Unlocking Creative Potential: The Value of AI and Human Creativity for the Long-Term, Getty Images

Ai4

AI-assisted summary of Unlocking Creative Potential: The Value of AI and Human Creativity for the Long-Term, Getty Images — original source. Check the original for full context.

Source published: 2025-08-29

Detailed notes

## Background Published on YouTube by the Ai4 channel on August 29, 2025, this session features Aditi Arora, VP of Product at Getty Images. According to the page description, the presentation explores the interplay between artificial intelligence and human creativity, highlighting the critical role of clean training datasets, commercial safety, copyright protection, and recurring creator compensation.

## TL;DR Aditi Arora presents Getty Images' vision for a collaborative AI ecosystem where generative technology amplifies rather than replaces human creativity. She demonstrates how rapidly shifting cultural and visual trends require continuous inputs of fresh, authentic human photography to avoid algorithmic bias and model collapse. To resolve enterprise concerns regarding copyright infringement and unfair labor practices, Getty Images developed a commercially safe AI model trained exclusively on its fully licensed catalog with built-in legal indemnification. Arora emphasizes that maintaining human creator compensation is vital because today's genuine human photography serves as the indispensable foundation for tomorrow's AI models.

## Atmosphere & style The presentation delivers a crisp, pragmatic, and polished keynote tone typical of an enterprise technology conference. Arora speaks with authoritative clarity and pacing, using concrete visual comparisons and proprietary trend data to ground high-level generative AI concepts. The delivery feels educational and reassuring, balancing the commercial momentum of AI adoption with a firm, principled advocacy for creator ethics and legal rigor.

## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=4s) Introduction and the Scale of Getty Images

Aditi Arora opens the session by welcoming the audience as midday approaches, promising to keep the presentation engaging and fast-paced to leave ample time for audience questions. The core focus of her talk centers on the evolving relationship between artificial intelligence and human creativity. Before diving into the technical and creative implications of generative models, she provides background on Getty Images and its broader portfolio of brands, which includes iStock and Unsplash. The platform serves as a primary marketplace and creator of visual content for both creative and editorial endeavors, spanning news, media, and entertainment coverage.

On the creative side, Getty Images caters to a spectrum ranging from small and medium-sized businesses (SMBs) to major global enterprises, supplying visual assets for multi-channel marketing campaigns, brand identity systems, and outdoor advertising like billboards. She underscores the immense operational scale powering the platform:

- **Editorial Scope:** Covers approximately 160,000 events annually across news, media, and entertainment. - **Contributor Base:** Over 600,000 active contributors worldwide, with 82,000 working exclusively with Getty Images. - **Catalog & Traffic:** More than 625 million visual assets hosted on the platform, fielding roughly 2.7 billion searches each year. - **Visual GPS Research:** A proprietary global research initiative surveying over 7,000 professionals across 25 countries to track shifting cultural trends and visual perception insights.

```mermaid flowchart TD accTitle: Getty Images Ecosystem Architecture accDescr: Visual diagram showing the creator inputs, platform core assets, research, and outputs to business customers.

subgraph Inputs["1. Visual Sourcing"] A["600K+ Global Contributors<br/>(82K Exclusive)"] B["160K Editorial Events<br/>(News, Media, Sports)"] end

subgraph Platform["2. Getty Images Platform Core"] C["Catalog: 625M+ Assets"] D["Search Engine: 2.7B Queries/Yr"] E["Visual GPS Research<br/>(25 Countries / 7K+ Pros)"] end

subgraph Solutions["3. Outputs & Products"] F["SMB & Enterprise Campaigns"] G["Commercially Safe AI Generator"] end

Inputs --> Platform Platform --> Solutions ```

---

[02:13](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=133s) The Generative AI Boom and the Collaborative Ecosystem

Arora addresses why generative AI has rapidly surged to the very top of mind for visual artists, marketers, and enterprise brands alike. Industry research reveals that 68% of creative professionals have integrated generative AI into their workflows over the past six months, while nearly half of global marketers currently rely on AI tools daily to generate images or video. Economically, the generative AI market is on a steep upward trajectory, expanding from an estimated $66 billion today toward $442 billion within the next five years. Moreover, the raw output volume has exploded: in 2024 alone, more than 12 billion AI-generated images were produced worldwide—far surpassing the total annual output of traditional professional photography.

Faced with this volume, Arora introduces Getty Images' core thesis: the future of visual content creation must be structured as a "collaborative AI ecosystem" rather than a zero-sum replacement of human effort. Generative AI models are only as capable as the foundation they are built upon, fundamentally requiring an ongoing stream of authentic, high-quality, and culturally nuanced visual material to produce meaningful and relevant results.

---

[03:47](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=227s) Cultural Evolution and Real-World Visual Shifts

To illustrate why static or automated data pools quickly become obsolete, Arora walks the audience through a visual retrospective of the past decade. Using data from Getty’s Visual GPS program, she demonstrates how real-world visual trends directly mirror rapid societal shifts. At the onset of the COVID-19 pandemic in 2020, top platform downloads shifted abruptly to images of individuals wearing face masks in darkened, isolated home environments. Pre-pandemic, such imagery would have evoked confusion or alarm, whereas by 2022, visual trends normalized hybrid workspaces, digital overlays, and prominent representation of women of color in professional settings. By 2024, corporate visuals evolved further toward casual workplace collaboration and diverse women in executive leadership roles.

Similar generational transitions appear across multiple categories: - **Energy and Infrastructure:** Top visuals evolved from 2014 industrial oil rigs to 2024 imagery of electric vehicles plugged into curbside chargers. - **Mental Health:** Depictions moved away from clinical, isolated clichés of solitary individuals in dark rooms toward nuanced portrayals of employees actively managing stress within modern work environments. - **Family Dynamics:** Visuals diversified from traditional nuclear structures to multigenerational households, LGBTQ+ families, and single parents.

Arora engages the audience with a side-by-side comparison of two images—one shot by a camera and one generated by AI. When asking the audience to vote on which was real, she explains that while the second image was AI-generated, "it doesn't really matter if it's camera shot or AI generated. What matters is the emotion it evokes for the audience." As technology progresses and the visual gap closes, the decisive element becomes human intentionality and understanding the cultural context behind the visual.

---

[06:51](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=411s) Data Quality, Model Collapse, and Creator Incentives

Arora details the mechanics of diffusion models, showing how high-quality, real-world images are transformed during model training through the addition of mathematical noise, which the network subsequently learns to denoise against semantic text captions. Because the model relies on this process to generate new imagery, the downstream output is directly bounded by the caliber of its inputs.

