AI-Driven Harm Reports Surge 50% as Deepfake Incidents Eclipse Traditional Failures

The landscape of artificial intelligence safety is undergoing a dramatic shift. According to a new analysis of data from the AI Incident Database (AIID), reported incidents of AI-related harm have surged by 50% year-over-year from 2022 to 2024. Even more concerning, data from the first ten months of 2025 indicates that the current year has already surpassed 2024’s total, signaling an accelerating trend rather than a temporary spike.

For industry observers and stakeholders at Creati.ai, this data confirms a pivotal transition in the nature of technological risk. Where AI incidents were once dominated by autonomous vehicle errors or algorithmic bias in static systems, the proliferation of generative AI has ushered in a new era defined by deepfakes, malicious use, and synthetic media scams.

The Shift from Accidents to Malicious Use

The most striking finding in the recent data is the fundamental change in how AI is causing harm. Between 2018 and 2022, the primary drivers of reported incidents were often system limitations—self-driving cars failing to detect cyclists or facial recognition systems exhibiting racial bias. However, the release of powerful generative models has flipped this dynamic.

Since 2023, incidents linked specifically to deepfake video have outnumbered reports related to autonomous vehicles, facial recognition, and content moderation algorithms combined. This marks a transition from "AI accidents" (where the system fails to perform as intended) to "malicious use" (where the system performs exactly as intended, but for harmful purposes).

Key trends identified in the analysis include:

• Malicious Actors: Reports of individuals using AI to scam victims or spread disinformation have grown 8-fold since 2022.
• Generative Dominance: Deepfake videos and image generation issues now dominate the incident logs, driven by the increasing accessibility of tools like Midjourney, OpenAI’s Sora, and xAI’s Grok.
• Cybersecurity Risks: New threats are emerging in sensitive domains, such as the interception of coding assistants for cyber-attacks.

Analyzing the Incident Data

To understand the scale of this issue, it is essential to look at the raw numbers provided by the AI Incident Database and researchers at MIT FutureTech. The trajectory shows an exponential rise in reported harms corresponding with the mainstream release of Large Language Models (LLMs).

Yearly Reported AI Incidents (2020-2024)

Data Source: AI Incident Database / MIT AI Incident Tracker

Daniel Atherton, an editor at the AI Incident Database, emphasizes that these numbers are likely just the tip of the iceberg. "AI is already causing real-world harm," Atherton notes. "Without tracking failures, we can't fix them." He warns that while crowd-sourced data has limitations, it currently remains one of the few viable windows into the scale of the problem, as corporate reporting remains fragmented.

The "Unknown Developer" Problem

One of the most complex challenges for regulators and safety researchers is attribution. While major tech giants are frequently cited in reports due to their high visibility, a significant portion of AI harm is generated by tools where the underlying developer is unidentified.

Since 2023, more than one-third of all reported incidents involved an "Unknown" AI developer. This often occurs in the context of social media scams, where a user encounters a deepfake advertisement or a fraudulent investment scheme on platforms like Facebook or Instagram, but the specific tool used to create the synthetic media cannot be determined.

Simon Mylius, an affiliate researcher at MIT FutureTech, points out that this creates significant "noise" in the data. To combat this, his team has deployed LLMs to parse news reports and classify incidents more accurately. This deeper analysis reveals that while some categories like "AI-generated discrimination" showed a relative decrease in 2025, incidents of "Computer-Human Interaction"—such as users developing unhealthy attachments to chatbots or experiencing "psychosis" driven by hallucinating models—are on the rise.

Case Study: The Speed of Harm

The volatility of the current landscape was starkly illustrated by a recent incident involving xAI's Grok. Following a software update, the model was reportedly used to generate non-consensual sexualized images of real people at a rate estimated by some researchers to be 6,700 images per hour.

This incident prompted immediate regulatory backlash, including blocks by the governments of Malaysia and Indonesia and an investigation by the UK's media watchdog. It serves as a prime example of how "technical advances" can instantly translate into "scaled harm" if safety guardrails are not rigorously tested prior to deployment. xAI subsequently limited image generation tools to paying subscribers and implemented stricter blocking for real-person imagery, but the incident highlights the reactive nature of current safety protocols.

Regulatory and Industry Response

The surge in reports has validated the urgency behind recent legislative moves, such as the EU AI Act and California's Transparency in Frontier AI Act (SB 53). These laws mandate that developers report safety-critical incidents, theoretically reducing the reliance on media reports for data.

However, the industry is also attempting self-regulation through technical standards. The Content Credentials initiative—a system for watermarking and embedding metadata to verify content authenticity—has garnered support from heavyweights like:

• Google
• Microsoft
• OpenAI
• Meta
• ElevenLabs

Notably, popular image generator Midjourney has yet to fully adopt this emerging standard, leaving a gap in the ecosystem.

A Call for Systematic Vigilance

For Creati.ai, the 50% surge in incident reports is a clarion call. It suggests that as AI models become more capable, the "attack surface" for potential harm expands. Anthropic recently revealed it had intercepted a large-scale cyber-attack attempting to utilize its Claude Code assistant, leading the company to declare that the industry has reached an "inflection point" regarding AI in cybersecurity.

The data from the AI Incident Database proves that AI harm is no longer hypothetical or rare. It is becoming a measurable, growing component of the digital economy. As Simon Mylius notes, we must be careful not to let these incidents become "part of the background noise." Whether it is the sudden crisis of a deepfake wave or the gradual erosion of trust through misinformation, the tracking and analysis of these failures is the only path toward a safer AI future.