How are IP Protections Changing in a Global AI Landscape?
The second article in this three-part series. koralli unravels the complex relationship between artificial intelligence (AI) and intellectual property (IP), exploring what it means for you and your business. We’ll guide you through the various intersections between these two resource domains, building a comprehensive understanding of the challenges and opportunities that lie ahead. By the end, you'll be equipped with six actionable strategies to minimise risks associated with Generative AI and optimise your IP assets in this rapidly evolving landscape.
Photo by Google DeepMind by Pexel
How are IP Protections Changing in a Global AI Landscape?
Whilst this diagnosis is increasingly recognised—that current LLM and GAI training methods may infringe upon large swathes of intellectual property, as discussed in part one of this series—the path to a solution remains elusive. The risks to developers, businesses, and end-users are still not fully understood, and as new advances in AI are made daily, emerging uncertainties must be navigated.
Government & International AI-Regulation
The issue of IP infringement in AI training is a thorny challenge that extends beyond national borders. Understanding global regulation trends is the critical next step in securing your business interests and ensuring client security. International legislation currently presents a spectrum of regulatory approaches, ranging from strict controls and classification at one end to more flexible voluntary guidance at the other.
Photo by Darya Sannikova from Pexels
The Spectrum of Regulatory Jurisdictions
On the stricter end of the spectrum, the EU and China are leading the way. In China, a series of major recommendations and regulatory drafts published between 2021 and 2023 have sounded the starting gun initiating tightening controls on LLM ingestion. Notable actions include mandating that companies notify users when AI algorithms are in operation, requiring visible labelling of synthetically generated content (Deep Synthesis Regulation, 2022), and enforcing participation in China’s algorithm registry (Sheehan 2023). Meanwhile, in Europe, the European Parliament passed a draft of the EU’s AI Act in June 2023. The Act classifies AI tools into to four risk levels, each with corresponding obligations related to security, transparency, and accountability. The draft also proposes protections for rights holders, requiring all GAI models to publish summaries of copyrighted data used in training.
At the opposite end of the spectrum are the more flexible regulatory approaches, currently most notably adopted by the US and Japan. These strategies aim to foster AI innovation by leveraging the flexibility of voluntary guidance. For example, Japan’s 2017 amendment to its Copyright Act specified that the downloading or processing of data via the internet for AI development does not constitute copyright infringement. Such light-touch approaches to data ingestion stand in stark contrast to jurisdictions focused on IP protection, and pose significant challenges to businesses that rely heavily on the UK’s robust IP export market.
UK Policy Position
In the context of the international landscape, the UK’s AI strategy appears to fall short of the global standards. The AI white paper (2023) advocates for a voluntary approach to regulation, a stance that is increasingly out of step with the EU’s trajectory and in conflict with the UK's obligations under the Berne Convention and the Data Free Flow with Trust (DFFT) initiative. This legislative inertia places strain on the rapidly growing £335 million Text and Data Mining (TDM) licensing market, which currently operates at the intersection of publishing interests (e.g., datasets, scholarship, long-form content, etc.) and AI platforms.
Photo by Suzy Hazelwood from Pexels
Where the white paper does align with the EU is in its call for a regulatory framework for high-risk AI systems and the requirement for companies to conduct risk assessments before deploying AI tools. These provisions offer some reassurance to businesses, allowing them to plan ahead accordingly. However, it’s important to note that without formal legislation, these principles remain unenforceable. The potential impact of the European Parliament's AI Act on UK businesses cannot be overlooked, given the bloc’s geographic proximity and significant market influence. As a result, UK businesses are strongly advised to formulate their AI/IP strategies with EU policies in mind as a prerequisite for long-term success.
Legal Challenges
The profile of AI companies potentially breaching IP law has risen significantly over the last few years, largely due to high-profile legal challenges. Two major cases that are poised to influence the future trajectory of IP legislation are outlined below:
Andersen v. Stability AI et al. (2022)
In Andersen v. Stability AI et al. (2022), three illustrators sued multiple GAI platforms—Midjourney, DeviantArt Inc., and Stability AI—for training their models on the artists’ original works without proper licensing (Chen 2023). If the courts find that these platforms produced works that are insufficiently transformative, they could incur substantial infringement penalties.
Getty Images Inc. v. Stability AI Inc. (2023)
Getty Images Inc. v. Stability AI Inc. (2023) is another pivotal example that could set a precedent for future disputes. Getty, a leading image licensing service, accused Stable Diffusion (Stability AI) of the unlicensed ingestion of over 12 million photos to train their LLM (Brittain 2023).
