Advancing AI: Highlights from November

November was a pivotal chapter for Sony AI, marked by the debut of our ethics project FHIBE, a consent-driven benchmark years in the making, and a trio of blogs that traced its foundations, its construction, and the people who brought it to life. Together, these stories highlight how ethical data practices, technical rigor, and human-centered design can reshape AI’s relationship with the world.

Below is a look back at the work that defined the month.

What If Fairness Started at the Dataset Level?

We opened the month by examining a key question: what would change if fairness began with the data itself? This piece lays the philosophical and scientific groundwork for FHIBE by tracing years of Sony AI research into data collection practices, representational gaps, and the risks of building models on scraped or unbalanced datasets.

It highlights practical tools developed through this work — such as Targeted Augmentations for Bias Mitigation (TAB) — and introduces structured frameworks for defining and measuring diversity with more rigor.

Read the full blog here: https://ai.sony/blog/What-If-Fairness-Started-at-the-Dataset-Level/

Introducing FHIBE: A Consent-Driven Benchmark for AI Fairness Evaluation

Shortly after, we shared FHIBE, a new benchmark published in Nature and created to help developers uncover bias before deployment. With 10,000+ images collected through consent, fair pay, and global representation, FHIBE fills a long-standing gap: an ethical, technically robust dataset designed specifically for fairness evaluation.

The benchmark includes detailed demographic metadata, environmental conditions, and precise annotations supporting tasks like pose estimation, face detection, and vision-language audits across foundation models.
FHIBE shows what becomes possible when fairness is treated as a design requirement rather than a post-training fix.

Read the full story here: https://ai.sony/blog/Introducing-FHIBE-A-Consent-Driven-Benchmark-for-AI-Fairness-Evaluation/

And to watch our short film, A Fair Reflection, visit: FairnessBenchmark.ai.sony

The FHIBE Team: Data, Dignity, and the People Who Made It Possible

We closed the month with a deep look at the people behind FHIBE. This story follows the three-year journey of building the benchmark, and the decision-making, infrastructure, and interdisciplinary collaboration it required.

It traces how revocable consent was introduced, how compensation policies were tied to documentation, how privacy safeguards were designed, and how the team redefined representation at scale through 1,234 intersectional identity groups.

The piece also highlights the operational reality of the work: vendor coordination, QA processes, custom engineering tools, legal considerations, and the constant negotiation between utility and privacy.

Together, these perspectives show how FHIBE became not just a dataset, but a model for long-term stewardship.

Meet the Team, dive into the full interview: https://ai.sony/blog/The-FHIBE-Team-Data-Dignity-and-the-People-Who-Made-It-Possible/

November was a milestone month for FHIBE. Following its publication in Nature, the benchmark quickly drew global attention, sparking more than 25 articles across 15 countries. The coverage showcased the new benchmark tool and the importance of improving the data foundations of modern AI. Some coverage included:

Nature: Nature published the FHIBE research alongside an editorial feature – a distinction reserved for select studies – that examined the research behind FHIBE and emphasized the importance of the AI industry collaborating to tackle widespread data collection issues. It stated, “[Alice] Xiang and her research team have shown how to produce and test responsible AI systems. They have chosen a tough problem, but this should not be their fight alone. Others must join the effort so we can build AI applications according to the highest standards of accuracy and ethics.” Read the article here.

The Register: Thomas Claburn interviewed Alice Xiang about the importance of FHIBE and the impact she hopes it will have on the AI industry. She noted, “At a time when data nihilism is increasingly common in AI, FHIBE strives to raise the standards for ethical data collection across the industry. FHIBE doesn't fully solve this problem since there's still the scalability issue (FHIBE is a small evaluation dataset, not a large training dataset), but one of our goals was to inspire the R&D community and industry to invest more care and funding into ethical data curation. This is an incredibly important problem – arguably one of the biggest problems in AI now – but far less attention is paid to innovation on the data layer compared to the algorithmic layer." Read the article here.

Engadget: Will Shanklin provided readers with an in-depth overview of FHIBE in Engadget’s feature coverage, commenting, “[FHIBE] tests the degree to which today's AI models treat people fairly. Spoiler: Sony didn't find a single dataset from any company that fully met its benchmarks.” Read the article here.