```mermaid flowchart LR accTitle: Generative Training and Ecosystem Feedback Loop accDescr: Diagram illustrating the lifecycle from human capture to training, generation, and contributor compensation.

subgraph Capture["1. Human Creation"] A["Photographers capture<br/>evolving real-world trends"] end

subgraph Training["2. Model Training"] B["High-quality clean data"] C["Noise addition & denoising"] end

subgraph Generation["3. Generative Output"] D["Authentic & unbiased visual output"] end

subgraph Feedback["4. Sustainable Loop"] E["Recurring contributor compensation"] end

A --> B --> C --> D D -.-> E -.-> A ```

Training datasets require accuracy, cultural diversity, and continuous updates to avoid skewed outputs or algorithmic bias. Arora warns against the trap of prioritizing volume over quality, highlighting the acute risk of **model collapse**: when newer generations of AI models are trained predominantly on synthetic, AI-generated content rather than genuine human photography, visual artifacts compound and overall model quality rapidly deteriorates.

This dynamic places human creators at the very center of AI sustainability: - Not paying human photographers disincentivizes them from capturing new, emerging real-world cultural moments. - Without fresh human photography, training data stagnates and AI models lose fidelity with modern society. - As Arora emphasizes: *"Today's pictures are tomorrow's training data set and the future of better models."*

---

[09:17](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=557s) Commercially Safe AI and Legal Protection

Addressing enterprise adoption barriers, Arora outlines the primary friction points reported by Getty Images' corporate customers: legal exposure and copyright infringement lawsuits, lack of visual authenticity, and ethical concerns regarding creator compensation. In response, Getty Images developed a proprietary commercially safe generative AI model built strictly on its fully licensed, clean catalog.

Demonstrating real-world side-by-side outputs, Arora shows that prompting an open web-scraped AI tool for "running shoes" frequently renders copyrighted trademarks like the Nike swoosh because the underlying model scraped unlicensed commercial branding. In contrast, Getty's engine outputs trademark-free designs safe for commercial deployment. Similarly, prompts for medical staff in scraped models often default to stereotypical middle-aged white male doctors, whereas Getty's model produces representative healthcare teams that mirror modern hospital staff.

---

[11:45](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=705s) Key Takeaways for Creators and Brands

Arora wraps up the session by directing attendees interested in deeper technical and commercial discussions to connect with her colleague Bailey, who is in attendance. She summarizes the fundamental guiding principles for working with generative media:

- **Know Your Why:** AI is an assistive amplifier for creative vision, not an outright substitute for human imagination and taste. - **Keep Models Relevant:** Because human culture and societal norms constantly shift, model training pipelines must continuously absorb contemporary human work. - **Know Your Tools:** Understand the provenance of training data, how the models were constructed, and what legal protections apply. - **Demand Fair Compensation:** Creators must receive recurring, fair payment for their work to sustain the underlying creative economy that feeds future generative systems. - **Stay Authentic:** A creator's personal viewpoint, contextual understanding, and unique human perspective remain their single greatest competitive advantage.

## Video description Aditi Arora, VP of Product, Getty Images

Abstract

Join Aditi Anora, VP of Product at Getty Images, for an engaging discussion on the interplay between AI and human creativity. This session will delve into how training data plays a crucial role in the effectiveness and quality of outputs—specifically how AI models can generate more authentic and culturally relevant visual content that is better reflective of modern society today. Aditi will highlight the importance of models being trained on transparent and clean data sets to avoid legal risks and safeguard intellectual property rights, as well as the nuances of developing AI models and tools that are high-quality and commercially safe. Lastly, the session will dive into the imperative of supporting content creators and recurring compensation to ensure that models are fueled responsibly, and technologies can be harnessed to push the boundaries of AI creativity versus hinder it.

[Source](https://www.youtube.com/watch?v=OYDbhzmiKeE)

中文简报

释放创意潜能：AI 与人类创造力的长期价值（Getty Images） Aditi Arora 阐述了 Getty Images 对人机协同 AI 生态系统的愿景，主张生成式技术应赋能增强而非取代人类创造力。她指出，瞬息万变的文化与视觉潮流需要源源不断注入新鲜、真实的人类摄影作品，以规避算法偏见和模型崩溃风险。为化解企业对侵犯版权及不当劳工做法的担忧，Getty Images 开发了一款商业安全的 AI 模型，该模型完全依托其获得完全授权的图库进行训练，并提供法律赔偿保障。Arora 强调，保障人类创作者的收益至关重要，因为当下真实的人类摄影正是未来 AI 模型不可或缺的发展基石。 Getty Images 产品副总裁 Aditi Arora 表示，生成式 AI 市场规模预计将在未来五年内从 660 亿美元跃升至 4420 亿美元；面对海量生成内容，视觉创作必须构建为“人机协同生态”而非单纯替代人类。 (2:13) 扩散模型的效果取决于输入质量，Arora 警告过度使用合成数据会导致“模型崩溃”（Model Collapse），唯有持续吸纳记录现实与文化变迁的真实人类摄影，才能维持模型的敏锐度与真实性。 (6:51) 针对侵权诉讼与偏见风险，Getty Images 依托完全授权且标注规范的图库构建了商业安全 AI 模型，可过滤商标侵权元素，提供商业法律赔偿保障，并呈现更多元化的人口构图。 (9:17) Arora 呼吁必须向创作者提供持续的合理报酬，以维持良性的模型训练数据闭环，同时强调人类独特的文化理解和情感视角始终是无法被替代的核心优势。 (11:45)

