June 2026 Summaries

The EU AI Act mandates that AI-generated content, including audio, text, images, and video, must be marked with machine-readable watermarks by August 2, 2026, to ensure traceability and compliance in European markets. The regulation stipulates that these watermarks should be robust, interoperable, and reliable enough to survive through normal content handling processes like compression, re-encoding, and platform uploads. There are two main approaches to embedding watermarks: model-level embedding at the point of generation and post-generation embedding. However, the complexity lies in balancing robustness with preserving content quality and ensuring interoperability across different systems. Additionally, the act distinguishes between watermarking, labeling, and disclosure, assigning distinct responsibilities to AI providers and deployers to ensure both technical and user-facing transparency. As teams prepare for compliance, they must document and test watermarking practices thoroughly, considering potential degradation points and ensuring that watermarks remain detectable after content undergoes typical processing and distribution.

Proteus is an automated framework developed by Resemble AI for testing the robustness of audio deepfake detectors against everyday audio transformations, which can affect the accuracy of such detectors. Introduced at ISSMAD 2026, Proteus systematically applies combinations of common audio processing techniques such as codec transcoding, added noise, and VoIP simulation to assess their impact on the detector's reliability. The study highlights a significant asymmetry where genuine audio is more easily misclassified as fake, a phenomenon that can be exploited by bad actors in the form of a "liar's dividend." Proteus tested over 17,000 audio transformation chains, with 72% being rejected due to quality degradation, but 4,847 chains successfully manipulated the detector's verdict while maintaining audio intelligibility and speaker identity. The findings, which are integrated back into the model's training data to improve its performance, underscore the security implications of audio deepfake technology and the challenges of maintaining detector accuracy in real-world conditions.

The Deepfake Watchlist for the week of June 19-25, 2026, highlights various incidents and developments in the realm of synthetic media, focusing on unauthorized cloning of voices and likenesses, legislative advancements, and the burgeoning issue of deepfake scams. A key incident involved the unauthorized cloning of the late voice actor Phil Sayer's voice, prompting discussions on intellectual property rights as the U.S. Senate Judiciary Committee advances the NO FAKES Act to establish federal rights over personal likenesses. Meanwhile, the Lovo case, a significant lawsuit addressing unauthorized voice cloning, collapsed due to the company's bankruptcy, highlighting the slow pace of legal remedies. Additionally, there is a surge in deepfake scams involving celebrity likenesses, while Cate Blanchett's nonprofit launched a Human Consent Registry to help individuals protect their identities from unauthorized AI use. These developments underscore the need for proactive measures such as marking and authorizing digital content, as well as the challenges of enforcing rights post-violation.

PerTh Multimodal is a comprehensive watermarking solution designed to address the increasing need for robust, imperceptible watermarking across audio, video, image, and text mediums, in response to regulatory requirements like the EU AI Act. Originally launched as an audio-only watermarker, PerTh has evolved to a multimodal platform with enhanced capabilities, achieving near-perfect accuracy in media watermark recovery. It embeds tamper-resistant signatures using modality-specific encoding techniques, such as psychoacoustic methods for audio and pixel-level modifications for images and videos, while linguistic rewriting is used for text watermarking. The new version rectifies limitations of the original PerTh audio model through a novel training attack system and curriculum design, incorporating diverse data augmentations to ensure robustness against real-world transformations. PerTh's integration capabilities are bolstered by ONNX support for edge deployment, ensuring interoperability and compliance with legal mandates, while offering detailed threshold settings to balance detection rates and false positives. The platform also complements metadata solutions like C2PA for dual-layer provenance, ensuring content integrity even after re-encoding or format changes. Despite its advancements, PerTh's explicit watermarking model requires further evaluation in non-speech audio and real-world settings to fully understand its potential and limitations.

Deepfake detection in forensics plays a crucial role in helping security and investigation teams assess the authenticity of audio, image, and video evidence for potential manipulation before they influence significant case decisions. Given the rise of AI-related scams, as illustrated by the FBI's 2025 Internet Crime Report, which recorded tens of thousands of complaints and substantial financial losses, the need for rigorous evidence verification has never been more critical. The forensic workflow involves examining the consistency of media files, their metadata, and their context within a case, and this process must be supported by human judgment and structured documentation. Resemble AI aids in this effort by offering a multimodal detection layer that provides explainable forensic context, helping teams understand anomalies and make informed decisions. By incorporating tools like DETECT-3B Omni and Resemble Intelligence, forensic teams can not only detect potential deepfake media but also document and align their findings with legal and compliance standards, ensuring that high-impact decisions are based on reliable and thoroughly reviewed evidence.

Resemble AI's Deepfake Watchlist for the week of June 5-11, 2026, highlights several alarming instances of synthetic media being used for fraud and harassment. Notably, deepfake videos featuring fabricated altercations between public figures like the Bank of England Governor and Nigel Farage were promoted as ads on platforms like X. This reflects a trend where scam networks leverage paid placements to bypass the need for organic credibility. In parallel, the UK government is pushing for Apple and Google to block nude images on children's phones, shifting child protection to the device level. Additionally, platforms like Reddit have seen deepfake BBC news segments used as ads for fake investment platforms, revealing a sophisticated structure of impersonation attacks. In a different context, Oklahoma's Sand Springs Public Schools shut down student email systems after AI-generated images of administrators circulated, demonstrating the challenges schools face in addressing AI-driven harassment. Lastly, a report by CrowdStrike links nearly half of state-backed intrusions into the US tech sector to North Korean operatives using AI-generated deepfakes to infiltrate companies. These developments underscore the growing need for more robust upstream enforcement and regulation to combat synthetic media threats effectively.

