Sony AI is a research organization that Sony says was established in April 2020, and its two best-documented outputs are Gran Turismo Sophy, a racing agent described in the journal Nature on February 9, 2022, and FHIBE, a consent-based image dataset for testing bias in computer vision, published in Nature on November 5, 2025. Both are peer-reviewed results with stated limits, so they are a firmer basis for describing the lab than product announcements or press coverage.
Key takeaways
- Sony AI lists six research areas on its website: gaming and interactive agents, imaging and sensing, AI ethics, robotics, scientific discovery, and AI for creators.
- Gran Turismo Sophy used model-free deep reinforcement learning and beat four of the world's top Gran Turismo drivers in a head-to-head event, according to the 2022 Nature paper.
- A racing agent for one video game is evidence about that game, not about driving on public roads.
- FHIBE holds 10,318 images of 1,981 people from 81 countries and areas, and the authors report that it confirmed known biases in existing models.
- The FHIBE authors state their own limits: high collection cost, less visual variety than scraped web images, and results that can differ between datasets.
What the organization says it is
Sony AI's website states that it was established in April 2020 to pursue research in artificial intelligence, with the aim of AI that works in harmony with people. It names six flagship areas: gaming and interactive agents, imaging and sensing, AI ethics, robotics, scientific discovery, and AI for creators. The Nature paper on Gran Turismo Sophy lists all of its authors as working at Sony AI, with teams in New York, Tokyo, and Zurich.
The site does not publish staffing, budget, or revenue figures, and none are given here. Descriptions of the lab's size or commercial impact that appear elsewhere could not be traced to a primary source.
Gran Turismo Sophy: what the paper reports
The paper, titled "Outracing champion Gran Turismo drivers with deep reinforcement learning," appeared in Nature volume 602 on February 9, 2022. It describes an agent called Gran Turismo Sophy, trained with model-free deep reinforcement learning, which means the agent improves by trial and error against a reward signal rather than by copying recorded laps. The authors also used what they call mixed-scenario training, and they wrote the reward function so that the agent stays fast while following the loosely defined sportsmanship norms of racing. The abstract reports that the agent beat four of the world's top Gran Turismo drivers in a head-to-head competition.
The setting is a racing simulation. The abstract does not name the human drivers or the cars, and the extended data captions mention the tracks Seaside, Sarthe, and Maggiore. Results in a game with a fixed physics model say little about how the same method would behave with sensor noise, weather, or pedestrians.
From research paper to a playable mode
On February 20, 2023, Polyphony Digital and Sony AI announced Race Together, an event inside Gran Turismo 7 on PlayStation 5 and described as the first global release of the agent. Players needed Collector Level 6 and the 1.29 update. The event offered four races, each at Beginner, Intermediate, and Expert levels, with every race placing the player against four Gran Turismo Sophy cars of different strength, plus a one-versus-one mode with identical car setups. At announcement it ran until late March 2023. Any later availability was not confirmed for this page.
FHIBE: a dataset built for testing bias
The second paper, "Fair human-centric image dataset for ethical AI benchmarking," appeared in Nature volume 648 on November 5, 2025. It introduces the Fair Human-Centric Image Benchmark (FHIBE), meant for checking whether computer vision models perform unevenly across groups of people. Many earlier datasets were scraped from the web without the consent of the people pictured. FHIBE instead collected images from paid participants who also reported their own age, pronouns, ancestry, and skin tone.
| Item | What the paper reports |
|---|---|
| Images and people | 10,318 images of 1,981 individuals, including 1,711 primary subjects who submitted pictures of themselves |
| Geography | 81 countries and areas across five regions, with images taken between May 2011 and January 2024 |
| Annotations | Self-reported attributes, bounding boxes, 33 keypoints, and 28 segmentation categories |
| Models tested | Eight narrow vision tasks and two vision-language models, CLIP and BLIP-2 |
| Cost | About USD 308,500 for images plus roughly USD 450,000 in fixed costs |
What the FHIBE authors found, and what they caution
The authors confirmed previously documented biases, such as face detection working better on images labeled with she/her/hers pronouns than he/him/his, and reported new ones, including weaker face parsing for older people. They also found stereotyped outputs from BLIP-2, with toxic responses more common for some ancestry and skin-tone groups. The largest gaps appeared for intersectional groups, meaning combinations of attributes.
The limits are stated in the paper. Consent-based images show less visual variety than web-scraped ones, because subjects are more often centered and closer to the camera. Paying participants can attract fraudulent submissions, and confirming that a submitter is the person pictured is hard. Some sensitive attributes are reported only in aggregate, and bias results can conflict between datasets, so the authors recommend case-by-case testing. Access to the data requires agreeing to terms of use.
Questions to ask about any lab's claim
- Is the result peer-reviewed, and does the abstract say the same thing as the press summary?
- What was the test setting: a simulation, a benchmark, or a live deployment?
- How many people or cases were involved, and who chose them?
- Do the authors list limits, and does the coverage repeat them?
- Is a date attached to any claim about product availability?
FAQ
Is Sony AI the same as Sony's consumer electronics business?
No. Sony AI's site describes it as a research organization established in April 2020. How its work is used inside other Sony products was not confirmed from primary sources for this page.
Did Gran Turismo Sophy beat professional drivers?
The Nature abstract says the agent beat four of the world's top Gran Turismo drivers in a head-to-head competition. That result concerns a racing game, not real-world driving.
Can anyone use the FHIBE dataset?
Coverage of the release says it is publicly available, but users must create an account and agree to the terms of use. The terms were not reviewed for this page.
Does FHIBE prove that AI models are biased?
It adds evidence for specific models and tasks. The authors tested eight narrow vision tasks and two vision-language models, and they caution that results can differ across datasets.
Related reading
- DeepSeek and how its models are described
- Baidu's Ernie Bot chatbot profile
- OpenAI as a research company
- How Tesla uses AI in its vehicles
More in AI Tools and Industry.
Last verified: October 2026
Sources
- Outracing champion Gran Turismo drivers with deep reinforcement learning, Nature 602, February 9, 2022
- Fair human-centric image dataset for ethical AI benchmarking, Nature 648, November 5, 2025
- Sony AI home page, Sony AI, accessed October 27, 2026
- Gran Turismo 7 news: Race Together event announcement, Polyphony Digital and Sony AI, February 20, 2023