Reading summary

Aditi Arora of Getty Images argues that generative visual AI must operate within a collaborative ecosystem that continuously integrates and fairly compensates human creators to prevent model collapse and cultural irrelevance. Rapid Generative Adoption and Ecosystem Scale (2:13) Arora highlights a rapid expansion in visual content generation, citing industry research that 68% of creative professionals integrated generative AI into their workflows over the past six months, while nearly half of global marketers use AI generation tools daily. The broader generative AI market is projected to grow from an estimated $66 billion to $442 billion over the next five years, with more than 12 billion synthetic images produced worldwide in 2024 alone, outstripping traditional professional photography output. Contrasting this volume with Getty Images' underlying footprint—over 625 million catalog assets, 2.7 billion annual searches, and more than 600,000 active global contributors—Arora presents a model of a "collaborative AI ecosystem." Rather than viewing generative AI as a zero-sum replacement for human artists, she asserts that generative tools remain fundamentally reliant on an ongoing stream of authentic, culturally nuanced human imagery to produce relevant visual assets. Cultural Evolution and the Danger of Model Collapse (3:47) Drawing on Getty's Visual GPS research across 25 countries, Arora shows how commercial imagery mirrors rapid societal evolution. Top downloads shifted from 2014 depictions of industrial oil rigs to 2024 visuals of electric vehicle chargers, while workplace imagery progressed from isolated individuals in face masks in 2020 to hybrid setups and diverse women in executive leadership by 2024. Depictions of family structures and mental health have similarly transitioned toward nuanced, contemporary realities. Diffusion models function by learning to denoise mathematically perturbed imagery against textual descriptions, meaning generative outputs are inherently bounded by the fidelity and timeliness of their training data. Arora warns that training newer models predominantly on synthetic AI content rather than fresh human photography leads to "model collapse," in which visual artifacts compound and output fidelity rapidly degrades. Maintaining visual relevance requires continuously feeding newly captured real-world imagery into training pipelines. Sustaining Creator Economics and Data Quality (6:51) Arora emphasizes that the long-term technical health of generative models depends on economic incentives for human photographers. If creators are uncompensated, they will stop capturing emerging cultural and societal shifts, depriving training pipelines of fresh data and causing models to drift out of touch with modern society. As Arora summarizes, current human photography forms the foundation of future training datasets. To preserve this pipeline, Getty Images implements a model where catalog contributors receive recurring financial compensation when their work informs generative training. Arora advises creators and brands to view AI as an assistive amplifier of human vision rather than an autonomous replacement, noting that human intentionality and contextual understanding remain a creator's core competitive advantage. Addressing Enterprise Copyright Liability and Bias (9:17) Corporate adoption of generative tools faces friction around copyright infringement risks, visual authenticity, and ethical creator treatment. Arora notes that models trained on uncurated, scraped web data frequently output protected intellectual property—such as trademarked Nike logos on footwear queries—or reproduce demographic stereotypes, such as defaulting exclusively to white male doctors for medical prompts. To counter these risks, Getty Images built a commercial generator trained exclusively on its fully licensed, curated catalog. This architecture filters out infringing third-party intellectual property, provides commercial legal indemnification for enterprise users, and generates demographic representations that reflect modern workforces. Arora asserts that as the technical fidelity between camera capture and AI generation narrows, adherence to legal safety and cultural accuracy will dictate commercial utility. Arora's framework positions commercial indemnification and recurring creator royalties as operational imperatives for maintaining model quality and legal safety, though enterprises must still weigh these licensed systems against the widespread availability of low-cost, web-scraped alternatives.

精炼阅读摘要

Getty Images副总裁Aditi Arora指出，生成式人工智能的发展不应演变为对人类摄影师的零和替代，而应构建一个人机共生的协作生态；AI图像模型的输出质量与生命力直接依赖于输入数据，必须由人类创作者持续提供反映真实文化演变的高质量素材以避免模型退化。视觉生成规模爆发与协作生态构建 (2:13) 生成式AI工具已迅速渗透进商业创意与品牌营销流程。行业统计显示，68%的创意专业人士在过去六个月内将生成式AI整合入工作流，近半数的全球营销人员日常依赖AI工具生成图像或视频。从经济规模来看，生成式AI市场预计将在未来五年内从660亿美元扩张至4420亿美元。仅在2024年，全球生成的AI图像就已突破120亿张，在纯产出数量上大幅超越了传统专业摄影的年度规模。面对合成内容的爆发式增长，Getty Images提出核心主张：视觉内容的未来必须建立在“人机协作生态”之上，而非将AI视为对人类劳动力的单一替代。Arora强调，生成式AI模型的能力完全受限于其底座训练集；若缺乏持续流入的真实、高水准且具备文化细节的视觉素材，算法模型便无法自主产出具有社会相关性与现实意义的内容。社会文化演进与人类真实表达的价值 (3:47) 基于覆盖25个国家、针对7000余名专业人士的Visual GPS研究数据，Arora展示了过去十年真实社会文化变迁如何直接映射在视觉需求中。以2020年新冠疫情为例，平台热搜素材迅速转向昏暗隔绝环境与佩戴口罩的画面；而到了2022至2024年，视觉趋势进一步演变为混合办公场景以及由少数族裔女性担任高管职位。类似的演变广泛出现在能源形态（从传统油井转向电动汽车充电桩）、心理健康（从孤立无助的陈旧刻板印象转向职场压力管理）以及家庭结构（吸纳隔代抚养与多元家庭形式）等各个领域。 Arora通过现场盲测对比摄影照片与AI生成图像说明，随着生成技术不断精进，判定画面是相机捕捉还是算法合成已不再关键；决定视觉资产长久价值的，是图像能否激发受众的情感共鸣，以及创作者对现实文化背景的洞察力与表达意图。扩散机制、合成数据陷阱与激励闭环 (6:51) 扩散模型在技术训练中主要依赖于向高质量真实图像中添加数学噪声，再训练神经网络依据文本描述进行去噪复原。Arora指出，这决定了模型的输出质量严格受制于输入数据的优劣。一味追求训练数据的绝对数量而牺牲质量会导致严重的算法偏差，尤其当后代模型过度依赖AI合成内容进行自循环训练时，视觉瑕疵与伪影会成倍放大，从而引发“模型崩溃”（model collapse），使生成效果全面退化。这一机制使得人类创作者成为AI可持续运作的关键节点。Arora提出：“今天的照片就是明天的训练数据集，也是未来打造更优质模型的基石。”如果平台与下游企业未能向记录时代变化的摄影师支付合理报酬，创作者将丧失记录现实生活的经济动力，进而导致上游训练数据陷入停滞，最终使AI模型与现实社会脱节。企业商用风险与合规数据资产保障 (9:17) 针对企业在应用生成式AI时面临的阻碍，Arora归纳为三大痛点：知识产权侵权与版权诉讼风险、视觉内容缺乏真实度，以及创作者报酬缺失带来的伦理指责。从公开网络抓取未授权数据训练的模型，容易在生成日常物品时输出具有法律争议的第三方商标（例如输入跑鞋提示词却生成Nike勾状标志），且往往会在特定职业形象上固化偏见（如默认医生均为中年白人男性）。对此，Getty Images依托其包含6.25亿视觉资产的授权库与全球60万内容贡献者体系，推出了商用安全型生成引擎。该工具仅使用完全获得授权的数据集进行训练，主动过滤第三方知识产权标志以提供全面的商业法律赔偿保障，并在算法中纠正人口统计学偏见，同时通过经常性版权分成回馈人类贡献者，以实现商业合规与创作者经济的良性循环。 Arora总结建议，品牌与创作者应将AI视作放大人类审美意图的辅助工具，在使用前必须严加核查训练数据的来源合法性与法律保护条款，并确保创作者获得合理回报以维系生态韧性。不过，该分享未进一步说明在开源无授权模型成本持续降低的市场竞争下，合规商业授权方案如何长期保持价格竞争力，亦未披露针对贡献者版权分成的具体财务分配比例。