Chatterbox Multilingual v3 is a significant release in the evolution of multilingual text-to-speech models, supporting 25 languages and featuring embedded PerTh watermarking to meet regulatory requirements and enhance trustworthiness. This version retains the 0.5B Llama-based backbone of its predecessor but improves on speaker similarity, reduces hallucination rates, and enhances conversational naturalness through refined data training and language-specific specialization. The PerTh watermarking system, designed to be imperceptible yet robust against manipulations, aligns with upcoming regulations like the EU AI Act, ensuring audio provenance and reducing reliance on listener-level detection, which has become unreliable. The model's Character Error Rate (CER) evaluations highlight its strengths and weaknesses across different languages, with Italian and German performing exceptionally well, while Korean and Vietnamese require further data enhancements. The release includes general-purpose and Single-Language Pack models to address specific language demands, reflecting deployment insights that favor dedicated models for high-volume languages. Chatterbox v3 is optimized for enterprise-scale deployment via NVIDIA NIM, offering significant improvements in latency and throughput, with a focus on continuous enhancement in language support, subjective quality metrics, and watermarking capabilities.

Candidate fraud in remote hiring involves applicants misrepresenting their identity, credentials, or skills, often exacerbated by the use of AI technologies such as deepfakes and voice spoofing, as highlighted in the FBI's 2025 IC3 Annual Report. This type of fraud poses significant challenges, especially in roles with access to sensitive information, as it complicates verification processes across different stages of hiring, from application to post-hire. Effective detection requires a layered approach, integrating recruiter observations, structured interviews, work-sample verification, identity checks, and live meeting protection, with particular attention to roles that involve high access risk. Tools like Resemble AI can assist by providing live detection of synthetic audio and video during interviews, although they are not a substitute for comprehensive HR judgment and should be part of a broader verification strategy that maintains a balance between security and candidate experience. Ultimately, the goal is to implement a consistent and proportional verification process that identifies fraud without unfairly treating all candidates as suspicious, ensuring the integrity of the hiring process while safeguarding sensitive access.

The Deepfake Watchlist, curated by Resemble AI, provides a weekly analysis of synthetic media incidents that influence the news cycle, highlighting cases where AI-generated content is used to manipulate public perception and institutional trust. A key feature this week is an AI-generated video of Jose Mourinho falsely endorsing Florentino Pérez's Real Madrid presidential campaign, showcasing how synthetic media can fabricate consent in political contexts. The report also discusses various incidents, including a hack of Obama's White House Instagram account, widespread exposure to deepfakes before UK elections, and an increase in AI voice cloning scams resulting in significant financial losses. Additionally, Labour MP Jess Asato's legal action against xAI over unauthorized sexualized deepfakes signifies a shift towards litigating AI platform liability. The analysis underscores the need for proactive harm prevention, as the gap between detection and prevention remains a significant challenge, with legislative and regulatory frameworks struggling to keep up with the rapid evolution of AI-generated content.

TrueFoundry has integrated Resemble AI with its AI Gateway, enabling seamless voice cloning and text-to-speech (TTS) capabilities within the same infrastructure used for large language models (LLMs) and other AI traffic. This integration allows teams to utilize Resemble AI as a primary TTS provider through a native SDK pass-through, ensuring centralized authentication, access control, and cost tracking without altering client-side code. The gateway facilitates requests to Resemble's synthesis and streaming endpoints, maintaining the native API structure while offering routing and failover across multiple TTS providers through Virtual Models. The integration supports detailed observability and cost tracking, capturing input character counts and synthesis durations, and ensures that existing Resemble client code remains compatible by simply adjusting the base URL and authentication tokens. This architecture preserves Resemble's capabilities, such as Chatterbox model selection and audio precision controls, while TrueFoundry manages deployment, routing, and operational oversight within a unified AI governance framework.

In 2026, the rise of AI-generated media has led to concerns about content authenticity, prompting the need for effective watermarking solutions. Resemble AI and Steg.ai are two leading platforms addressing this issue, each offering distinct approaches. Resemble AI integrates watermarking during the content generation process, providing traceability and authenticity for audio, video, and images, and is particularly suited for enterprises focused on real-time multimodal AI content creation. In contrast, Steg.ai offers robust forensic watermarking across a variety of media, including images, videos, and documents, making it ideal for organizations requiring strong content protection and traceability after media distribution. Factors such as security, integration capabilities, and scalability are crucial when choosing between these platforms, as each caters to different stages of the content lifecycle. Resemble AI emphasizes embedding trust at the point of creation, while Steg.ai focuses on forensic protection and tracing of distributed media. Pricing models and licensing vary, with Resemble AI offering a flexible, usage-based approach, while Steg.ai provides a subscription model with enterprise options. The choice between these platforms depends on specific use cases, such as creating verifiable content in real-time or securing assets across distribution channels.

Deepfake interviews present significant challenges in the realm of remote hiring by potentially involving manipulated video, synthetic audio, proxy participation, or stolen identity details, making it difficult for hiring teams to ensure continuity between a candidate's profile and their live interview presence. Traditional interview checks often fall short in detecting these manipulations due to their focus on assessing answers and confidence rather than verifying the authenticity of live media. The increasing sophistication of deepfake technology necessitates a more integrated approach to risk management, involving structured interviews, live work walkthroughs, and enhanced identity checks, especially for roles with high access to sensitive systems. Tools like Resemble AI provide a layer of detection by reviewing audio and video in real-time, helping to identify potential deepfake risks before a candidate proceeds to onboarding or access approval. This comprehensive strategy aims to mitigate risks by examining patterns across the hiring process, ensuring that any inconsistencies are thoroughly vetted, and maintaining a clear documentation trail for any suspicious activities detected during interviews.