This is supplementary material for our published full report. Its objective is not to summarise the entire report, but rather, to augment the report with data visualisations that could not be meaningfully captured in a 2D format. Correspondingly, this page begins by briefly summarising the key research findings and goes on to discuss the five technology clusters. The graphics in this section are interactive.

Main results

According to the experts surveyed:

• All twelve technologies will reach operational deployability by 2040 in the United States, Russia, and China. By then, these technologies are expected to be useable on the battlefield.
• Hypersonic weapon systems and directed energy weapons will make a significant jump in terms of their deployability by 2030. For hypersonic weapon systems, this could be reflective of an ongoing arms race.
• All twelve technologies will weaken either international stability or human security, and half of the technologies examined will predominantly weaken (rather than strengthen) both. The net impact of all the technologies was negative.
• Anti-satellite capabilities and AI-enhanced information warfare show uniquely negative impacts, particularly for crisis stability. Their possible second-order effects on human security raise serious concerns.
• AI and quantum technologies for command, control, communications, computers, intelligence, surveillance, and reconnaissance purposes are expected to have the most positive effects of any of the examined technologies. By increasing clarity and situational awareness, these military applications could strengthen crisis stability.
• Similarities in impact do not necessarily mirror technological similarities. Some technologies have very similar effects on international stability and human security, though they share no technical characteristics with one another.
• Great power competition will be the main driver behind future research and development in the United States, Russia, and China. A possible China-Russia alliance could confront the United States with serious military-technological challenges.

We term these combined effects ‘negative multiplicity’, reflecting the predominantly negative, concurrent, and in some cases, similar first- and second-order effects that emerging technologies are expected to have on international stability and human security.

Our clustering exercise revealed that the technology cluster with the most negative impact on international stability and human security consists of technologies that share few—if any—technical characteristics. We thereby conclude that a narrow research or policy focus on a single technology or a single area of impact obscures the similar effects of seemingly dissimilar technologies. Such a narrow focus also risks obscuring the significant second-order effects on human security that their use could generate. These issues become more pronounced when one considers those technologies that are expected to mature more rapidly over the next ten years. For example, the expected jump in the deployability of hypersonic weapon systems and directed energy weapons is worrying. For hypersonic weapon systems, this rapid change in deployability could reflect and exacerbate perceptions of an already ongoing arms race.

Our study underscores the negative environment created by geopolitical competition. Combined with the expected extent of technology trajectory alignment between the United States, Russia, and China, we find evidence of an ongoing arms race for emerging military technologies. Consequently, future arms control efforts will likely face resistance from the great powers, and existing agreements might come under additional pressure. Inversely, the expected technological alignment of the three countries could benefit arms control efforts. Arms control may seem more attractive when adversaries begin to close the competitive gap. In any case, technology governance in light of negative multiplicity will require recognising the interactive potential of emerging technologies. The window for preventive initiatives could soon close, however, if the deployability forecasts of our experts are correct.

The five technology clusters

In our study, technologies are clustered by their effects on stability and security, rather than their technical characteristics. Occasionally there is overlap between the technical characteristics and the technology clusters. Where they diverge, technology clusters offer a new way of thinking and talking about similarities and dissimilarities of impact in the emerging technological landscape.

Machine learning (ML) was used to identify the technology clusters. As discussed in the methods section of the report, the technology scoring exercise comprises 48 numerical features (i.e., numerical scores given by experts) per technology and twelve emerging technologies. When responding to the technology scoring exercise, the 30 expert respondents collectively generated approximately 8,000 data points. Whereas humans struggle to recognise patterns and identify relationships in datasets of this volume, ML is well suited to the task.

We created a three-dimensional impact graph by inputting the numerical scores from the technology scoring exercise into a clustering algorithm. The technology clusters are shown in Figure 1. The three axes correspond to the three dependent variables: arms race stability, crisis stability, and humanitarian principles. The five technology clusters are denoted by the colour of the dots.

Figure 1: Average scores (mean) for all technologies across all three dimensions, clustered into five groups

Each technology cluster comprises emerging technologies that were scored similarly by experts across the three dependent variables. Put differently, the technology clusters provide groupings of technologies that exist in the same three-dimensional ‘impact space’ on Figure 1. All three axes move in the same direction, meaning that the higher the value on each axis, the more impactful experts believe a given emerging technology is on that dependent variable. Although the technology scoring exercise gave expert respondents equal opportunities to express how emerging technologies might strengthen and weaken international stability and human security, ‘high impact’ here has a negative connotation. The higher the value on the axis, the more experts foresaw a given technology negatively affecting arms race stability, crisis stability, and/or compliance with humanitarian principles.

Figure 2 examines each dependent variable individually.