Underpinning these and many other cases are the semantics of IP law, particularly the interpretation of the ‘fair use’ doctrine. Although the specific implications of fair use vary between US and UK jurisdictions, its overarching purpose is to permit, in certain circumstances, the use of copyrighted work without the need for rightsholder licensing. These exceptions include criticism (including satire), commentary, news reporting, teaching, scholarship, or research (Appel et al.).
Equally pivotal to these legal debates are the concepts of ‘derivative work’ and ‘transformative use,’ which help define what constitutes new content. ‘Transformative use’ is likely to be a key defence strategy in AI companies’ legal arsenal. This is the same principle Google successfully employed to defend its Digital Library and against the Authors Guild in 2015 (Merkley 2023).
Advice For Business: mitigating liability for IP infringement
As regulation remains in flux, reaching a global consensus on AI’s impact on intellectual property rights (IPRs) seems a distant prospect. The current landscape is marked by two competing schools of thought: one advocating for stronger AI controls to safeguard IP exports, and the other pushing for relaxed, voluntary regulation to maximise AI’s potential. As new protections largely hinge on the outcomes of a few key legal challenges, businesses must take proactive steps to mitigate their risk of unintentional IP infringement.
Consequently, it is critical that businesses implement strategies to assess and reduce their risk of unintentional infringement on others’ intellectual property. Below are six key areas where your business may be vulnerable to IP infringement while using GAI applications, along with recommendations on how to mitigate against them:
1. AI-Training Practice Risks
Businesses employing GAI services should thoroughly understand the training methods and data used by the platforms they rely on. This practice is essential for navigating the legal risks associated with AI use. If an AI platform is deemed in breach of a rightsholder’s IP, your business could be liable for ‘willful infringement,’ leading to steep financial penalties. Additionally, in the near future, insurance requirements may include a comprehensive breakdown of AI in use and an audit of its IP contributors (Appel et al.).
2. Content Creation Risks
The use of AI-generated content in commercial work currently risks contravening ‘derivative use’ precedents. Until legal cases such as Getty Images Inc. v. Stability AI Inc. (2023) and Andersen v. Stability AI et al. (2022) establish that GAI content is sufficiently ‘transformative,’ can businesses should proceed with caution. Publishing AI-generated content without legal clarity could result in IP infringement liabilities, including penalties for "willful infringement," even if training data practices are eventually settled.
Photo by Google DeepMind from Pexels
3. Code Generation Risks
When using GAI to produce code, software designers should closely examine the model in use, particularly focusing on its training methods. If the model uses code with attached licences, you should ensure that the necessary licences are purchased and that proper credit is given to the original source. Similar to content creation risks, if the code generated fails to significantly ‘transform’ the data used in the model’s training, you may be liable to infringement penalties.
4. Demand Terms of Service from GAI platforms
Ensure these terms of service confirm official licensure of the ingested training data. Additionally, due to the inherent difficulties in tracing data provenance, it is advisable that these terms include broad compensation for potential IPR short fallings. If AI platforms cannot provide training audits for IP-protected data or prove that datasets are subject to open-source licences, it is highly advisable to cease using their tools (Appel et al.).
5. Trade Secret Leaks
Depending on the platform, some GAI models ingest user input data, which can then be used in further model training, either immediately or at a later date. This poses a significant risk that your data could be reproduced elsewhere for other users. For instance, OpenAI’s terms of service state that their LLMs are developed using three primary sources of information, including “information that our users or our human trainers provide” (Open AI, n.d.).
Even the API, considered the commercial and “safer” version of ChatGPT, carries risks:
“….OpenAI may securely retain API inputs and outputs for up to 30 days ….After 30 days, API inputs and outputs are removed from our systems, unless we are legally required to retain them.
…access to API business data stored on our systems [includes] … specialised third-party contractors (Open AI, n.d.).”
In other words, the platform has the right to access, analyse, and, in some cases, share input data with third parties (Norton Rose Fulbright 2023; Ioet 2023).
Photo by Nejat Gunduc from Pexels
A clear example of these risks occurred in 2023 when Samsung Semiconductor employees unintentionally disclosed confidential information through interactions with ChatGPT. Due to the platform’s training methods, which were learned from prompts and text inputs, this information was potentially accessible to other users. In response, Samsung implemented a new policy banning the use of generative AI tools, including Google Bard and Bing, over concerns that these platforms were storing data on external servers (Gurman 2023).
