Over the past year, artificial intelligence (AI) has become increasingly widespread across the globe, revealing a growing potential for adverse consequences linked to its use. According to the AI Incident Database, a crowd-sourced repository aggregating media accounts of AI failures and mishaps, reports of AI-related incidents increased by 50% from 2022 through 2024. This upward trajectory continued into 2025 with reported incidents surpassing the entire previous year’s total within the first ten months alone. These incidents encompass a diverse range of issues, including scams facilitated by deepfake technology and instances of disordered thinking provoked by conversational AI chatbots.
Daniel Atherton, an editor at the AI Incident Database, asserts that AI is already creating tangible harms in the real world, stressing the importance of monitoring failures to enable corrective actions. The database compiles information primarily through the collection of news articles documenting AI-related events, consolidating multiple reports on the same occurrence into single entries to provide clarity.
While acknowledging the limitations inherent in crowd-sourced media data, Atherton notes that increased media attention partly contributes to the observed rise in incident reporting. However, he emphasizes that news coverage remains currently the most accessible public source detailing AI-associated harms. Despite this, only a portion of actual incidents receive journalistic focus, and among those, not all are captured in the database. He highlights that global reporting and documentation represent only a fraction of the real experiences of individuals encountering AI-induced harms.
In regulatory terms, legislation such as the European Union's AI Act and California’s Transparency in Frontier AI Act (SB 53) mandate that developers report certain AI incidents to authorities, yet the reporting threshold typically encompasses only the most severe or safety-critical events.
Challenges in Understanding Incident Patterns
Artificial intelligence encompasses a broad range of technologies including autonomous vehicles, chatbots, and content moderation systems. The AI Incident Database aggregates all incidents under this umbrella without detailed categorization, complicating efforts to discern overarching patterns within the dataset. Simon Mylius, an affiliate researcher at MIT FutureTech, underscores this difficulty and has contributed towards developing a tool that leverages natural language models to analyze news reports linked to each incident. This tool categorizes incidents by the type and severity of harm.
Although this AI-assisted classification framework is still undergoing validation, it aims to assist policymakers in managing vast volumes of incident reports and in identifying emerging trends. Recognizing that media reports often contain 'noise,' Mylius’s team is designing analytical methods adapted from disease surveillance to better interpret the data. Their goal is to provide regulators with more reliable insights and enable proactive responses to AI-related harms ahead of major crises akin to those experienced in social media governance.
Differential Trends Among AI Harms
An application of this taxonomy reveals that the increase in AI-related incidents is not uniform across all categories. For example, in 2025, reports of AI-generated misinformation and discrimination declined, whereas incidents associated with ‘computer human interaction’ — encompassing issues like the psychological impacts related to ChatGPT usage — increased. The most notable growth pertains to the utilization of AI by malicious actors, primarily for perpetrating scams or spreading false information, with such reports amplifying eightfold since 2022.
Prior to 2023, the systems most commonly implicated included autonomous vehicles, facial recognition technologies, and content moderation algorithms. More recently, however, deepfake videos have emerged as the predominant source of reported incidents, surpassing those three categories combined. Notably absent from these figures are deepfakes generated subsequent to a December update to xAI’s Grok chatbot. This update enabled widespread sexualized image production targeting real women and minors, with rates estimated at 6,700 images per hour. The severity of the issue incited governmental responses from Malaysia and Indonesia, which blocked the chatbot, and prompted investigations by the U.K. media watchdog. British officials have also outlined plans to enact legislation criminalizing the generation of non-consensual sexualized imagery, explicitly referencing Grok’s activities.
Following public backlash, xAI restricted Grok’s image generation capabilities to paying subscribers and implemented content filters barring the manipulation of images involving real individuals in revealing attire.
Technological Advances and Emerging Security Concerns
The surge in deepfake-related incidents coincides with rapid improvements in the quality and accessibility of such synthetic media. This evolution illustrates that while some AI-related failures are due to system limitations — for example, autonomous vehicles failing to recognize cyclists — other harms stem from advancements in AI technology itself. The continued progress of AI, especially in areas like code generation, is likely to introduce new risks.
In November, Anthropic reported intercepting a significant cyberattack exploiting their Claude Code assistant, emphasizing an inflection point where AI tools serve both defensive and offensive roles within cybersecurity domains. Mylius predicts a surge in cyberattacks facilitated by AI, expecting substantial financial losses resulting from these sophisticated threats in the near future.
Accountability and Industry Responses
Given their prominent market positions, the largest AI companies figure most frequently in incident reports. Nonetheless, over one-third of incidents since 2023 have involved AI developers that are unidentified. Atherton highlights the complexity of attribution in scams circulated on platforms like Facebook and Instagram; while Meta is implicated as the platform provider, the specific AI tools used to create scams often remain unreported.
In 2024, Reuters disclosed that Meta anticipated approximately 10% of its advertising revenue would be derived from scam and banned goods ads. Meta described this estimate as overly inclusive and part of an internal effort to tackle such fraudulent content, cautioning that the disclosures present a selective narrative that misrepresents its approach.
Efforts to enhance AI accountability have garnered support from major industry players. For instance, 'Content Credentials,' a system embedding watermarks and metadata to verify content authenticity and flag AI-generated media, is endorsed by companies including Google, Microsoft, OpenAI, Meta, and ElevenLabs. ElevenLabs also offers a tool purported to detect AI-generated audio samples created with its technology. However, Midjourney, a widely used image generation platform, does not currently participate in this emerging standard.
Ongoing Vigilance and Broader Concerns
Atherton cautions against allowing present-day AI harms to be regarded as 'background noise,' underscoring the necessity to maintain vigilance. Mylius concurs, remarking that some negative impacts surface abruptly, whereas others develop incrementally and are less perceptible on an individual basis. Issues such as societal disruption, privacy erosion, and the amplification of misinformation may not be immediately visible in singular incidents but cumulatively result in significant damage.