In the third quarter of 2026, the Dutch Hub collected 19,053 online messages between July and September. The number of monitored messages increased significantly throughout the quarter, from 4,702 in July and 5,181 in August to 9,170 in September. September therefore accounted for almost half (48.1%) of all messages collected during the quarter. Compared to August, the number of messages increased by as much as 77%. This increase indicates a substantial rise in the volume of online discussions captured by the Hub's monitoring activities. The largest peak was recorded on 16 September. However, without further analysis of the messages’ content, this peak cannot be directly linked to a specific social or political event.
Despite the substantial increase in messages, the average toxicity score remained remarkably stable. Across all three months, the average toxicity score was 0.20, which is relatively low. The proportion of highly toxic messages, with a score above 0.8, also remained limited. Over this quarter, 45 messages fell into this category, representing 0.24% of the total. However, this does not mean that harmful expressions were largely absent. The analysis shows that online hate does not manifest exclusively through explicit threats or extremely hostile language, but also through more subtle forms of contempt, stereotyping, insults, and discriminatory statements. These less explicit forms of hostility can contribute to the normalisation of negative and hostile perceptions of particular population groups.
Comparing the individual months reveals interesting shifts in the nature of toxic expressions. In July, racism, profanity, political hostility, and sexism were prominent characteristics of the monitored messages. Notably, political content was also relatively frequently associated with threatening language. In August, the emphasis shifted toward expressions of contempt, identified in about 15% of messages, alongside ridicule, insults, and racism. A notable combination of political content and hostility towards Jewish people was also identified during this month. In September, contempt, political hostility, and offensive language remained prominent, while political and religious topics once again formed the most frequent combination. These developments illustrate how different forms of online hate intersect with broader social and political discussions.
Another area of concern is the presence of threatening language and explicitly antisemitic expressions. The proportion of messages containing threatening language was approximately 3% in July and August, increasing to 4% in September. Despite what appears to be a modest increase, threatening language remains a recurring feature of online discussions during periods where the average toxicity score remains stable. Furthermore, several explicitly antisemitic expressions stood out among the toxic keywords identified in September, including direct calls for violence against Jewish people and references to the Holocaust. The intersection of political discussions, religious topics, and antisemitic language deserves particular attention, as it may indicate how hostile stereotypes and more extreme expressions circulate within broader online discussions.
The third-quarter results therefore highlight an important distinction between the volume of online discussions, average toxicity levels, and the specific forms of hate expressed within them. While the number of monitored messages increased substantially in September, average toxicity remained unchanged. At the same time, the monthly analyses reveal shifts in the nature of hostile expressions, with political and religious topics, contempt, and antisemitism as recurring figures. This underscores the importance of monitoring approaches that go beyond average toxicity scores. By examining the content, context, and interconnectedness of harmful narratives, it becomes possible to better understand how online hate evolves and which societal tensions these discussions reflect.
Recurring Patterns in Toxic Bot Content
A comparison of the four word clouds representing toxic bot content shows that violence, political polarisation and identity-related issues consistently recur across the monitored messages. Words such as geweld (violence), dood (death), moord (murder), slachtoffers (victims) and vermoord (murdered) are prominent in almost all datasets. In addition, terms such as politiek (politics), links (left-wing), rechts (right-wing) and steekpartij (stabbing incident) appear frequently, indicating a strong overlap between violence-related language and political discussions.
The recurring presence of terms such as boeren (farmers), islamitische (Islamic), moslims (Muslims) and Nederlander (Dutch person) is also noteworthy. This suggests that national identity, religion and societal divisions are important themes within the monitored bot content. Although the emphasis varies somewhat between datasets, for example through the greater prominence of moslims (Muslims), genocide and islamisering (Islamisation) in the third word cloud, the overall pattern remains remarkably consistent.
The combination of violence-related terminology, political divisions and references to specific population groups illustrates how different polarising issues converge within this online environment. At the same time, caution is needed when interpreting these findings: the presence of words such as violence or murder does not automatically mean that messages glorify or encourage violence. Additional contextual analysis is necessary to make such determinations. The findings primarily provide a basis for further investigating the extent to which bot content contributes to the dissemination and amplification of hostile and polarising narratives.
Another interesting finding for the Hub's analysis is that the four word clouds not only reflect similar themes but also show a striking repetition of specific terms. In particular, boeren (farmers), islamitische (Islamic), geweld (violence) and slachtoffers (victims) are consistently prominent. This may point to a recurring thematic pattern within the monitored bot content.
Developments in the Nature of Toxic Bot Content
A comparison of the different toxicity categories reveals several shifts in the nature of the monitored toxic bot content during the third quarter of 2026. August stands out in particular due to higher scores in the categories of racism, religion and antisemitism. The score for racism increased from 0.35 in July to 0.40 in August, while the score for religion doubled from 0.15 to 0.30. The most notable development occurred in the category of antisemitism, where the score increased from 0.20 in July to 0.50 in August. In September, these scores declined again, reaching 0.35 for racism, 0.20 for religion and 0.25 for antisemitism.
These fluctuations are particularly interesting because the average toxicity score across all monitored messages remained stable at 0.20 throughout the quarter. This demonstrates that significant differences between specific forms of harmful content can exist beneath a relatively constant overall toxicity score.
The associated keywords also reveal interesting patterns. Within the racism category, terms such as islamitische (Islamic), moslims (Muslims) and dikke (fat) recur frequently, while in August, words such as islamitische (Islamic), jihad and Israël (Israel) were also prominent within the religion category. During the same month, the antisemitism category was particularly associated with the terms Hamas, Israël (Israel) and antisemitisme (antisemitism).