The landscape is evolving rapidly. As of October 2023, Microsoft researchers Ronen Eldan and Azure’s Mark Russinovich claimed to have successfully modified the Llama 2-7b model to selectively unlearn information, using the Harry Potter books as a test case. If this approach proves scalable, it could significantly reduce the risks businesses face from leaked trade secrets via AI generation tools (Eldan and Russinovich 2023).
6. Patenting Risks
Under current US patent law, inventions where AI is the only named inventor are not eligible for patent protection. A 1993 ruling established that ‘only natural persons can be ‘inventors’,’ a principle reaffirmed in Thaler v. Vidal (2022) (Schneider 2023). This leaves the legal interpretation of patent eligibility, ownership, and inventorship in flux. Until these nuances are clarified, inventions developed using AI-generation tools remain vulnerable to having their patents invalidated.
This issue is particularly relevant to the pharmaceutical industry, especially in the area of drug design. AI has been on the biotech agenda for many years, but recently, major industry players like Moderna and GSK have publicly emphasised their intent to leverage AI capabilities. As these developments unfold, this will be an area to watch closely (The Economist 2023).
In the final part of this series, we will outline six strategies to prepare your business for these challenges and capitalise on the opportunities presented by the AI-IP intersection. These actionable steps will help ensure your business stays ahead in an evolving AI landscape.
Links to articles 1 and 3 in the series:
Sources [Accessed August 12, 2024]:
Appel, Gil, Juliana Neelbauer, and David A. Schweidel. 2023. “Generative AI Has an Intellectual Property Problem.” Harvard Business Review.
https://hbr.org/2023/04/generative-ai-has-an-intellectual-property-problem.
Brittain, Blake. 2023. “Getty Images lawsuit says Stability AI misused photos to train AI.” Reuters. https://www.reuters.com/legal/getty-images-lawsuit-says-stability-ai-misused-photos-train-ai-2023-02-06/.
Chen, Min. 2023. “Artists and Illustrators Are Suing Three A.I. Art Generators for Scraping and 'Collaging' Their Work Without Consent.” Artnet News. https://news.artnet.com/art-world/class-action-lawsuit-ai-generators-deviantart-midjourney-stable-diffusion-2246770.
The Economist. 2023. “Big pharma is warming to the potential of AI: But some worry the Terminator is coming.” The Economist. https://www.economist.com/business/2023/07/13/big-pharma-is-warming-to-the-potential-of-ai?utm_medium=cpc.adword.pd&utm_source=google&ppccampaignID=18156330227&ppcadID=&utm_campaign=a.22brand_pmax&utm_content=conversion.direct-response.anonymous&gad_source.
Eldan, Ronen, and Mark Russinovich. 2023. “Who's Harry Potter? Making LLMs forget.” Microsoft. https://www.microsoft.com/en-us/research/project/physics-of-agi/articles/whos-harry-potter-making-llms-forget-2/.
Gurman, Mark. 2023. “Samsung Bans Generative AI Use by Staff After ChatGPT Data Leak.” Bloomberg UK. https://www.bloomberg.com/news/articles/2023-05-02/samsung-bans-chatgpt-and-other-generative-ai-use-by-staff-after-leak?leadSource=uverify%20wall.
Ioet. 2023. “Avoid Exposing Sensitive Data to ChatGPT: Tips and Tricks for Safe AI Interaction.” LinkedIn. https://www.linkedin.com/pulse/avoid-exposing-sensitive-data-chatgpt-tips-tricks-safe-ai-interaction/.
Merkley, Ryan. 2023. “On AI-Generated Works, Artists, and Intellectual Property.” Lawfare. https://www.lawfaremedia.org/article/ai-generated-works-artists-and-intellectual-property.
Norton Rose Fullbright. 2023. “Everyone is using ChatGPT: What does my organization need to watch out for?” Norton Rose Fullbright. https://www.nortonrosefulbright.com/en-gb/knowledge/publications/f2457585/everyone-is-using-chatgpt-what-does-my-organisation-need-to-watch-out-for.
Open AI. n.d. “How ChatGPT and our language models are developed.” OpenAI Help Center. Accessed August 12, 2024.
https://help.openai.com/en/articles/7842364-how-chatgpt-and-our-language-models-are-developed.
Schneider, Jacob W. 2023. “The Coming Shift from Patent to Trade Secret Protection for Generative AI Inventions.” Holland & Knight. https://www.hklaw.com/en/insights/publications/2023/12/the-coming-shift-from-patent-to-trade-secret-protection-for-generative.
Sheehan, Matt. 2023. “China's AI Regulations and How They Get Made.” Carnegie Endowment for International Peace. https://carnegieendowment.org/2023/07/10/china-s-ai-regulations-and-how-they-get-made-pub-90117.