AI traces its roots back at least as far as the 1950s, when Alan Turing began exploring the concept of machine intelligence. His proposition—now known as the Turing test—was that a machine was “intelligent” if a person interacting with it could not tell whether he or she was engaging with a human or a computer. Even with several decades of AI research behind us, we have not yet arrived at a commonly accepted definition of AI. Like many of the computing waves before it, there remains a lot of hype and confusion as we try to understand its contours. For purposes of this article, AI is defined as a set of technologies that enable machine intelligence to simulate or augment elements of human thinking. AI technologies include machine learning, neural networks, probabilistic reasoning, and other “intelligent” technologies, many of which learn from experience (e.g., data), much like humans. Such technologies enable computers to simulate or replicate human cognitive capabilities, including vision, speech, knowledge, problem-solving, and other skills. Accordingly, computers, which could once only follow preprogrammed routines, are now able to interact with humans in ways that feel much more natural and responsive.
There is general agreement that certain conditions have led to the present widespread adoption of AI: the increased availability of data, the development of sophisticated algorithms, and the increase in computing power provided by the cloud. Nearly everything around us is digitized these days, and inexpensive sensors in our devices and environment collect more data all the time. AI algorithms help us sort through this massive amount of data to identify valuable information, facilitating more informed decision-making. Analyzing this data efficiently and at scale requires significant computing power, which is provided by recent advances in cloud computing. With these benefits, AI can and is producing breakthroughs in disciplines as wide-ranging as medicine, agriculture, education, and manufacturing.1
AI technologies, while encompassing a number of different techniques, often share certain characteristics. In many fields of AI, progress is happening rapidly, and today’s algorithms will be quickly replaced by better solutions. Moreover, in many cases, AI is naturally hidden from view. AI often comprises processes that reside in the cloud and are never exposed to users. For example, a translation algorithm may generate a translation from one language to another, without providing any indication about the steps that were used to produce the translation. In many cases, these decisions are relatively innocuous—translation, speech recognition, image classification, etc.—save for edge cases where particular classes of speech or images are misclassified. However, in other cases, the very nature of an algorithm making decisions is inherently controversial. Examples can include determinations about who qualifies for a loan or medical insurance, or profiling by law enforcement.
We are, however, still in the early days of AI. Current AI systems can solve narrow tasks like playing a game, returning relevant search results, or finding the quickest route to a destination. Significant technical progress is still needed to create AI systems that have more general intelligence. For example, while computers can simulate human vision, speech, and other capabilities, they have difficulty understanding the nuances of human behavior such as tone, emotion, intuition, and empathy. Moreover, while today’s AI often produces results that meet or exceed human capabilities—for example, in image recognition and speech processing—it has its limits. It is easy to find examples where AI algorithms fail. For example, image classifiers and speech recognizers both have difficulty operating in noisy environments. AI is also routinely criticized for included biases—whether they are designed into the system intentionally or inadvertently. Moreover, AI systems may be vulnerable to manipulation, as adversaries attempt to steal or alter code, or deliberately induce errors.
It is against this backdrop, in part, that trade secrets are receiving renewed attention. A trade secret comprises information that provides a competitive advantage because it is not known to others, and for which reasonable safeguards are maintained to protect its secrecy. Trade secrets are particularly suited to technologies that are not capable of independent discovery or reverse engineering, technologies that are rapidly replaced by new innovations, and technologies that cannot be described without expending significant effort, all of which are especially prevalent in AI.
While no one can seem to agree on where trade secret law originated—be it contract, tort, property, or even criminal law—trade secrets have always been important to protecting competitive business information. Trade secrets can protect a wide range of proprietary information, including product designs and technical data, as well as nontechnical information, such as customer lists, marketing strategies, and sales techniques. Almost any information can be protected by trade secret, as long as it confers a competitive advantage, although such protection does not extend to general knowledge or an employee’s skills or abilities.
Trade secrets have long been underrepresented in IP scholarship and legal protection. There is often limited discussion on trade secrets in IP textbooks, journals, and even law school curricula. Trade secrets significantly lagged behind the other IP disciplines in legal protections as well, primarily being addressed by common law and then individual state laws, while patents, copyrights, and trademarks were protected by federal statute.
