI-Report 2026

Personellstatistikk for Forsvaret – perioden 2018–2025

The Norwegian Armed Forces is a major employer in the Norwegian labour market, and the long-term defence plan calls for a substantial increase in personnel. This expansion requires targeted efforts, and the ability to recruit, develop and retain skilled personnel will be critical to operational capability. This report analyses developments in selected key indicators for personnel policy in the Armed Forces. It provides a status assessment at the start of the long-term plan period and a fact-based foundation for monitoring progress and informing future decisions and priorities. The report first analyses developments in key indicators for military and civilian personnel. It then examines attrition patterns and internal mobility, including who leaves and the scale and nature of job rotation and promotions. The analysis covers 2018–2025, with an emphasis on 2025. The personnel indicators show that the Armed Forces is progressing well against its planned increase in both military and civilian personnel. The share of military personnel employed on long-term contracts has risen, alongside the share of women in the military workforce. At the same time, the age profile points to a future need for recruitment and knowledge transfer, particularly among officers. Young specialists and women have higher attrition rates than other groups. The officer corps also faces a persistent challenge, as many experienced officers will reach the Armed Forces’ special retirement age in the coming years. Although overall officer attrition is low, it is highest among those aged 30–39. Sustained attrition in this experienced mid-career group could eventually result in shortages of qualified officers for key leadership roles positions requiring specialised expertise. The analysis of internal mobility shows that job rotation has increased over the past two years for all military personnel groups except senior officers. Average tenure in a position is also short. As new personnel are primarily recruited and trained at lower organisational levels while growth occurs across several levels, substantial internal mobility will be needed to fill new positions. The underlying structural drivers therefore suggest that job rotation and promotion rates are unlikely to decline. Instead, continued growth and attrition may require sharper prioritisation between positions that need longer tenure, personnel who should rotate frequently, and positions that may have to remain vacant for extended periods.
FFI-Report 2026

Environmental reporting and greenhouse gas inventory of the Norwegian defence sector for 2025

The reports in the series ‘Environmental reporting in the Norwegian defence sector’ are published annually by the Norwegian Defence Research Establishment (FFI). These present data reported by the defence sector and associated partners to the Norwegian Defence Environmental Database (NDED). The reports provide an overview of results and trends for environmental aspects of the defence sector’s operations, including waste production, energy expenditure, fuel consumption, use of ammunition, water consumption, consumption of chemicals, and acute pollution. Greenhouse gas emissions are presented in a greenhouse gas inventory. Waste generation is reported to the NDED by associated waste management companies contracted within the various regions of the Norwegian Defence Estate Agency (NDEA). The total amount of waste produced in 2025 was 18,320 tons, which represents an 8 % increase compared to 2024. The degree of waste sorting was 64.8 %, a reduction of 0.3 percentage points compared to the previous year. 31.4 % of the waste was recycled, while 59.3 % was processed with energy recovery. Energy consumption associated with the defence sector’s buildings and properties in Norway is reported by the NDEA through statistics from suppliers. The total energy consumption in buildings and other properties is estimated to 702.5 GWh in 2025. This represents a marginal reduction compared to 2024. Of the energy used in 2025, 96 % came from renewable sources, which is about the same as the previous year. Fuel consumption connected to the use of vehicles, aircraft, vessels, and auxiliary power units was 100 066 m3 in 2025. This is an increase by approximately 2.4 % compared to 2024. Fuel consumption on aircraft and vessels represents 90.1 % of the total fuel consumption in the defence sector. The use of ammunition is reported and specified on a digital form (DBL-750) by organizational unit, shooting range, and ammunition type. A total of 22 223 554 units of ammunition were reported used in 2025, which is 8.4 % more than in 2024. The degree of reporting is the relationship between ammunition provided to the Armed Forces and the proportion reported being used. The degree of reporting in 2024 was 87 % (excluding blank ammunition), which is an increase of 2 percentage points compared to 2024. The reported use of lead-based small arms ammunition has decreased with 163 000 units, or 12 %, from 2024 to 2025. The estimated emission of lead is 5.5 tons in 2025, compared to 6.1 tons in 2024, a decrease of 11 %. Water consumption is reported by the NDEA based on measured and estimated volumes. The total water consumption in 2025 was 2.27 million m3, an increase of 4 % compared to 2024. The use of chemicals is reported from establishments within the sector where chemicals are used on a regular basis. Apart from de-icing fluids, it is insufficiently reported. 40 517 kg of aircraft deicing and 292 700 kg of runway deicing fluids were reported in 2025. The relative usage of urea to the total usage of runway deicing fluids was 96 % in 2025, an increase of 34 percentage points compared to 2024. The greenhouse gas inventory consists of reported fuel and energy use and emission factors associated with the various materials. Emissions from the defence sector’s activities were estimated to 264 512 tons of CO2-equivalents in 2025 in Scope 1 and 2, and 2 095 522 tons of CO2-equivalents when including indirect emissions not mandatory to reporting (Scope 3). Emissions in Scope 1 and 2 have increased by 2.4 % compared to 2024. There is a close relation between the demands and prerequisites which dictate the sector’s volume and pattern of activity and its total impact on the environment. It is therefore relevant to assess this impact in light of the tasks assigned to the defence sector within a dynamic political framework.
FFI-Report 2026