Detailed notes — zh

## Background 该视频由 YouTube 频道 Ai4 于 2025 年 8 月 29 日发布，记录了 Getty Images 产品副总裁 Aditi Arora 的主题演讲。演讲探讨了人工智能与人类创造力之间的相互作用，深入分析了干净、透明的训练数据对生成真实且具备文化相关性视觉内容的重要性，并阐述了如何规避法律与知识产权风险以及建立对创作者的持续报酬机制。 ## TL;DR Aditi Arora 首先介绍了 Getty Images 庞大的视觉生态资产与生成式 AI 在创意领域的爆发式应用，指出人类与 AI 应当建立协作而非替代关系。她结合过去十年社会变迁的视觉演进指出，AI 模型若想具备真实的情感共鸣与多元代表性，必须依赖摄影师对现实世界的实时捕捉，否则反复以 AI 生成内容进行训练将引发“模型退化”危机。通过对比通用 AI 模型在商标侵权与刻板印象上的缺陷，她展示了 Getty Images 基于干净版权数据打造的商业安全级 AI 方案，并呼吁行业确保创作者获得合理经济补偿，让 AI 真正成为人类独特创意的放大器。 ## Atmosphere & style 整场演讲专业洗练、逻辑清晰，既有企业高管对行业前沿趋势的宏观把控，又充满对创作者权益的人文关怀。讲者通过生动的现场盲测互动与鲜明的行业案例对比，将复杂的数据合规与模型崩溃技术原理讲得深入浅出。现场氛围在冷静理性的商业与技术剖析中逐渐升温，最终收尾于对创作者独特性与公平回报的有力呼吁，具有极强的说服力与鼓舞人心的力量。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=4s) Getty Images 依托全球数十万创作者与海量资产，通过 Visual GPS 调研持续捕捉社会文化演进与全球视觉趋势。 - [02:13](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=133s) 生成式 AI 视觉市场正在爆发式增长，创作者与品牌应构建互利共生的“协同 AI 生态系统”，而非将 AI 视作替代品。 - [03:47](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=227s) 视觉呈现深刻映射了过去十年社会在混合办公、清洁能源及家庭多元性上的变迁，AI 仍必须依赖真实摄影师捕捉时代切面。 - [03:47](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=227s) 随着 AI 生成图像逼真度大幅提升，视觉资产的核心价值已转向其能否唤起真实的情感共鸣与还原深层社会语境。 - [06:51](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=411s) 输入数据的质量决定模型输出上限，具备高准确度、多样性与代表性的优质样本远比盲目堆砌海量数据更关键。 - [06:51](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=411s) 警惕“模型退化”（Model Collapse）危机：过度依赖上一代 AI 生成的合成数据进行二次训练，会导致模型输出表征走向崩溃。 - [06:51](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=411s) “今天的照片就是明天的训练数据集”，必须建立合理公平的创作者补偿机制，保障原创者持续生产高质量真实数据的积极性。 - [09:17](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=557s) 商业安全级 AI 模型必须基于干净、授权的合规数据进行训练，从源头杜绝商标与 IP 侵权，并提供法律免责赔偿保障。 - [09:17](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=557s) 通过高质量且具代表性的数据集，能够有效消除通用模型在性别与族裔上的职业刻板印象，真实呈现当代社会风貌。 - [12:01](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=721s) 创作者与企业必须深入了解所采用 AI 工具的数据来源与训练透明度，以长远眼光维护数据生态的健康度。 - [12:01](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=721s) 无论技术媒介如何演变，人类创作者基于真实生活体验所拥有的“独一无二的视角”，始终是 AI 无法替代的核心竞争力。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=4s) Getty Images 的业务版图与视觉生态 Getty Images 产品副总裁 Aditi Arora 在开场中简要活跃现场气氛，表示希望在中午时段为大家带来一场既生动高效又能留足问答时间的分享。在正式探讨人工智能与人类创造力的深层价值之前，她首先向在场观众系统介绍了 Getty Images 的业务架构。作为一个全球领先的视觉内容创作者平台与交易市场，Getty Images 旗下汇聚了包括 Getty Images、iStock 和 Unsplash 在内的知名视觉品牌，为客户在创意与编辑项目中提供丰富的视觉资产。在编辑业务板块，平台每年覆盖约 16 万场活动，横跨新闻、媒体与娱乐领域；而在创意业务板块，其客户群体则涵盖从中小型企业（SMB）到大型跨国企业的广泛客群，广泛应用于品牌营销战役、企业形象物料构建及户外广告牌设计等场景。 Aditi Arora 列举了一组庞大的业务数据来展示其生态规模：平台拥有约 60 万名全球内容供稿人（Contributors），其中 8.2 万名更是 Getty Images 的独家签约创作者；平台每年承接约 27 亿次搜索请求，库内储备了 6.25 亿件视觉资产。为了精准洞察视觉趋势，Getty Images 运营着一项名为 Visual GPS 的全球调研项目，通过对 25 个国家的 7,000 多名专业人士开展问卷调查，捕捉视觉内容领域的最新趋势与消费者心态，并将深度洞察反馈给客户群体。此外，她还特别提及了平台自主研发的“商业安全级”AI 生成器（Commercially Safe AI Generator），为后续的深入探讨埋下伏笔。 - **核心品牌矩阵**：Getty Images、iStock、Unsplash - **编辑业务广度**：年均报道覆盖约 160,000 场新闻、媒体及娱乐活动 - **供稿人与资产体量**：超 600,000 名创作者（其中 82,000 名独家签约），库藏资产达 6.25 亿件 - **年搜索量**：全球每年 27 亿次检索 - **Visual GPS 调研体系**：覆盖全球 25 个国家、逾 7,000 名受访者，持续输出行业视觉趋势报告 --- ### [02:13](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=133s) 创意加速与协同 AI 生态系统的兴起生成式人工智能（Gen AI）正在成为全球创作者与品牌营销人员不可回避的核心议题。数据显示，在过去 6 个月内，已有 68% 的创意专业人士在日常工作中使用生成式 AI；而在全球营销人员中，近半数每天都会借助 AI 进行图像或视频内容的创作。从市场规模来看，整个生成式 AI 视觉市场当前估值约为 660 亿美元，预计在未来 5 年内将激增至 4,420 亿美元。单看 2024 年一年，由 AI 生成的内容总量就突破了 120 亿件，这一数量已经远远超过了全球专业摄影师在一年内所能拍摄交付的全部照片总量。