However, trade secret law has been bolstered in recent years, and IP scholarship seems to be catching up as well. In the United States in particular, the 2016 Defend Trade Secrets Act (DTSA) created a federal civil cause of action for trade secret misappropriation,2 better aligning trade secrets with other IP rights. The DTSA’s federal remedy provides a more efficient and sensible way for a trade secret owner to pursue infringers, who often act across state or even national borders.
In enacting the DTSA, Congress was particularly influenced by reports that trade secret theft was reaching unprecedented levels, primarily from current or former employees. Accounts ranged from the loss of hundreds of millions of dollars to trillions of dollars, in addition to millions of U.S. jobs.3 A precise calculation of the impact of trade secret theft is difficult to determine, as it is relatively anonymous and difficult to detect; a company may not realize information has been stolen until years afterward. Moreover, it is difficult to measure the economic value of information. Finally, even if a company knows that a trade secret was stolen, accusing an employee, business partner, or competitor of trade secret theft carries reputational and relational risks of its own.4
Internationally, trade secrets are protected by article 39 of the 1995 Trade-Related Aspects of Intellectual Property Rights (TRIPS) Agreement. Although the term “trade secret” is not used, the agreement protects “undisclosed information,” which is described in terms similar to trade secrets. While World Trade Organization members are required to enact trade secret laws under TRIPS, the effectiveness and enforcement of these laws has been an ongoing challenge, particularly in large emerging economies, such as China, India, and Brazil.5 As AI companies and technologies increasingly operate across international borders, a company’s ability to protect its trade secrets may be significantly impacted by weak protection and enforcement abroad.
Most trade secrets are stolen by someone known to the company, often a current or former employee. This is particularly relevant in the age of AI, where employees are moving between companies at phenomenal rates. The nature of AI work is such that it requires employees with specialized skills, often for which there is an ongoing shortage. It is estimated that by 2020, 30 percent of technology jobs will remain unfilled due to lack of qualified candidates.6 To further complicate matters, traditional employment relationships are increasingly being replaced by alternative models, in which workers may contract with multiple companies. With stiff competition for skilled employees and frequent movement of employees between companies, it is important for companies to ensure that these employees do not take with them confidential company information. Equally important, companies need to deter new employees from using trade secrets belonging to former employers.
Trade Secrets and Other IP Rights
Trade secrets have a number of advantages, relative to other forms of intellectual property. Trade secrets generally protect broader subject matter than the other IP rights, reaching both technical and nontechnical information, ideas, and even facts, such as names and phone numbers on a client list. The information protected by trade secret need not be novel or original. Trade secrets are also protectable immediately, without the cost or lengthy registration timelines required by other forms of intellectual property. Perhaps most desirably, a trade secret’s status lasts as long as the information is commercially valuable and can be maintained as secret, unlike the limited terms that adhere to patents and copyrights.
Unlike other IP rights, however, a trade secret does not give its owner a monopoly over the subject of the trade secret. The information is only protected against misappropriation—improper acquisition, use, or disclosure. As soon as trade secret information is disclosed, whether intentionally or accidentally, its status as a trade secret is lost. Lack of registration, moreover, is not without its risks. In many cases, no formal registration means that trade secrets are rarely well-defined. Moreover, the volume of proprietary information within a company makes it difficult to inventory and properly maintain trade secrets. Finally, critics argue against the propriety of trade secrets at all, suggesting that owners should not be rewarded for keeping secret information that may not be original or innovative.
At first glance, patents and trade secrets may seem to be mutually exclusive. Patents are premised on disclosure, while trade secrets are premised on lack of disclosure. Patents and trade secrets may, however, be used in complementary ways. For example, a patent may protect a core invention, while trade secrets may protect research results, know-how, and data sets associated with the invention. Trade secrets are particularly useful before and after patent application filings. Prior to a patent application’s filing, trade secrets can protect research and development (R&D), including negative results (i.e., information that indicates or proves that a certain technique will not work). After the patent application’s filing, trade secrets can be used to protect further R&D, including any improved techniques developed after filing. To the extent that these improvements are not disclosed in follow-on patent applications, the information may be retained as trade secrets.
