Convening at the White House last July, seven leading artificial intelligence (AI) companies made a series of voluntary commitments to “move toward safe, secure, and transparent development of AI technology.”1 Among the eight commitments was a promise to invest in developing mechanisms, like watermarking, to label AI-generated content.2 While the White House categorized this commitment as a consumer protection measure,3 AI companies could stand to benefit too. If AI companies commit to watermarking (or are required to watermark) AI-generated works, they may argue that they ought to also receive the copyrights to those outputs by default. Although it may not have been the White House’s explicit intention, this copyright framework could be a good thing. If such a scheme were implemented, the value of those copyrights could, in turn, incentivize companies to continue developing and maintaining effective watermarking tools. In this way, the White House commitment — if taken seriously — could accomplish a subtle trade: exchanging the economic value of copyrights for accurate and reliable AI identification.
The commitments are part of the Biden-Harris Administration’s initiative to “seize the tremendous promise and manage the risks” of AI,4 a novel technology with broad ramifications. The commitments center on three principles of safety, security, and trust,5 and reflect an ever-growing interest in the governance of artificial intelligence.6 In its statement announcing the commitments, the Biden-Harris Administration also highlighted its broader approach to safe and responsible AI development.7 The statement described a variety of meetings that the President and Vice President had convened with AI companies, researchers, and stakeholders; noted the publication of a framework for AI rights;8 and detailed a recent Executive Order9 that addresses algorithmic bias in technologies including AI.10 These initiatives suggest a swing toward a proactive regulatory response to AI that’s unlike the more reactionary measures that have been directed toward other digital-era developments, like social media.11
However, in part due to the disconnect between fast-paced technical development and slower-paced government action,12 early attempts to regulate AI have thus far failed to bring down the hammer, at least in the United States.13 The new White House commitments have already received criticisms to this effect.14 Commentators have critiqued the commitments as vague, “sensible-sounding pledge[s] with lots of wiggle room”15 — pledges, in other words, that don’t actually require meaningful action from the companies. The commitments are not “backed by the force of law” and have no accompanying enforcement mechanism.16 The lack of accountability metrics also effectively takes the pressure off companies to solve difficult technical challenges, like detecting AI-generated outputs after they’re released to the public.17 These critiques — and the fact that many of the commitments are only iterations of precautions already taken by AI companies18 — suggest that the White House commitments ultimately lack teeth.
However, while the related White House press release does not offer many specific solutions on how to achieve the commitments, the fifth commitment — to identify AI-generated content — is a notable exception: it specifies watermarking, in particular, as a potential identification mechanism.19 Unlike tools that allow viewers to discern whether something was AI-generated after it’s been created and shared,20 watermarking occurs at the point of generation — that is, the AI companies themselves, rather than downstream viewers, take responsibility for authentication. Watermarking is a concept initially found in copyright law and is particularly important in the digital sphere.21 It acts as a type of “copyright management information,” which can be used to identify a copyrighted work,22 and under the Digital Millennium Copyright Act,23 removing or altering a watermark without the copyright owner’s consent is prohibited.24
Given this connection between watermarking and copyright, the commitment to develop technical mechanisms to label AI-generated content may open the door for AI firms to secure economic value through copyrights, which could have a meaningful and positive effect. Legally, who owns AI-generated works is an unsettled question.25 But watermarks, as a form of copyright management information, are a typical sign of copyright ownership.26 The established connection between watermarking and copyright, backed by momentum from the White House commitment, could lay the foundation for a legal regime in which AI companies retain ownership of copyrights to AI-generated works, at least as a default matter. Clarifying AI ownership could be beneficial: it may provide a more administrable framework for judicial oversight, incentivize the implementation of effective identification mechanisms for AI-generated works, and lay the groundwork for more robust regulatory schemes in the future.
The question of who will own the copyrights to AI-generated content remains open to the courts, but commitments to add watermarks could impact the debate because watermarks are already understood as a type of copyright management information. An intuitive inference is that if watermarks symbolize ownership and if companies must authenticate and watermark content at the time it’s generated, then the companies have a claim to copyright ownership for that content. In fact, if the courts already associate watermarking with copyright ownership, it would be unusual to award copyright ownership to someone other than the entity responsible for creating the watermark, especially given the associated research and development costs of watermarking technology. To the extent AI-generated works are copyrightable,27 the White House commitment regarding watermarking may ultimately suggest ownership rests with the AI companies.
Even small suggestions like this could have a significant impact because the relationship between AI and copyright law is already live in the courts.28 Painting in broad strokes, copyright concerns about AI fall into one of two buckets: (1) whether AI companies infringe on existing copyrights by training their models on copyrighted material; or (2) whether the content generated by AI models is copyrightable and, if so, to whom those copyrights belong.29 The latter of these, the “output” question, is familiar. The Supreme Court took up a similar question in 1884 when it decided in Burrow-Giles Lithographic Co. v. Sarony30 that photographs were constitutionally eligible for copyright.31 At the time, people questioned whether photographs produced by photographers were similar to “writings” produced by “authors,” protectable under the Intellectual Property Clause.32 Ultimately, the Court took an expansive reading of both terms and determined that “[t]he only reason why photographs were not included in the [Copyright A]ct of 1802 is, probably, that they did not exist.”33 The parallels between photography and artificial intelligence are limited, but, at its core, the Court’s reasoning in Burrow-Giles suggests that copyright law is meant to adapt to new technologies and that the scope of copyright law can be expanded. The question today is how that should happen with respect to AI.
