Quarterly Insights: Online Hate and Toxicity Trends (Q3 2025 Report)

This report is the fourth in a series providing a quarterly analysis of online harm trends. This edition examines patterns observed in July, August, and September 2025.

Data for July, August, and September was based on 49,8M messages in 27 languages, across 20 social media platforms: Reddit, X, 4chan, Gab, YouTube, Facebook, Thread and Instagram.

Toxicity over time 

This timeline illustrates the average toxicity level on social media in Q3 2025.

In July, the average ranged between 0.18 and 0.20, starting slightly higher and gradually decreasing to around 0.16-0.17 towards the end of the month. August showed similar stability, fluctuating within a narrow range between 0.17 and 0.19. In September, levels rose, reaching around 0.25 on 11-12 September before returning to approximately 0.20-0.21 by the end of the month. Overall, the period was characterised by relatively stable toxicity, with brief increases in early to mid-September.

Are you unsure about the terminology we use in this report? Visit our FAQ for more details.

Regional breakdown

This section underlines toxicity trends across four different European regions: Western Europe, Southern Europe, Northern Europe, and Eastern Europe. 

During this period, toxicity levels across European regions remained broadly stable, with moderate but consistent differences between regions.

• Western Europe showed the highest levels throughout, averaging around 0.20-0.22 in July and August, before rising to approximately 0.26-0.27 in mid-September. This gradual uptick marked the most notable change across regions during the quarter.

• Southern Europe followed a comparable but slightly lower trajectory, maintaining values between 0.16 and 0.20, with limited week-to-week variation and no pronounced spikes.

• Northern Europe continued to register the lowest toxicity levels, generally between 0.08 and 0.10, with only slight increases toward late September.

• Eastern Europe remained relatively stable, averaging 0.10-0.13, showing brief, minor increases in mid-July and again around mid-September.

Overall, regional differences persisted, with Western Europe recording the most sustained rise, while Northern and Eastern Europe remained comparatively stable and low.

VLOPs vs non-VLOPs

The following graph compares the average toxicity levels on Very Large Online Platforms (VLOPs) - including Facebook, X, YouTube, Instagram, and TikTok - with those on non-VLOPs such as 4chan, Gab, Reddit, and Thread, between April and June 2025.

Across all monitored platforms, toxicity levels remained broadly consistent with previous months, though clear differences persisted between VLOPs and non-VLOPs.

Non-VLOPs such as 4chan, WordPress, and Gab continued to register the highest toxicity levels overall. 4chan remained the most toxic environment, averaging between 0.40 and 0.44 throughout the period, with only minor day-to-day variation. WordPress followed closely, with average toxicity levels around 0.34-0.40, and several peaks above 0.42 in August. Gab showed slightly lower but still elevated values, maintaining between 0.27 and 0.33, with peaks up to 0.38 in late August and mid-September. Reddit’s toxicity scores generally ranged between 0.23 and 0.27, indicating moderate but steady levels across the observed period, though several data gaps, particularly from late July and much of September, limit the continuity of this assessment. In contrast, Telegram remained considerably less toxic within this group, with averages around 0.14-0.20.

Among VLOPs, toxicity levels were noticeably lower and more stable. Twitter (X) maintained moderate but consistent levels around 0.21-0.23, reaching an average of 0.27 in mid-September. YouTube averaged 0.13-0.15 between July and August, showing an increase of 0.10 score in the first half of September.

Toxicity levels on both Facebook and Instagram remained low and relatively steady throughout the period. Facebook averaged between 0.13 and 0.17, showing mild day-to-day variation but no major spikes. A slight increase was observed in mid-September, when values rose from around 0.14 to 0.17, though the change was temporary and returned to baseline by the end of the month. Instagram maintained consistently low toxicity, ranging between 0.12 and 0.14 for most of the period. Small fluctuations were visible late August, when values briefly rose to 0.16–0.17, before stabilising again. TikTok consistently displayed the lowest average toxicity of all monitored platforms, hovering between 0.11 and 0.13 throughout the period, with an increase of 0.10 score mid-September. 

Overall, non-VLOPs continued to present significantly higher toxicity scores than VLOPs, with 4chan and WordPress maintaining elevated toxicity. In contrast, platforms under the DSA’s VLOP designation demonstrated more consistent moderation patterns and generally stable, lower toxicity averages across the quarter.

Hate speech by category

The data reveal distinct thematic overlaps across the six monitored baselines, highlighting how hate narratives are shaped by intersecting categories of discourse and expression.

• Antisemitic content shows the strongest connection to racism (79.9%), followed by religion (44.4%) and politics (39.4%), illustrating how antisemitism is frequently embedded in ideological and religious framing. Threat-related language (19.3%) also appears with notable regularity, pointing to a confrontational tone.

• Anti-Muslim narratives are primarily driven by religious framing (75.1%) and racism (46.9%), reflecting a pattern similar to antisemitism but with a more pronounced religious focus. Political themes (26.6%) and threats (18.0%) are also prevalent, underlining the intersection between identity-based hostility and sociopolitical discourse.

• Anti-LGBTQ+ content is strongly associated with sexism (58.8%), indicating the centrality of gendered narratives. Additional overlaps with ridicule (12.5%) and obscenity (8.2%) suggest that mockery and demeaning language play a key role in perpetuating this form of toxicity.

• Sexism-related discourse exhibits substantial internal overlap with sexism (37.0%) itself, but also intersects with ridicule (10.6%), threats (11.7%), and obscenity (5.4%), indicating a combination of degrading and hostile tones. Political (13.8%) and religious (7.9%) elements appear less prominently but signal occasional moral and ideological framing.

• Anti-refugee narratives are dominated by political (45.8%) and racist (38.7%) content, reflecting the persistent association of migration discourse with identity-based polarisation. The presence of threat (15.8%) and contempt (5.7%) further underscores the exclusionary and antagonistic tone of such discussions.

• Anti-Roma discourse remains racist (37.6%), coupled with substantial political (10.4%) content.