Figure 2: Average scores (mean) for all technologies across all three dimensions, separated onto 1-dimensional axes

Nuancing the technology clusters

The clustering algorithm used to group the technologies together treats the five technology clusters as mutually exclusive. In an effort to nuance the technology clusters, we sought to understand which technologies constitute the ‘core’ elements of each technology cluster and which technologies are ‘borderline’ cases that could conceivably fit into more than one technology cluster. To do this, we did two things: First, we used an internal validation technique called ‘Silhouette score’, which is an average ratio between the distances between data points within individual clusters (i.e., compactness) and the distance to the closest data point in separate clusters (i.e., separation). This score is the mean distance to data points within the cluster divided by the mean distance to data points outside its cluster. A lower Silhouette score for a given technology signifies that it is further from the other technologies in its cluster and closer to technologies in other clusters.

Figure 3 shows the Silhouette scores for each technology.

Figure 3: Silhouette scores for all technologies based on five clusters

Figure 3 highlights which technologies are least cohesive with their cluster and share features with technologies in other clusters. This includes directed energy weapons, AI for C4ISR and hypersonic weapon systems. Synthetic biology has a Silhouette score of zero (i.e., it does not have a bar) because the technique measures the distance between points in a group and synthetic biology is the sole technology in its cluster.

Second, we ran a clustering algorithm on each dependent variable independently. The outcome of this exercise is shown in the graph with the overlapping bubbles in Figure 4. We also include the binary technology clusters in Figure 4 to facilitate comparison.

Figure 4: Illustrating overlap between clusters based on clustering run separately on each dimension

In Figure 4, the bubbles contain the technologies that were clustered together along all three axes. The technologies at the intersection of two bubbles can be understood as ‘borderline’ technologies, which were clustered with one group or the other, depending on the axis. For example, synthetic biology was grouped with cognitive and physical human enhancement technologies on the crisis stability and humanitarian principles axes, but with quantum for hardening & exploiting systems and quantum for C4ISR on the arms race stability axis. This is what designates it as a borderline technology. Borderline technologies are useful because they help to nuance our understanding of these technologies relative to one another. They also underscore the conceptual utility of technology clusters more broadly.

The most interesting insight from this exercise relates to Cluster 1. One might assume that in a cluster that includes ASAT capabilities, directed energy weapons, hypersonic weapon systems, and AI for information warfare, that the latter would be the borderline technology, on account of its non-kinetic character. However, this is not evident in the data. Instead, experts determined that the borderline cases in Cluster 1 are directed energy weapons and hypersonic weapon systems, which share features with Cluster 2 and Cluster 4, respectively. We can thus conclude that AI for information warfare and ASAT capabilities are the core technologies that constitute Cluster 1.

Combining technology impact assessment with technology forecasting

Once we had used ML to identify five clusters and their potential impact, we wanted to know when technologies’ impact might become most acute. To illustrate this, we combined the impact results from the scoring exercise with the technology readiness levels (TRL). The full report discusses how our experts assessed that most of the twelve technologies will be deployable in an operational environment by 2040. Thus, the more interesting years are 2021 and 2030. Figures 5, 6, and 7 show all twelve technologies in a three-dimensional impact graph, for 2021, 2030, and 2040 respectively. The three axes correspond to the three dependent variables. Each technology is represented by a dot, which comes in the colours of the respective cluster to which each technology belongs. The larger the dot, the higher the TRL for a given year.

Figure 5: Average scores (mean) for all technologies across all three dimensions, clustered into five groups, with relative TRL as bubble size (2021)

For 2021 (shown in Figure 5), the experts indicated that three high-impact technologies showed the highest level of deployability: ASAT capabilities, AI for information warfare, and AI for cyber operations. As mentioned above, experts explained the high level of deployability of ASAT capabilities in their qualitative comments by referring to the growing proliferation of state- and commercially owned satellites and the parallel military competition between the United States, Russia, and China, including in space. Regarding AI for information warfare, experts identified Russia as the main driver and anticipated that the United States will primarily invest in counter-influence AI measures to minimise the risks that information operations could pose. Further, experts clarified that there has been significant interest and corresponding investment in AI for cyber operations in the United States, Russia, and China, driven mainly by human limitations in this area and the potential of this technology to strengthen both offensive and defensive operations.