To close the month, researcher Chieh-Hsin Lai announced a new monograph, The Principles of Diffusion Models: From Origins to Advances, co-authored with Yang Song, Dongjun Kim, Yuki Mitsufuji, and Stefano Ermon. The work offers a clear, systematic walkthrough of diffusion modeling. It unravels how the early formulations emerged, how today’s techniques achieve controllability and speed, and the ideas shaping the next generation of models.

The monograph traces the core concepts that helped establish diffusion methods and explains the reasoning behind the formulations now common in modern generative modeling. It’s designed to serve as an accessible, reliable reference for both newcomers and advanced practitioners who want a unified view of the field’s foundations.

*The authors have also noted that a markdown version will be released for deeper discussion and community feedback. Follow Sony AI Researcher, Chieh-Hsin (Jesse) Lai , at x.com/JCJesseLai for updates.

Access the monograph: The Principles of Diffusion Models

NeurIPS 2025

In December, Sony AI will join the global research community at NeurIPS 2025 in San Diego and Mexico City. Our teams will present new work spanning music evaluation, text-to-image modeling, diffusion, and audio compression, reflecting ongoing progress across several flagship areas.

This year’s contributions explore how models learn from creative data, how training signals shape generative behavior, and how evaluation methods can better reflect real-world use. A full roundup of our NeurIPS papers and workshops will be shared next month.

For now, please rediscover our contributions to NeurIPS in our prior year’s round ups:

- From Data Fairness to 3D Image Generation: Sony AI at NeurIPS 2024

- Exploring the Challenges of Fair Dataset Curation: Insights from NeurIPS 2024 – Sony AI

Gran Turismo 7 Power Pack DLC Announced

In November, Sony Interactive Entertainment announced the upcoming Gran Turismo 7 Power Pack DLC, arriving December 4 on PS5. The update brings several new racing experiences to players, including the return of 24-hour endurance races, 50 new challenges across 20 themed categories, and full race-weekend formats that introduce practice, qualifying, and main events. It also debuts Gran Turismo Sophy 3.0, offering more realistic, competitive racing behavior in a wide range of scenarios.

The DLC includes 5,000,000 in-game credits and a dedicated game mode designed for players who want a more technical, motorsport-driven experience.
A separate Spec III update with new tracks, cars, and features will follow later in December.

Read the full announcement here:
Gran Turismo 7 Power Pack DLC coming December 4 – PlayStation.Blog

Yuki Mitsufuji, Lead Research Scientist and Vice President of AI Research at Sony AI, presented his work on multimodal generative AI frameworks for creators at the SANE 2025 workshop at Google, New York, NY, on November 7, 2025.

More info on the SANE workshop series:
http://www.saneworkshop.org/
To watch Mitsufuji’s Talk, visit: https://youtu.be/HAQeX7AMt3k?si=8PTcHDuJuYYXgTVY

About:

“This talk explores how cutting-edge generative AI is transforming creative workflows in music, cinema, and gaming. Led by Dr. Yuki Mitsufuji, the Music Foundation Model Team at Sony AI has developed multimodal frameworks such as MMAudio, which generate high-quality, synchronized audio from video and text inputs. Their research, recognized at top venues like NeurIPS, ICLR, and CVPR, has contributed to both content creation and protection, with practical demos integrated into commercial products. The session will highlight key innovations, including sound restoration projects and the future of AI-powered media production.”

As November wraps, preparations continue for NeurIPS 2025, where our researchers will present new contributions across generative modeling, multimodal evaluation, and responsible AI research. A full roundup will arrive in December, alongside additional FHIBE resources and updates.

Thank you for following along. November underscored a clear message: fairness isn’t a step in the pipeline: it’s a practice shaped by people, process, and purpose. FHIBE stands as a reminder of what careful, collaborative dataset building can achieve.

Connect with us on LinkedIn, Instagram, or X, and let us know what you’d like to see in future editions. Until next month, keep imagining the possibilities with Sony AI.