面对内容爆发式增长的现状，Aditi Arora 提出了对未来视觉内容创作的思考：人类与 AI 的关系并非简单的替代，而应当构建一个“协同 AI 生态系统”（Collaborative AI Ecosystem）。她指出，AI 模型要输出高品质、真实且具有文化相关性的视觉结果，其底层的根本依赖正是高质量、真实、具有当代语境的人类原创内容。 ```mermaid flowchart LR accTitle: 协同 AI 生态系统逻辑流 subgraph S1["人类真实世界"] A["现实社会演变与文化脉动"] --> B["专业创作者捕捉真实画面"] end subgraph S2["模型训练体系"] B --> C["高品质、真实合规训练集"] C --> D["生成式 AI 模型学习"] end subgraph S3["商业应用交付"] D --> E["生成具情感共鸣的安全资产"] E --> F["创作者获得持续经济回报"] F -.->|激励再创作| B end ``` --- ### [03:47](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=227s) 视觉像素维度的十年社会演进与盲测启示回顾过去十年，工作环境、技术工具以及人际协作模式发生了深刻的重构。Aditi Arora 借助 Visual GPS 过去 5 年的数据展示了视觉搜索与下载趋势的变迁：在 2020 年新冠疫情初期，平台下载量最高的图片大多是佩戴口罩的女性，或是昏暗居家环境中的办公特写；这些画面如果在疫情之前展示给受众，会引发完全不同的心理感受。到了 2022 年，视觉呈现逐渐正常化并走向多元，更多展现混合办公模式（Hybrid Work）、视觉叠加特效以及有色人种女性的职场形象；而到了 2024 年，画面趋势进一步演进为办公室中轻松自然的互动与女性高管的领导力展现。类似的变化在各领域随处可见：能源领域的视觉从 2014 年的海上石油钻井平台，转向 2024 年路边充电的电动汽车；心理健康的视觉表达从孤独坐在房间里的男性，转变为在真实职场中一边穿行一边应对压力与挑战的现代职场人；家庭结构的视觉呈现也更加包容，广泛纳入了跨代同堂家庭、LGBTQ+ 群体以及单亲母亲等形象。在现场互动环节，Aditi 展示了两张视觉风格极为逼真的特写照片让观众竞猜：一张是相机实拍，另一张由 AI 生成。当揭晓第二张其实是 AI 生成后，Aditi 总结道，随着技术的飞速进步，肉眼辨别实拍还是 AI 生成正变得越来越困难。但真正的关键在于画面能否为受众唤起真实的情感共鸣，"真正重要的是谁在拍摄这张照片，以及他们是否理解拍摄当时的真实语境"。如果未来十年社会仍将加速进化，AI 模型就必须依赖身处真实世界中的摄影师去捕捉这些不断演进的社会切面。 | 视觉主题 | 过去十年呈现（约 2014 / 2020） | 当前演变趋势（2024） | |:--- |:--- |:--- | | **职场与办公** | 疫情初期口罩遮面、昏暗居家隔离办公 | 混合办公、女性高管领导力、轻松自然的职场社交 | | **能源基础设施** | 海上石油钻井平台等传统化石能源 | 路边充电桩与电动汽车（EV）的普及场景 | | **心理健康** | 单一男性在密闭房间内抱头独坐 | 现代职场人直面日常挑战与心理调节的真实动态 | | **家庭结构** | 传统单一核心家庭构成 | 多代同堂、LGBTQ+ 家庭、单亲母亲等包容性呈现 | --- ### [06:51](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=411s) 训练数据质量、模型退化危机与创作者价值闭环 Aditi 强调了模型开发中的核心公理：输入给训练数据集的数据质量，直接决定了模型输出结果的品质上限。训练数据必须具备高准确性、多样性与代表性，才能避免生成结果产生偏见与失真。当下模型训练的关键在于平衡“规模与质量”，行业更需要的是高品质的范例样本，而非盲目堆砌数据数量。同时，训练集必须保持实时更新，以映射真实社会的变化。她在此处发出了关于“模型退化”（Model Degradation / Model Collapse）的严厉预警：如果未来的模型主要使用由上一代 AI 生成的内容来进行训练，随着迭代周期的推进，模型的生成质量将不可避免地走向崩溃。随后她解析了扩散模型的基本原理：模型在人类实拍的真实狗照片上不断添加噪声，再通过理解图像的语义标注进行逐步去噪，从而学习并生成出新的图像。基于这一技术现实，Aditi 指出摄影师与创作者正是整个生态系统的绝对核心。如果企业在抓取这些图像用于训练时不给予摄影师合理的报酬，就会扼杀创作者探索新创意的积极性。她给出了一个至关重要的论点："今天的照片就是明天的训练数据集，更是未来更优质模型的基石"（"Today's pictures are tomorrow's training data set and the future of better models"）。 - **数据质量法则**：高准确度、多样性与代表性是消除偏见、提升生成质量的关键 - **样本质量高于纯粹规模**：需要更优质的高维特征样本，而非简单的无序数据爬取 - **模型退化（Model Collapse）风险**：AI 生成内容反复自我训练会导致模型输出退化与表征崩溃 - **创作者经济保障**：必须对提供训练源数据的创作者提供持续经济补偿，维护生态健康 --- ### [09:17](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=557s) 商业安全实践：IP 侵权规避与社会多元表征为了解决企业客户在使用 AI 图像生成时面临的三大核心顾虑——法律诉讼风险、内容真实性欠缺以及创作者补偿争议，Getty Images 构建了一套全流程商业安全的 AI 解决方案。该模型完全基于干净合规、已获授权的自有版权数据进行训练，确保生成的图像不会侵犯第三方知识产权（IP），同时为客户提供法律免责赔偿保障（Legal Indemnification），并在商业收益中切实向贡献数据的创作者提供分成。在对比演示中，Aditi 展示了行业通用 AI 与 Getty Images AI 的差异： 1. **商标侵权风险**：当输入“跑鞋”（Running Shoes）的提示词时，行业通用生成器由于使用了未授权的互联网抓取数据，生成的跑鞋侧面赫然出现了耐克的 Swoosh 标志，给商业使用带来了巨大侵权隐患；而 Getty 的生成模型则完全避开了专有商标侵权。 2. **社会偏见与真实度**：当输入“医生与护士”（Doctors and Nurses）的提示词时，行业通用生成器依然倾向于输出刻板的中年白人男性医生形象；Getty 的生成器则更符合当下真实医院的工作环境，呈现出更多元的性别与族裔分布。在演讲最后，Aditi 呼吁与会者明确自身的目标定位——"明确你的初衷，AI 应当服务于你的远景，而非成为替代品"（"Know your why. AI should serve your vision. It is not a replacement"）。AI 不是人类创造力的替代者，而是放大器；只有不断以反映真实世界变迁的高质量数据滋养模型，AI 技术才能长久保持生命力。 - **干净授权数据训练**：从源头杜绝未经许可的网络爬虫数据污染 - **全面商业安全与法律赔偿**：免除商业落地中的知识产权纠纷隐患 - **消除刻板印象**：通过高质量代表性数据集，还原真实世界的多元职业与族裔面貌 - **核心定位认知**：将 AI 视为人类灵感的放大器与协作工具，而非人类创造力的代用品 ### [12:01](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=721s) 创作者的核心准则：了解工具来源、坚持合理报酬与保持真实视角在演讲的收尾阶段，Aditi Arora 向所有创作者和行业从业者提出了几项至关重要的建议与呼吁。