These findings point to the strong presence of religious and geopolitical topics within the classified bot content. Furthermore, terms such as omvolking (Great Replacement), genocide, elite and de overheid (the government) appear within the untruth category. This provides opportunities for further research into the presence of conspiracy theories and polarising narratives. However, keywords alone are insufficient to determine whether messages are genuinely antisemitic, racist or misleading; the context of individual messages must also be examined.
Other categories show a relatively stable pattern. Contempt remains an important form of toxicity throughout the quarter, with scores ranging from 0.30 to 0.35. Ridicule, sexism and threatening language also remain recurring elements. In particular, words such as boeren (farmers), hond (dog), belachelijk (ridiculous) and onzin (nonsense) frequently appear in the contempt and ridicule categories, while terms such as geweld (violence), slachtoffers (victims), moord (murder) and vermoord (murdered) recur within the threat category.
Notably, the score for threatening language remained unchanged at 0.20 across all three months, despite the substantial increase in the number of monitored messages in September.
Taken together, these results demonstrate that toxic bot content is not characterised exclusively by explicit threats or violent language, but also by recurring expressions of contempt, insults and political and identity-related divisions. The temporarily elevated scores for racism, religion and antisemitism in August deserve particular attention. These findings underline the importance of monitoring approaches that go beyond average toxicity scores and examine which forms of hostility become more prominent at particular moments.
Political Discussions, Contempt and Threatening Language Dominate Monitored Bot Content
The distribution of monitored messages across different toxicity categories shows that political topics played a central role in online discussions during the third quarter of 2026. Politics represented the largest category in all three months, with a clear increase in the absolute number of messages in September.
However, a large proportion of messages classified as political did not have an elevated toxicity score. This demonstrates that although political topics are strongly represented within the monitored bot content, political content is not necessarily toxic in itself. Its overlap with other categories, such as contempt, racism or threatening language, is therefore important for better understanding the nature of these discussions.
When focusing specifically on messages with elevated toxicity scores, a different picture emerges. Contempt and ridicule stand out in particular. In July and August, these categories contained relatively large numbers of messages with toxicity scores above 0.20. Both categories also remained clearly present in September.
This suggests that hostile bot content does not manifest exclusively through explicit threats or discriminatory statements, but also through insults, derogatory remarks and the ridicule of individuals or groups. Such forms of hostility can contribute to an online environment in which aggressive and polarising communication becomes increasingly commonplace.
Threatening language also remains a recurring feature of the monitored content. In all three months, messages were identified within the threat category, including both messages without elevated toxicity scores and messages scoring above 0.20. Although this category remains smaller in absolute terms than politics, it is clearly present alongside racism, religion and sexism.
Combined with the earlier keyword analysis, in which terms such as geweld (violence), dood (death), moord (murder) and slachtoffers (victims) appeared regularly, this underscores the importance of further contextual analysis of violence-related language. Not every reference to violence constitutes a threat or an incitement to violence.
The monthly comparison also shows that the increase in the total number of monitored messages in September was not accompanied by a proportional increase in highly toxic content. Instead, the graphs indicate that most messages within the individual categories did not have elevated toxicity scores, while only a small proportion of all monitored messages reached a score above 0.80.
Combined with the previously established stable average toxicity score of 0.20, this confirms that an increase in the volume of online content does not automatically mean that the average severity of toxicity increases.
Taken together, the findings show that political discussions provide an important context for the monitored bot content, while contempt, ridicule and threatening language are recurring characteristics of more toxic messages. These findings emphasise that monitoring online hate should not focus exclusively on the most extreme expressions, but should also consider less explicit forms of hostility circulating within broader political and societal discussions.
Conclusions
The monitoring results of the Dutch Hub for the third quarter of 2026 show that online hate and polarising expressions are recurring features of the monitored bot content. Although the number of collected messages increased substantially throughout the quarter, from 4,702 in July to 9,170 in September, the average toxicity score remained stable at 0.20. This demonstrates that an increase in the volume of online messages does not necessarily correspond to an increase in average toxicity. At the same time, the analyses reveal that important differences in the nature and severity of harmful expressions can exist beneath this stable overall score.
One notable finding is the central role of political and societal divisions. Political topics represented the largest category throughout the quarter, while contempt, ridicule and offensive language were recurring characteristics of more toxic messages. The results also show that violence-related terminology, racism and identity-related topics appeared regularly. In particular, recurring references to religion, national identity and political divisions stand out. This suggests that different forms of hostility intersect within the same online discussions and may contribute to the amplification of polarising narratives.
The monthly comparison further demonstrates that specific forms of toxicity can fluctuate considerably, even when average toxicity remains stable. August stands out in particular due to higher scores for racism, religion and antisemitism. The recurring presence of terms such as islamitische (Islamic), moslims (Muslims), Israël (Israel), Hamas and genocide highlights the importance of further examining how religious and geopolitical topics are discussed. At the same time, keywords and automated classifications alone do not provide sufficient evidence to establish whether discriminatory or hateful narratives are actually present. A qualitative analysis of the context remains necessary.
The main conclusion is that developments in online hate cannot be assessed solely through the number of messages or average toxicity scores. The recurring presence of contempt, political hostility and identity-related divisions deserves attention, even when such expressions are not classified as extremely toxic. The findings therefore emphasise the importance of continuous monitoring that examines both quantitative developments and the content and context of online discussions. For the Dutch Hub, the added value lies not only in identifying explicit hate speech and threats, but also in uncovering recurring patterns that may contribute to the normalisation of hostile and polarising communication.