Patent strategy undoubtedly has effects on trade secrets. In particular, perceptions on the effectiveness of the patent system may affect the interest in trade secrets. As patents are deemed less effective, trade secrets may be seen as more attractive, and vice versa. In the present environment, in which Alice and its progeny have limited patentable subject matter, companies are considering other strategies. For undetectable AI technologies in particular, the patent bargain may not be a good one. A patent application requires that the invention be described in detail, while the company may never be able to discern whether a competitor is using the disclosed technique.
As with any technology, AI can be protected with a variety of IP assets—patents, copyrights, trademarks, and trade secrets. Optimal protection may involve pursing assets in all of these categories. Doing so allows the coverage of the broadest possible subject matter, where certain IP assets protect categories of subject matter excluded by other assets. Further, multiple asset types broaden the range of remedies available in a dispute or litigation, and provide additional protection if a particular asset is found invalid.
With a number of IP assets to choose from, when should a company select trade secret? In general, trade secrets are most suitable for technologies that cannot be reverse engineered or independently developed without difficulty, technology areas characterized by rapid development, and technologies that cannot be described without expending significant effort. For this reason, AI technologies are particularly well suited to protection by trade secret. Over time, however, the calculus may change. For example, while AI technologies may be difficult to detect today, future technological developments may aid discovery. As competitors’ abilities to reverse engineer increase, firms may choose to patent rather than retain trade secrets. This has been observed in other industries, as fundamental scientific advances have led to an observable uptick in patent filings.7 The same effect may be had as technological development in AI becomes easier to describe and as its rate of acceleration begins to slow.
How to Protect Trade Secrets
The ins and outs of protecting trade secrets are beyond the scope of this article. Briefly, however, companies must identify the proprietary information they wish to protect, and then adopt and implement reasonable measures to ensure its ongoing secrecy. Such measures may include physical barriers (e.g., fences, walls, locks, and security guards), technical features (e.g., encryption and passwords), entry and exit interviews for employees, and confidentiality agreements. Further, the information maintained as trade secret should be regularly audited and updated as new proprietary information is added and obsolete information is removed.
AI and Trade Secrets
AI and trade secrets do not operate in a vacuum as technical issues only. They exist in a legal and social environment governed by standards of fairness, privacy, and other obligations. If we can begin to see the potential implications of trade secrets in the early days of AI, we will be better able to shape those impacts with laws, regulations, and social standards.
Does Keeping AI Technologies Secret Hinder Innovation?
That trade secrets are seen by many as hampering innovation is not new to the age of AI, but it may be particularly acute for AI technologies. Trade secret law may be seen at odds with patent and copyright law, for example, which reward disclosure of innovation. Such disclosure enables others to build on or design around the innovations of others, spurring further advancement. If information is kept secret instead, this virtuous cycle is interrupted. Inventors cannot build on ideas that are hidden from view, and many companies will continue to research in the same area, often duplicating efforts. Moreover, because trade secret protection adheres immediately, e.g., to research and preliminary data, and does not require that an innovation ever be realized, valuable ideas that could lead to larger discoveries may go unshared.8 This is of particular consequence for AI technologies, for which the spread of innovation may already be difficult. AI requires significant investments in R&D, access to data sources, and the participation of highly skilled employees, all of which can limit the opportunities for significant AI advances.
However, others argue that trade secrets may actually encourage disclosure9 and, thereby, innovation. For example, the protections of trade secret law can substitute for investments in secrecy that companies would otherwise undertake. There is empirical evidence that companies overinvest in secrecy measures absent trade secret law.10 For example, in countries without strong legal protection or enforcement of trade secrets, companies tend to make business decisions that inefficiently limit disclosure, as there is no remedy to prevent the misuse of shared information. These include significant physical measures, such as walls, fences, and armed security guards. Further, companies may be less willing to contract with third parties for manufacturing or development if it requires sharing secret information, even where the third party could operate more efficiently.11 Such restrictions on the flow of information between potential business partners slow down innovation and commercialization.
There are concerns that even companies that own trade secrets suffer from reduced innovation. For example, if employees’ rights are—or appear to be—overly restricted relative to assets developed for the company, employees may be less motivated to be more innovative. In fact, employees may be less likely to work for a company at all if its IP policies, including those regarding trade secrets, are deemed overreaching.
Can AI Workers Move Freely between Employers?