首先，她着重强调了“了解你的工具”（"Know your tools"）这一基本原则。使用者必须清楚地理解所采用的 AI 模型其数据究竟来自何处、背后又是如何进行训练的。她警示大家必须建立长远的眼光，因为“今天的照片就是明天的数据集”（"Today's photos are tomorrow's data sets"），当前生成与拍摄的每一份视觉资产，都将构成未来模型训练的基石，因此对数据源头的审视刻不容缓。随后，Aditi Arora 点明了整场演讲最核心的诉求——公平的报酬机制（"Fair compensation is the heart of the entire conversation we had today"）。她明确表示，只要创作者投入了创意并产出价值，就理应获得相应的报酬（"If you are being creative, you need to get paid for it"），这一原则在技术迭代中不容动摇。最后，她勉励所有创作者坚持真实性（"stay authentic"），并指出无论作品是由 AI 辅助生成，还是通过相机实景拍摄，创作者自身“独一无二的视角始终是最大的竞争优势”（"Your unique perspective is your biggest advantage"）。演讲在这一充满鼓舞的呼吁与现场热烈的掌声中圆满结束。 - **理解工具与数据溯源**：深入掌握 AI 工具的训练集构成与获取渠道，建立对底层数据透明度与合规性的认知。 - **数据循环与长远影响**：当下生产的照片与图像正是未来 AI 模型的训练养料，决定了未来技术生态的健康度。 - **创作者权益与公平报酬**：为创意付费是推动良性生态的核心，必须确保内容创作者在 AI 时代获得持续、公正的回报。 - **坚守真实与独特视角**：无论采用何种媒介或技术手段，创作者个人的独特视角才是无法被替代的核心竞争力。 ## Notable quotes > [03:47](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=227s) 真正重要的是谁在拍摄这张照片，以及他们是否理解拍摄当时的真实语境 — Aditi Arora > [06:51](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=411s) 今天的照片就是明天的训练数据集，更是未来更优质模型的基石 — Aditi Arora > [06:51](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=411s) Today's pictures are tomorrow's training data set and the future of better models — Aditi Arora > [09:17](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=557s) 明确你的初衷，AI 应当服务于你的远景，而非成为替代品 — Aditi Arora > [09:17](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=557s) Know your why. AI should serve your vision. It is not a replacement — Aditi Arora > [12:01](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=721s) Know your tools — Aditi Arora > [12:01](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=721s) Today's photos are tomorrow's data sets — Aditi Arora > [12:01](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=721s) Fair compensation is the heart of the entire conversation we had today — Aditi Arora > [12:01](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=721s) If you are being creative, you need to get paid for it — Aditi Arora > [12:01](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=721s) stay authentic — Aditi Arora > [12:01](https://www.youtube.com/watch?v=OYDbhzmiKeE&t=721s) Your unique perspective is your biggest advantage — Aditi Arora ## Mentioned Aditi Arora, Getty Images, iStock, Unsplash, Visual GPS, Nike, Ai4, 扩散模型 (Diffusion Models), 生成式人工智能 (Generative AI) ## Video description Aditi Arora, VP of Product, Getty Images Abstract Join Aditi Anora, VP of Product at Getty Images, for an engaging discussion on the interplay between AI and human creativity. This session will delve into how training data plays a crucial role in the effectiveness and quality of outputs—specifically how AI models can generate more authentic and culturally relevant visual content that is better reflective of modern society today. Aditi will highlight the importance of models being trained on transparent and clean data sets to avoid legal risks and safeguard intellectual property rights, as well as the nuances of developing AI models and tools that are high-quality and commercially safe. Lastly, the session will dive into the imperative of supporting content creators and recurring compensation to ensure that models are fueled responsibly, and technologies can be harnessed to push the boundaries of AI creativity versus hinder it. [Source](https://www.youtube.com/watch?v=OYDbhzmiKeE)

