SonyAI

What happens when interaction with software starts to feel like a relationship? AI companions are increasingly part of people’s social lives. Some users treat conversational agents as confidants. Others describe them as friends or even romantic partners. A new research paper by W. Bradley Knox, Katie Bradford, Samanta Varela Castro, Desmond C. Ong, Sean Williams, Jacob Romanov, Carly Nations, Peter Stone, and Samuel Baker explores how the design traits of AI companions may shape both benefits and potential harms. Rather than focusing only on individual incidents, the paper proposes a framework for understanding how system behavior and product design choices can influence human–AI relationships. For this Sony AI blog feature, we spoke with W. Bradley Knox and Peter Stone, two of the paper’s authors, about the motivation behind the research and why clearer language is needed to study AI companionship. 👉 Read the full interview and research overview on the Sony AI blog: https://bit.ly/41WYEeE

What does it take to make sound and image feel like they belong together? In episode 2 of the Research Vault, Sony AI researchers Akio Hayakawa / 早川 顕生, Masato Ishii, Takashi Shibuya, and Yuki Mitsufuji, PhD walk through MMDisCo — a new approach to joint audio and video generation that rethinks how multimodal AI systems can be built. The core idea: instead of training a single large model to handle both modalities, MMDisCo introduces a lightweight coordinating network — a multi-modal discriminator — that connects two independently trained models and guides them to generate outputs that are semantically and temporally aligned. The result is synchronized audio and video generation without the cost of retraining from scratch. In this deep dive, the team unpacks the architecture, the reasoning behind key design decisions, and what this means for the future of multimodal generation. Watch episode 2 of the Research Vault → https://lnkd.in/gmE-YxYe Read the paper → https://lnkd.in/g4vxXBqd #SonyAI #AIResearch #MachineLearning #Multimodal #DiffusionModels #GenerativeAI #ResearchVault

March at Sony AI brought research, publications, and live conversations together in one busy month. Chieh-Hsin (Jesse) Lai joined us for a Q&A on The Principles of Diffusion Models — a new book co-authored with Yang Song, Dongjun Kim, and Stefano Ermon that traces the shared mathematical foundations across diffusion approaches. A must-read for anyone navigating this fast-moving field. Looking ahead to May: Sony AI will present 11 accepted papers at IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP) 2026 in Barcelona, spanning music structure analysis, audio-visual generation, source separation, and more. Read the full March recap on the Sony AI blog: https://bit.ly/4c4oBxO #SonyAI #AIResearch #MachineLearning #DiffusionModels #ICASSP2026 #Robotics #ReinforcementLearning

Localized AI image edits are harder to detect than full synthesis. SHIELD benchmarks how well vision-language models flag partial edits in a zero-shot setting. 👉 See the findings: https://bit.ly/4dRuHE3

I have been selected to join the Technical Exchange Network (TEN) of the AI Infrastructure Interchange (AIII), an initiative by the World Intellectual Property Organization – WIPO. While most of the current debate around generative AI IP and ethics is (rightfully) approached from a legal/policy angle, the technical/infrastructural side is overlooked. However, "solving" genAI also needs addressing technical challenges: tracing provenance, calculating attribution, implementing labelling, etc. The TEN brings together a small group of technical experts who will meet regularly to share knowledge and discuss these challenges. Looking forward to contributing my experience and engaging with colleagues across sectors to further discuss these challenges and explore solutions that might hopefully inspire future work in the field. Also, if you are interested in joining, I understand that we will welcome new members on a rolling basis - more info here. https://lnkd.in/ep3HQPXv

Protecting image datasets remains a challenge. EnTruth introduces a minimal, robust method to trace unauthorized use in text-to-image diffusion models without degrading image quality. 👉 Explore the paper: https://bit.ly/4sXLlpO

Alice Xiang has been recognised in AI Magazine's Top 100 Women in AI 2026. As Sony Group's Global Head of AI Governance and Lead Research Scientist at Sony AI, her work sits at the intersection of ethics, data, and accountability. FHIBE is a case in point. The Fair Human-Centric Image Benchmark is the first publicly available, consent-driven, globally diverse dataset for evaluating bias in human-centric computer vision. It was published in Nature Magazine. It's free to use. And it's already being adopted across industry. Bias in AI doesn't fix itself. Alice is building the tools to find it. Congratulations to Alice and all the leaders recognised this year. Discover the full list: https://bit.ly/4c0yHkm #Top100WomenInAI #WomenInAI #AIEthics #ResponsibleAI #SonyAI #AIResearch #Leadership #ArtificialIntelligence

Fine-tuning large models is powerful—but costly. StelLA introduces a structured low-rank adaptation method that keeps subspaces orthonormal, improving performance across language and vision tasks without increasing compute. 👉 Read the research paper: https://bit.ly/46PkK62

Alice Xiang, Global Head of AI Governance at Sony and Head of AI Ethics at Sony AI, will be featured on the MIT Sloan Management Review podcast "Me, Myself, and AI," airing March 10 at 8 a.m. ET. In the episode, Alice discusses one of the central challenges in responsible AI development: the lack of ethically sourced datasets for evaluating bias in AI systems. The conversation explores how Sony AI’s Fair Human-Centric Image Benchmark (FHIBE) was created to help address that gap, enabling practitioners to assess fairness in human-centric computer vision models using consent-driven, diverse data. Tune in to hear how improved benchmarks and data practices can support more trustworthy AI development across industry. Listen starting March 10: https://lnkd.in/djQqArmD Visit FairnessBenchmark.ai.sony  to learn more about FHIBE. #ResponsibleAI #AIGovernance #AIethics #SonyAI

Diffusion models are now a foundational tool in generative AI, but the field can be difficult to navigate once acronyms, objectives, and frameworks begin to multiply. In this Q&A, Chieh-Hsin Jesse Lai, Research Scientist at Sony AI, discusses The Principles of Diffusion Models — a newly released book he co-authored to help researchers and practitioners understand what actually stays constant as the field evolves. The conversation explores why the book was written, how it unifies DDPMs, score-based models, and flow-based methods under shared mathematical principles, and why focusing on fundamentals — not just results — helps diffusion research remain useful over time. At a technical level, these methods share a common backbone: learning a time-dependent velocity field that smoothly transports a simple noise distribution into real data. From this perspective, sampling amounts to solving a differential equation along a continuous trajectory. Building on this view, the book covers controllable generation, improved numerical solvers, and diffusion-inspired flow-map models such as Consistency Models and related approaches that learn larger time jumps for faster generation. If you work with generative models or want a clearer conceptual map of diffusion methods, this interview is a thoughtful and practical read. 🔗👉 https://bit.ly/40eOeq4 #DiffusionModels #GenerativeAI #AIResearch #MachineLearning #DeepLearning #AppliedMath #ProbabilisticModels