Forsvarsindustrien i Norge – statistikk for 2025

The report describes the scope and development of the Norwegian defence industry in 2025. It is based primarily on information provided by 92 companies. The respondents include the largest industry players and therefore represent a significantly larger share of economic activity than the 30 percent response rate would suggest. Total defence-related revenue amounted to NOK 61.4 billion, representing an increase of 33 percent compared with 2024. Revenue increased across all market segments. Sales to foreign customers grew by 32 percent and accounted for approximately 75 percent of total revenue, while exports increased by 31 percent to NOK 24.8 billion. Revenue generated in Norway rose by 37 percent to NOK 15.7 billion. Companies classified as large accounted for 75 percent of total revenue. Seven companies fall within this size category in this year’s survey, compared with four in the previous report. This should be considered when comparing results across size groups. Total defence-related R&D expenditure increased by 19 percent. Externally funded R&D grew by 28 percent and accounted for 73 percent of total R&D activity, while self-financed R&D remained largely unchanged. The companies reported 17,481 defence-related full-time equivalent (FTE) positions, 33 percent more than in the previous year. Of these, 10,092 were located in Norway and 7,389 in foreign subsidiaries. Order backlogs increased by 22 percent to more than NOK 171.7 billion. The majority of respondents expect continued growth in 2026.
FFI-Report 2026

Happy Days – sosiale erfaringer fra 100 dager patruljetjeneste i arktisk vinter

In 2025, two patrols from the Long-Range Reconnaissance Squadron in Finnmark conducted a 100-day winter exercise, FAM100. This report examines social conditions and interpersonal relationships during the exercise. The dataset consists of semi-structured interviews with all 13 operators, analyzed using an exploratory, grounded theory-inspired approach in MAXQDA 24. A key finding is that the operators largely experienced the exercise as positive, despite expecting significant mental and social strain. Following an initial adaptation phase, many developed a sense of flow, well-being, camaraderie, and humor, referred to as “Happy Days.” However, it became evident that this sense of well-being was linked to the safe framework and known end date of the exercise. It therefore does not fully reflect the realities of war. Several factors contributed to effective collaboration: extensive preparation, selection and voluntariness; shared culture and training; and managed expectations and established ground rules. Collaboration was strongest with a clear framework with defined roles and tasks requiring contributions from the entire team. Mission command, characterized by clear intent and a high degree of autonomy within each individual’s area of expertise, proved particularly effective. At the same time, the study identified several sources of strain. Sleep deprivation emerged as the single most critical factor, significantly affecting cognitive capacity, decision-making, and social tolerance. Additional stressors included physical wear and tear, skin problems, cold exposure, limited daylight, use of night vision devices, and persistent equipment challenges. Irritation primarily arose from minor issues, uneven workload distribution, and personality differences, amplified by prolonged close cohabitation. Relational challenges were largely managed through strategies such as rest, temporary isolation, shielding from strain, and explicit conflict management. Two different feedback models—structured “hot wash” sessions and more selective, ongoing correction—were practiced in the two patrols, without one emerging as clearly superior. Effective management appears to depend on context, group dynamics, and individual maturity. The study also identifies significant challenges related to equipment, logistics, and operational frameworks. Operators perceive the system as largely adapted to peacetime conditions, characterized by slow procurement processes and insufficiently adapted equipment. This affects both operational capability and motivation. Furthermore, a tension was observed between field-based risk understanding and the system-level need for control and safety. In conclusion, the operators highlight the need for broader transformation within the Armed Forces from a peacetime mindset to a more realistic understanding of war. This includes faster and more flexible procurement processes, greater preparedness for casualties and ethical dilemmas, as well as strengthened medical support and operational adaptation to demanding environments. The study thus provides important insights for planning and execution of future long-duration operations.
FFI-Report 2026

Kontraktsstrategier i Forsvarets investeringsprosjekter – etterlevelse av statens prosjektmodell

According to the long-term plan for the defence sector, there is need for extensive investment in new materials, buildings, and facilities. An important part of the basis for such public sector investment projects is a contract strategy. The contract strategy is a description of, amongst other things, the types of contracts that will be used, the planned compensation scheme for the supplier(s), and the distribution of tasks and responsibilities between the parties involved. This study investigates how contract strategies in the Norwegian defence sector comply with the state project model. To investigate contract strategy characteristics, we analyse the documentation for the conceptual appraisal phase (konseptvalgutredning – KVU) and the pre-project phase (sentralt styringsdokument – SSD) of a selection of projects that have recently completed their pre-project phase. To investigate how the projects are constrained in which contract strategies they can use, we analyse the mission letters for each phase. The contract strategies in our sample are in large parts decided before the pre-project phase. Through the analysis of mission letters we identify a number of decisions, the most common being: exemption from the Norwegian law for procurement, the investment being part of an international collaborative project, that procurement should be off-the-shelf, that procurement should be direct from potential suppliers, or that procurement should be carried out through the American Foreign Military Sales program (FMS). The identified decisions are often elaborated upon, but the elaborations are typically very brief. The state project model requires that at least two alternative contract strategies should be put forward and discussed. We find, however, that most contract strategies in our sample do not have alternative strategies laid out in the SSD. The lack of alternatives is mostly unexplained. There is also a requirement that the structure of the contract, the compensation format, and the distribution of tasks are discussed as part of the contract strategy. Though most of the strategies we studied include descriptions of structure, compensation format, and task distribution, there is generally no discussion and little or no elaboration of why certain structures, compensation formats, and task distributions were chosen. The SSDs in our sample generally lack descriptions and discussions of other requirements under the state project model, such as descriptions of contract type, incentive schemes, the degree of specification and so on.