How artists can protect their work from AI | Dr. Heather Zheng | TEDxChicago

Key Concepts

• Human Identity: Encompasses not only biometric data (face, voice, fingerprints) but also knowledge, experiences, relationships, and cultural background.
• Generative AI: Artificial intelligence models capable of creating new content, including realistic fake images, voices, and artistic styles.
• Adversarial Machine Learning: A field of machine learning focused on developing models that can withstand or exploit vulnerabilities in other AI systems.
• Data Scraping: The automated extraction of data from websites or digital sources, often used to train AI models.
• Copyright Infringement: The unauthorized use of copyrighted material.
• Deepfakes: Highly realistic synthetic media, typically videos or audio recordings, that depict people saying or doing things they never actually said or did.
• Glaze: A tool developed to protect the artistic style of visual artists by subtly altering their artwork to mislead AI models.
• Nightshade: A tool developed to protect individual objects or copyrighted characters within artwork by teaching AI models that these elements are something different.

The Threat of Generative AI to Human Identity and Creativity

Dr. Heather Jung, a computer science faculty member at the University of Chicago, discusses the profound impact of generative AI on human identity, a concept that extends beyond physical attributes to encompass knowledge, experiences, relationships, and cultural background. She highlights that identity is crucial for self-perception and societal positioning, and its theft constitutes a deep personal violation.

Generative AI's Impact on Biometric Identity

Generative AI models pose a significant threat by being trained on personal data to create convincing fake versions of individuals.

• Facial Identity: With just a few photos, AI models can generate realistic but fake images of a person within minutes. These models are readily available on platforms like Civ AI, particularly for public figures and celebrities.
• Vocal Identity: Short audio recordings can be used to mimic anyone's voice. This has led to instances like Tom Hanks appearing in unauthorized dental advertisements and Scarlett Johansson's voice being generated by OpenAI against her wishes.
• Real-World Exploitation: Scammers are leveraging these tools to create fake faces and voices of loved ones, such as grandchildren, to defraud seniors. Furthermore, these tools are used to generate harmful content, including "undressing apps" that superimpose AI-generated nudity onto photos of children.

Generative AI's Impact on Creative Identity

Beyond biometric data, generative AI is also undermining the identity of human creatives, which is shaped by their unique creations, imagination, and emotional expression.

• Mimicking Artistic Style: AI models are trained on artists' distributed artwork without their permission, enabling them to generate new pieces in the artists' distinct styles. Many such "style models" are available online, named after the artists they imitate.
• Devastating Consequences for Artists: Artists are experiencing job losses and income reduction, with AI mimics overshadowing their original work in search results. This has led some to quit their professions or reconsider their career paths.
• Legal and Corporate Challenges: While some artists are pursuing legal action or advocating for policy reform, these efforts are slow and face strong opposition from large corporations that advocate for unrestricted training on copyrighted content.
• The Artist's Dilemma: Artists are presented with a difficult choice:
  • Option 1: Share Artwork: This allows them to remain visible and connected but risks their work being scraped and used to train AI models that replace them.
  • Option 2: Withhold Artwork: This protects their work from exploitation but leads to disconnection from the creative community. For many, this is not a viable option.

Developing Solutions: Glaze and Nightshade

Motivated by the artist's dilemma, Dr. Jung and her team, leveraging their expertise in adversarial machine learning, developed two tools to protect visual artists: Glaze and Nightshade.

• Underlying Principle: AI models perceive artwork differently than humans. These tools introduce subtle, optimized modifications to artwork that are imperceptible to humans but cause AI models to learn incorrect information.
• Glaze: This tool specifically protects an artist's unique style. It modifies artwork so that AI models learn a significantly different artistic style. For example, a model trained on Glazed artwork might learn a style distinct from the original artist, such as that of Van Gogh.
• Nightshade: This tool focuses on protecting individual objects or copyrighted characters. The modifications teach AI models that the object is something different. For instance, a "nice shaded" image of a hat might teach an AI model that it is a round object with chocolate and strawberries.
• Accessibility and Impact: Glaze and Nightshade are released as free software tools. Since their release (Glaze in 2023, Nightshade in 2024), they have been downloaded over 10 million times by artists in more than 160 countries, serving as a first line of defense. The work has received significant media attention and recognition from both technical and artistic communities.

Broader Implications and Future Directions

The research and development of Glaze and Nightshade are ongoing efforts to strengthen these tools, improve their robustness, and explore new protective measures.

• Policy and Regulation: Dr. Jung has testified before state legislatures in California and Illinois regarding data and identity privacy, leading to the signing of regulatory bills into law in both states.
• Expanding Protective Tools: New forms of protective tools are being developed, including those designed to disrupt deepfake facial models and undressing apps. Other researchers are also contributing to this field, with work on protecting audio, voice actors, and music.
• The Question of Economic Efficiency vs. Identity: The work raises a critical question: are we willing to sacrifice our identities and those of future generations for economic efficiency?
• The Voice of the Next Generation: Interestingly, children have shown immense enthusiasm for these protective tools, viewing the defense of creativity and identity as essential for their future. This sentiment is captured by a poster made by Dr. Jung's daughter, stating, "AI don't generate my life."

Conclusion

Dr. Jung's presentation underscores the urgent need to protect human identity and creativity in the age of generative AI. By being mindful of data sharing and employing protective technologies like Glaze and Nightshade, individuals can begin to reclaim control over their digital selves and artistic expressions. The collective efforts of individuals, researchers, and policymakers are crucial to ensure that technology serves humanity without compromising its fundamental essence.