Figure 6: Average scores (mean) for all technologies across all three dimensions, clustered into five groups, with relative TRL as bubble size (2030)

According to experts’ estimates, by 2030 (shown in Figure 6), DEWs and hypersonic weapon systems will join the ranks of high-impact technologies with the highest level of deployability. In the qualitative comments, experts explained the high level of deployability of DEWs by reference to their potentially prominent role in missile defence and anti-satellite capabilities. They also referred to the defensive potential of this technology for countering rockets, artillery, mortars, hypersonic weapon systems, and (swarms of) unmanned aerial systems. Regarding hypersonic weapon systems, experts cited the exclusivity of this capability, the political prestige it confers, and perceived first-mover advantages as motivating the United States, Russia, and China to engage in an intensifying race to develop and field this technology. Experts added U.S. missile defence and military competition between China and the United States in the Indo-Pacific region as additional explanations.

Figure 7: Average scores (mean) for all technologies across all three dimensions, clustered into five groups, with relative TRL as bubble size (2040)

To what extent might these technologies strengthen and/or weaken stability and security?

Our study is driven by the assumption that technology is rarely entirely good or bad. Rather, the enabling technologies surveyed in this study have dual-use potential. While some technologies are inherently problematic, the effects of others are more contingent; they may be a challenge or an opportunity, depending on precisely how they are applied, under what conditions, and by whom. Thus, we phrased most of the questions in the scoring exercise in both positive and negative terms.

Thirty-two questions in the technology scoring exercise gave experts an opportunity to score both the extent to which these technologies are likely to strengthen or improve arms race stability, crisis stability, and humanitarian principles and the extent to which they are likely to weaken or deteriorate these same outcomes. The logic for including both options was that a given emerging technology could have both a positive and a negative impact, depending on its application.

In the data analysis stage, we evaluated the 32 questions that had an inverted option in more depth, to clarify two questions. First, to what extent might these twelve emerging technologies collectively strengthen or weaken international stability and human security? Second, how do individual technologies compare to each other (i.e., do certain technologies strengthen and weaken stability and security to equal or different degrees)?

Figure 8 endeavours to answer the first question: To what extent might these twelve emerging technologies collectively strengthen or weaken international security? It shows 24 dots in three-dimensional space. The 24 dots depict the twelve emerging technologies, each of which are represented by two dots: a green one that signifies the average score for questions that represent an opportunity to strengthen stability and security (hereafter referred to as ‘average strengthening score’), and a red one that signifies the average score for questions that represent a risk of weakening stability and security (hereafter referred to as ‘average weakening score’).

Figure 8: Average strengthening score (green) and average weakning score (red) for all technologies across all three dimensions 

Figure 8 indicates expert agreement that these twelve technologies will, in aggregate, weaken international stability and human security to a greater extent than they will strengthen international stability and human security. This can be ascertained from the red dots in the upper right corner of the three-dimensional graph.

Our next task was to compare the technologies themselves based on their expert-assessed strengthening or weakening effect. Figure 9 attempts to tackle the second question: Do all the technologies strengthen and weaken stability and security in equal measure or does the extent to which they strengthen or weaken stability and security vary?

In Figure 9, the y-axis shows the difference between the average strengthening and weakening scores for each technology. The taller the bar, the greater the difference between a given technology’s average strengthening and weakening scores. Red bars correspond to technologies that experts assessed as having a weakening effect that is greater than their strengthening effect across all three axes. Grey bars correspond to technologies for which the average strengthening score is greater than the average weakening score on at least one axis.

Figure 9: Distance between average strengthening score and average weakning score for all technologies, with technologies in red corresponding to those where the average weakning score is greater than the average strengthening score across all three dimensions, and technologies in grey corresponding to those where the average weakning score is greater than the average strengthening score across some—but not all—dimensions

Figure 10: Comparing average strengthening score and average weakning score for all technologies across all three dependent variables for the first three columns, with the colour green corresponding to a higher average strengthening score, and the colour red corresponding to a higher average weakening score. The fourth column (furthest right) represents the combined score, and contains the same information as Figure 9 above

Experts firstly agreed that all of the twelve technologies will weaken at a minimum one of the three dependent variables. Second, experts assessed that six of the technologies (i.e., AI weapons and effects, AI cyber operations, hypersonic weapon systems, directed energy weapons, AI information warfare, and ASAT capabilities) have weakening effects on all three dependent variables. Third, only two technologies (i.e., AI for C4ISR and quantum for C4ISR) show more of a strengthening than a weakening effect on two of three axes (shown in Figure 10).

It is important to be clear about what the data does and does not reveal. A technology with a higher rating of ‘mixed strengthening and weakening’ is not necessarily less problematic than a technology with a higher rating of ‘weakening’. There may be a technology with a far greater average weakening score than average strengthening score, but the properties of the technology that contribute to the weakening effect could be relatively easy to address, in full or in part. Conversely, there may be a technology with a greater average weakening score than average strengthening score, but the properties of the technology that contribute to the weakening effect could be severe and inherent to the technology itself.