Source quotes

[3:47] Aditi Arora: 真正重要的是谁在拍摄这张照片，以及他们是否理解拍摄当时的真实语境 [6:51] Aditi Arora: 今天的照片就是明天的训练数据集，更是未来更优质模型的基石 [6:51] Aditi Arora: Today's pictures are tomorrow's training data set and the future of better models [9:17] Aditi Arora: 明确你的初衷，AI 应当服务于你的远景，而非成为替代品 [9:17] Aditi Arora: Know your why. AI should serve your vision. It is not a replacement [12:01] Aditi Arora: Know your tools [12:01] Aditi Arora: Today's photos are tomorrow's data sets [12:01] Aditi Arora: Fair compensation is the heart of the entire conversation we had today [12:01] Aditi Arora: If you are being creative, you need to get paid for it [12:01] Aditi Arora: stay authentic [12:01] Aditi Arora: Your unique perspective is your biggest advantage

Original transcript

[0:04] Hello everyone. Can you hear everyone? [0:07] Um, good afternoon. We are almost at 12. [0:10] I'll try to make sure we keep the [0:12] session as entertaining as possible and [0:14] as quick as possible as well so that we [0:17] get enough time for questions. So I will [0:20] be talking about the value of AI and [0:22] human creativity. But before we deep [0:25] dive into this specific topic, let's [0:27] talk about Getty Images. For those who [0:29] you don't know what Getty does, um we [0:33] are a visual content creator and a [0:36] marketplace. You may have heard of [0:38] brands that we own, Getty Images, [0:40] Eyestock, and Unsplash. We customers [0:43] come to us um for getting access to [0:46] visual content for their creative as [0:49] well as editorial uh projects with [0:52] within editorial. We have about 160,000 [0:55] events that we cover. So all the way [0:57] from news, media and entertainment. Uh [1:01] today I'll talk mostly about creative [1:03] side of our business which serves pretty [1:06] much SMBs all the way to enterprises. [1:09] They use and leverage our content for um [1:13] all the way from marketing campaigns to [1:15] building out their corporate branding [1:17] material, billboards, etc. [1:20] So overall we have about 600,000 [1:23] contributors who contribute to our [1:25] platform out of which 82,000 are [1:28] exclusive to Getty Images. We get about [1:32] 2.7 billion searches per year. We also [1:36] have about 625 billion assets million [1:39] assets on our platform. So we also have [1:42] a research program called visual GPS and [1:46] what it does is it actually surveys [1:49] professionals around the globe in 25 [1:51] countries um as well as it's about over [1:54] 7,000 individuals capturing trends and [1:57] insights about what's trending or what's [2:00] on people's mind with respect to visual [2:03] content and that's reported to our [2:05] customer base and we also have a [2:07] commercially safe AI generator which [2:09] I'll touch on towards end of the [2:11] presentation. [2:13] So why are we talking about this topic? [2:16] Why is AI top of mind for every creator [2:19] as well as a brand? Um if you look at [2:22] some trends here, 68% of creative [2:25] professionals use Gen AI in the last 6 [2:27] months. [2:30] We also see acceleration of content [2:33] creation. Nearly half of marketeers [2:35] today worldwide use AI daily for image [2:39] creation or video creation. We are also [2:42] expecting the market to grow [2:43] drastically. So we are at 66 billion [2:45] today expected to grow to 442 billion in [2:49] the next 5 years. On top of it, if you [2:52] look at the sheer volume of content [2:54] created through AI in the last year in [2:57] 2024, it was over 12 billion. So which [3:00] is way more content that's captured [3:03] through photography uh professional [3:05] photography in a year. So what does it [3:08] leave us with the future of visual [3:10] content creation? Uh what I believe and [3:14] we would I would like to hear your [3:16] thoughts as well after the session is we [3:19] see this as a collaborative AI [3:21] ecosystem. So let's unfold this um in [3:24] the next few minutes. When I say a [3:26] collaborative AI ecosystem, what I'm [3:28] saying is training AI models require [3:33] highquality authentic relevant content [3:36] for them to pro uh provide good relevant [3:39] outputs. So let's step back and look at [3:42] how has our work and our environment [3:44] changed over the last 10 years. [3:47] So if you think about 10 years ago, what [3:50] kind of colleagues were you working [3:51] with? Um what kind of technologies were [3:55] you working with? Um if you look at it [3:58] overall we looked at some trends through [4:00] visual GPS in the last 5 years. This is [4:04] right when co started. You see a woman [4:06] in the mask. These were the top [4:08] downloads across Getty Images website. [4:10] You also see a dark room a home setting [4:14] which were quite popular at the time. [4:16] But let's say if I showed you these [4:18] visuals precoid what would you think? [4:21] What emotions would it evoke for you? In [4:24] 2022, we saw this normalized. We saw way [4:27] more visual overlays. We saw hybrid [4:30] working environments and we saw women of [4:32] color showing up in most of the images [4:35] that were downloaded. And 2024, fast [4:38] forward, we're looking at casual [4:42] interactions at work. We're also looking [4:44] at women in leadership. [4:48] Let's look at some other trends. energy [4:49] sources. [4:51] 2014, it was images like oil rigs that [4:54] were showing up. 2024, you're looking at [4:56] electrical vehicles charging at the [4:59] sidewalk. [5:00] Mental health, there's an image of a man [5:03] sitting in a room alone. And 2024, [5:07] you're looking at more images wherein [5:10] there's a person who is at work [5:13] literally walking through day-to-day [5:16] work as well as managing their [5:18] struggling uh struggling outputs what [5:20] they have with work. [5:23] Depiction of families, we see much more [5:26] multigenerational families. We see LGBTQ [5:29] plus families. We also see single [5:31] mothers all depicted in these pixels. [5:35] Same with this two shots that were [5:38] taken. One is camera shot and one is AI [5:40] generated. So let me ask you this. Can [5:43] you tell me the difference between the [5:44] two images? How many of you think the [5:47] first image is camera shot? [5:51] How many of you think the second image [5:53] is the camera shot? [5:57] Okay. Uh the second image is AI [5:59] generated. It doesn't really matter. The [6:01] qu answer here is it doesn't really [6:03] matter if it's camera shot or AI [6:05] generated. What