Policywise, a straightforward copyright scheme that incentivizes the proper identification and moderation of AI-generated content could solve real issues for AI regulation. Currently, concerns about unidentified AI-generated images, which can be “deepfakes” or otherwise manipulated, are particularly acute.34 At present, it’s not clear who owns AI-generated content,35 who is liable for AI-generated content,36 and what content exists in the public domain.37 With respect to something like a deepfake, then, it’s not clear who can be held responsible if, for example, someone’s likeness is used without their consent. Other complicated questions, like what happens if a model generates identical outputs for different users invoking different prompts,38 also remain unanswered. Not only do these open questions leave creators and developers in the dark with respect to their own protections,39 they also contribute to systematic inefficiencies within the broader copyright regime, as the legitimate reuse of potentially copyrighted works is made difficult by challenges in tracing the provenance of AI-generated works.40
Defaulting copyrights to AI companies may be a desirable option, in large part because it’s an administrable standard for courts. There are several ideas about how and to whom the copyrights for AI-generated content might be allocated.41 AI users have made claims to copyright ownership, arguing that prompting an AI model is sufficiently meaningful creative input to satisfy the requirements for authorship defined by the Copyright Act.42 However, the Copyright Office’s Zarya of the Dawn decision highlights the problem with this approach: it effectively requires that creators litigate the extent of their creative efforts.43 Professor Lawrence Lessig has argued that such an approach is ultimately worse for AI users and creators because it only further complicates copyright law, such that “copyright itself [becomes] the right to hire a lawyer.”44 By contrast, a system that defaults copyrights to AI companies could establish an industry standard that avoids complicated line-drawing exercises about the value of the creative input in any given AI-generated work.45
Defaulting copyrights to AI companies could also incentivize the implementation of effective identification mechanisms for AI-generated works. With the economic value of copyrights on the table, AI companies would have an incentive to effectively identify the content generated on their platforms, even if they later chose to contract around the default copyright delegations.46 Front-end identification matters because future regulations may require a reliable inventory of AI-generated works — having accurate identification mechanisms in place would significantly ease the administration of future regulatory schemes. Defaulting copyrights to AI companies could also help ensure that companies remain economically invested in maintaining proper identification mechanisms, despite any potential legal risks associated with claiming user outputs. In this way, the White House commitment could serve as a foreground to a more robust and efficient copyright system for AI-generated content in the future.
However, it’s difficult to predict precisely how this incentive structure would intersect with other corporate priorities. Companies may decide that the value of copyrights doesn’t justify assuming potential liability for the range of outputs that users create on their platforms. This risk would disincentivize companies from accurately identifying AI-generated content, and if this happened, the watermarking commitment would be, as the critics suggest, toothless. But companies are already pursuing identification mechanisms pursuant to the White House commitments.47 For example, Google recently rolled out a new tool called SynthID for watermarking content generated on its photorealistic text-to-image model, Imagen.48 Google’s early movement in this space is evidence that at least one major AI player has done the calculus and determined that investing in watermarking tools is worth potentially exposing their platform to new risks. Historically, it has paid off for tech companies to engage with regulators on the front end,49 and regulators may think about paying out — with copyrights as currency — once again. Failing to capitalize on momentum generated by the White House commitments could be a missed opportunity to promote a scheme that is administrable and mutually beneficial to users and companies alike.
That said, receiving all copyrights may be too big of a “reward” for AI companies’ efforts to identify and label AI-generated content, especially when such efforts could be statutorily required without exchanging anything in return. The underlying fear is that AI companies will be unduly compensated for the creative efforts of AI users and that creators and artists will suffer as a result.50 This concern is legitimate, especially because ensuring that creators are fairly compensated is one of the functions of copyright law generally.51 However, a straightforward default system does not necessarily disincentivize or devalue creation in copyright industries. To the contrary, a transparent copyright regime for AI-generated works can contribute to a well-functioning market for the copyrights themselves by ensuring that AI-generated works continue to be identified as AI-generated when they are reused and reproduced.52 AI companies can still contract away their copyrights, and the market for copyrights may in and of itself incentivize users to choose one platform over another, but establishing a transparent system that promotes effective identification of AI-generated content is the first step in creating a market that works for creators at all.53
Copyright law has always been about incentives.54 Exchanging the responsibility of AI identification for the value of AI copyrights would fit squarely within this system. The best way to cement these incentives would be to amend U.S. copyright law to enforce the watermarking requirement with the accompanying default of copyright ownership for AI companies. But by securing the initial commitment of AI companies to develop identification mechanisms, like watermarking, the White House may have tilted the debate regarding AI-copyright ownership toward AI companies and opened the door for this mutually beneficial trade to materialize.