Policymakers are currently addressing crucial issues concerning the data utilized to train AI models and intellectual property (IP) rights as the application of artificial intelligence (AI) continues to expand. The most extensive datasets available today consist of over a billion instances of human-generated text, representing the epitome of high-quality information. Presently, cutting-edge conceptual AI models are trained on vast volumes of data. Despite this, content creators have initiated multiple legal actions focusing on the online implications of deploying conceptual AI models to copyrighted content due to apprehensions regarding potential copyright infringements. While significant rulings are pending, a failure by the courts to recognize this as “fair use” of the material could hinder AI progress in the United States.
To strike a balance between safeguarding copyrights and meeting the demand for more data to educate AI systems, Congress and the Biden Administration have commenced exploring potential solutions. Policymakers should evaluate market-oriented solutions and observe how different countries have tackled these challenges as they craft their responses. For example, the European Union, Israel, and Japan have established frameworks to enhance accountability concerning the types of data employed by AI while concurrently upholding the use of copyrighted material for AI training. Similarly, companies like Adobe have introduced a feature enabling content creators to decide whether their data is utilized for AI advancement, potentially obviating the necessity for substantial governmental interventions.
This discourse delves into the education of various AI models, the intersection of current copyright laws with AI model training, and the broader legislative and regulatory landscape surrounding AI. It also examines how pivotal international and administrative decisions have influenced the development and utilization of AI in the United States.
Education on Legal Implications and Copyrighted Content
Education data, denoting the raw information utilized by models to make decisions, serves as the primary input for AI models. Data instruction manifests in diverse forms, such as images and videos crucial for enabling self-driving vehicles to recognize traffic signs and conversational interactions pivotal for enhancing customer service chatbots. While the quantity of training data is significant—top-tier models often leverage more than 45 terabytes (TB) of data—the quality is equally paramount, leading developers to incorporate datasets containing copyrighted material. For instance, books, valued for their extensive content diversity, are indispensable for text-generation models like ChatGPT. Presently, open-source datasets encompass nearly all published books, offering a readily accessible option for AI training. This practice extends beyond textual data to encompass music, videos, and images, all of which are instrumental in training AI models.
The methodology of utilizing copyrighted material for training has sparked considerable debate. Content creators have already initiated legal actions citing potential copyright violations. Notably, writer and comedian Sarah Silverman filed a lawsuit against OpenAI and Meta in July, alleging severe rights infringements by using her protected work without consent. Another lawsuit filed in late September raised similar concerns, contending that OpenAI’s Big Language Model ChatGPT replicated and disseminated copyrighted materials without authorization. The core argument posits that this training violates the Copyright Act due to the disapproval of rights holders.
The application of the “fair use” doctrine, a fundamental defense for AI developers, is poised to shape the outcomes of these legal disputes. Essentially, fair use allows for limited and transformative use of copyrighted material. Courts consider four factors when determining the applicability of a fair use defense:
- The purpose and nature of the use, whether for commercial or educational objectives.
- The characteristics of the copyrighted work.
- The proportion of the copyrighted material used.
- The impact of the use on the potential market for the copyrighted work.
Both AI developers and rights holders are grappling with uncertainty, given the absence of definitive rulings on the validity of fair use in AI model training. Concerns about AI’s potential to supplant artists and reshape creative industries have been raised by experts and creatives, underscoring the disruptive influence on creative markets. Conversely, AI holds significant promise in fostering artistic growth and accelerating creative industries. As authorities navigate these complexities in the realm of AI data training, it might be prudent for Congress to address AI and copyright issues specifically to furnish courts and relevant stakeholders with clarity, as advocated in a letter published by Creative Comments and endorsed by artists.
Looking ahead
As Congress engages with pertinent agencies and contemplates legislative solutions, lawmakers could draw insights from international counterparts and private industry to shape the future landscape. The U.S. Copyright Office, on August 30, issued a formal Notice of Inquiry soliciting input to “inform Congress” on potential future courses of action. The agency has embarked on an examination of “copyright law and policy issues raised by artificial intelligence technology.” Congressional hearings, focusing on rights and AI in academia, have been convened by the Senate Judiciary Committee. Moreover, the Federal Trade Commission is actively engaging with designers and creatives through roundtable discussions to delve into the diverse impacts of AI on their respective domains. A bipartisan group of lawmakers has deliberated on legislation addressing online content replication, underscoring the imperative of safeguarding individual creators. Notably, during a recent Senate Judiciary Subcommittee hearing on IP, a witness highlighted the adverse effects AI could have on individual creators, signaling a potential legislative response.
While formal actions in the U.S. are pending, regulators can glean insights from global regulatory frameworks and market-driven self-regulation tools on striking a balance between fostering AI innovation and upholding copyright protections. Diverse approaches adopted by international counterparts can offer valuable lessons for lawmakers. For instance, the European Union’s AI Act mandates transparency from model developers in disclosing the copyrighted materials used for training data, empowering artists and copyright holders to assert control over their works and demand compensation for their utilization. However, challenges persist in determining individual contributions to large training datasets and corresponding compensation. The Bipartisan Framework on Artificial Intelligence Legislation, advanced by Senators Hawley and Blumenthal, underscores the potential requirement for developers to disclose critical information about training data, limitations, accuracy, and safety of AI models to users and stakeholders, indicating congressional deliberation on related measures.
In contrast, Japan and Israel have diverged from the EU’s approach by permitting machine learning entities to train AI systems with copyrighted materials sans authorization, often prioritizing AI advancement over owners’ rights. Notably, the Chinese government has expanded fair use provisions for educational models focused on specific research areas or audio-visual applications, distinguishing between training models on copyrighted materials and the resultant output generated by a trained model. Israel’s Ministry of Justice has recently broadened fair use to encompass loosely training AI models, albeit with specified limitations, such as training a model solely on a single artist’s works. Japan and Israel aim to expedite AI development within their jurisdictions through these policy stances.
Lastly, private sector initiatives could obviate the necessity for substantial governmental interventions. Companies are exploring tools that enable creators to prevent AI firms from utilizing their works. Adobe’s initiative, allowing creators to label their data as “do not train,” signals a proactive stance in empowering content creators to safeguard their intellectual property. Researchers are also exploring “data poison attacks” that could disrupt image-generating AI models by introducing tainted images into training sets, while industry players are developing education blockers to prevent data scraping. Notably, OpenAI and a web3 protocol group have introduced tools to prevent the inclusion of individual data in training sets.
While the development of AI models hinges on leveraging copyrighted materials, policymakers must navigate the delicate balance between fostering innovation and respecting intellectual property rights. Drawing insights from global approaches and private sector innovations can inform lawmakers’ decisions on clarifying rights protections for materials utilized in AI model training, ensuring a harmonious coexistence between AI progress and copyright preservation.