matters is the emotion [6:08] it evokes for the audience who are [6:10] looking at this image. So as the [6:12] technology get better, we'll see that [6:14] it'll be hard to distinguish between [6:16] camera shot and AI shot. So what matters [6:19] is who is taking the picture and if they [6:22] understand the context in which this [6:23] picture was taken. [6:27] So within the last decade we saw the [6:30] pixels evolving. Within the next decade [6:32] we'll absolutely see the pixels evolving [6:35] more. Do you agree? Yes. No. We'll see [6:38] the society evolve and we would want our [6:42] um photographers to be able to capture [6:45] all that we see uh in the real world [6:48] being also showcased through the AI [6:49] models. [6:51] So in a summary nutshell, what I'm [6:54] trying to say here is data quality that [6:57] is provided to the training data sets [6:59] will drive the overall output quality of [7:02] the model. So the training data needs to [7:05] be accurate, diverse and representative [7:07] to be able to produce relevant results [7:11] and make sure that the outputs are not [7:13] skewed. We want to make sure we balance [7:16] scale and quality. uh we the training [7:19] datas need way more high quality [7:21] examples not just more examples. We also [7:24] want to make sure the models are up to [7:26] date. We just talked about it. We want [7:28] to make sure the AI models reflect the [7:30] changes in the real world. We want to [7:33] prevent model degradation. So what [7:36] happens when training of um the models [7:38] is done predominantly on AI generated [7:41] content over the next generation of [7:44] models you'll see the quality of the [7:46] model collapse. [7:49] So let's type talk about a bit more on [7:52] why training data is important. So on [7:55] the top as you see there's a picture of [7:57] a dog which is camera shot and real but [8:00] when the machine uses it for the [8:02] training data set it adds noise um to [8:05] the picture and then eventually it will [8:07] d noiseise the picture by understanding [8:09] the semantic captions of the image and [8:12] generate the bottom left um the image [8:15] that you see of the dog. So [8:19] the point here is it's really important [8:21] that we have good quality visuals, real [8:26] visuals created for uh training the [8:28] models. So what are we summarizing? [8:31] Creators are at the center of the [8:33] ecosystem. [8:34] Not compensating photographers put the [8:37] creative ecosystem at risk. Let's recap. [8:40] Society is evolving. We want to make [8:42] sure the photographers who capture these [8:45] depictions [8:47] are compensated. What happens is when [8:50] these photographers collect all of these [8:52] images and put it out there, they are [8:55] used as training data sets for AI [8:58] generative models. So not paying them [9:00] deincentivizes them for building new [9:03] creative ideas. [9:05] So the big point here is if anything you [9:09] took away from here was today's pictures [9:12] are tomorrow's training data set and the [9:14] future of better models. [9:17] So I want to talk a little bit about how [9:19] is Getty Images working through this and [9:23] how are we making sure that the um [9:26] entire creative ecosystem is thriving. [9:29] So we looked at the top concerns that [9:32] were um our customers reported on AI [9:35] image generation all the way from legal [9:38] concerns they don't want to get sued it [9:40] also was they want to see more authentic [9:43] content it was about the photographer [9:45] compensation. [9:47] So what we did was we built out [9:49] commercially safe model. What it means [9:51] is our model is trained on clean [9:55] licensed data. We make sure the [9:58] commercially safe output that you get [10:01] does not lead to infringing IP. We also [10:04] protect your original ideas. We make [10:06] sure that you get legal legal [10:08] indemnification. [10:09] And finally, the most important thing, [10:12] we are making sure the contributors are [10:14] getting compensated. [10:18] So if you would like to learn a bit more [10:20] about what we are doing uh with [10:21] generative AI, please um talk to Bailey [10:25] who's here with me today. and I'll be [10:27] happy to talk uh and listen to your [10:30] thoughts as well. So this is just a [10:33] brief example of what industry AI [10:35] generators generate as compared to Getty [10:38] Images AI generator. We prompted for [10:40] running shoes and as you can see on the [10:42] left hand side Nike swoosh shows up and [10:45] with Getty you would not see that [10:47] avoiding any legal concerns and the [10:50] reason why that happens with Nike swoosh [10:52] showing up in most of the industry air [10:54] generators is it is using scrape data. [10:56] They have not compensated or licensed [10:59] any of the data that they build the big [11:01] models on. [11:04] So last but not the least, we also [11:07] looked at industry AI generators. What [11:09] do they generate when you're asking them [11:12] for uh prompting them on doctors and [11:14] nurses? So even though it's partially [11:18] correct, you can see like there are [11:20] middle-aged white male doctors, but it [11:22] doesn't really reflect the society. And [11:24] within the Getty images, you see [11:26] something which is more real. This is [11:28] what you see in hospitals, and that's [11:30] what we care about the most. [11:33] So training data will absolutely reflect [11:36] society and that's what would lead to a [11:38] generative model that reflects your [11:40] actual outputs that come out of the [11:42] model. So key takeaways from my side, [11:45] know your why. AI should serve your [11:48] vision. It is not a replacement. AI is [11:52] not a substitute for human creativity. [11:54] Please make sure it is used to amplify [11:56] it. Keep models relevant. The world [11:59] changes and so our models need to [12:01] reflect those changes. Know your tools. [12:04] Please understand how that model was [12:06] sourced and how was the data trained on [12:08] it. Today's photos are tomorrow's data [12:12] sets. Fair compensation is the heart of [12:15] the entire conversation we had today. If [12:18] you are being creative, you need to get [12:20] paid for it. Whether it's AI generated [12:23] or whether it's through the camera shot [12:25] images, stay authentic. Your unique [12:28] perspective is your biggest advantage. [12:32] Thank you. [12:34] [Applause]