It is simple to say that a company’s proprietary information should stay with the company when an employee leaves, but that the employee should be able to take his or her skills and abilities with him or her. However, it can be difficult to determine where to draw this line. Company proprietary information is often intertwined with an employee’s knowledge and skill. Certainly not all valuable information learned on a job is protectable by trade secret. For example, to the extent that the employee’s knowledge and skills are similar to those possessed by others working in the industry, or would be generally known or ascertainable to competitors, they would not be protected by trade secret. Should the employee not be able to take some portion of the knowledge he or she acquires during his or her job to a new employer, the employee would effectively be barred from taking a new job in the industry in which he or she is most skilled and is best able to earn a livelihood.12 Employers, too, have an interest in allowing employees to move freely within an industry. The ability to obtain skilled employees and immediately put them to work on the latest developments is critical to a company’s competitive advantage.
Employer-employee relationships are addressed in part by the DTSA. The DTSA requires that any conditions imposed on new employment must be related to trade secret misappropriation; mere personal knowledge does not suffice.13 Moreover, relief granted under the DTSA cannot conflict with state laws promoting the lawful practice of a profession or trade.14 Of course, trade secret law is not the only body of law governing employee mobility. Employees and employers must also adhere to laws around invention assignments, work-for-hire arrangements, and noncompete and nonsolicitation agreements, among others.
All of these considerations support a strategy of well defining trade secrets within a company. It serves the interests of employers, so they can retain the commercial benefit from proprietary information while defining which information employees can otherwise take with them when they leave. Even so, trade secrets still pose some limits on job seekers. Employees may be prevented from showing prospective employers certain work performed at a prior job. Employers may even decline to hire certain employees or those from particular competitors who might place the employer at greater risk for allegations of trade secret misappropriation.
Does Secrecy of AI Comply with Standards of Transparency and Fairness?
AI poses unique challenges, as it informs and sometimes replaces human decision-making. AI guides decisions on loan eligibility, insurance coverage, medical procedures, and other significant issues. On the one hand, removing or reducing the elements of human decision-making, such as bias and errors, leads to more objective decisions. However, algorithms are not inherently more fair than human decision-makers. AI often depends on training data to arrive at a decision model. Such training can lead to biased results if the training data itself incorporates bias or prejudice, whether intentional or unintentional. Algorithms also depend on choices made by developers, for example, about which features should be included in decision-making. The inclusion or exclusion of certain features may lead to discriminatory results.
While businesses have legitimate interests in protecting proprietary information, such as the method by which a decision is made, individuals also have legitimate rights to know that AI algorithms are created and applied in a fair manner. For example, the Fourteenth Amendment requires procedural fairness for decisions made by government agencies, such as which taxpayers should be audited. In the AI context, that may mean that the same algorithm was applied to each individual and that it was not designed in a manner that disadvantages a particular group. Moreover, a growing body of privacy laws, most notably the European Union’s General Data Protection Regulation, require transparency about the collection, use, and storage of data. Transparency is not just an abstract value; it has practical consequences. If people do not understand, at least in a general sense, how AI makes predictions or recommendations, they will be less likely to trust or use AI.
Where legal or social standards require transparency, how does a company comply while protecting its proprietary information? Most agree that requiring a company to disclose its source code is neither necessary nor sufficient to demonstrate transparency. It is often difficult even for technical experts to understand how source code will behave. Even if it could be understood, many AI algorithms change over time as they are applied to specific data and learn from that data. Accordingly, algorithmic rules may be obsolete by the time they are analyzed.
Technical solutions do exist, however, to document where training data was obtained, identify its characteristics, and determine whether it is adequately representative. Further, there are techniques for verifying that AI systems behave as expected. Human analysis can also help identify where biases and blind spots may exist.
While transparency may be desirable in certain cases, widespread transparency comes with additional concerns. For example, compelling a business or agency to publicize decision-making rules, such as a method for determining when an individual should be audited, may allow strategic gaming of the system.15
We live in a time of exciting opportunity, as AI helps us make better use of our time and resources, contributing to advances that have never before been possible. Trade secrets are well suited for protecting these technologies, but their use has implications for innovation, transparency, employment, and other legal and social issues. One thing is for sure: as AI and trade secrets continue to evolve, the participation of lawyers who are skilled in both the technical and social sciences will be essential.
1. See, e.g., Microsoft, The Future Computed: Artificial Intelligence and Its Role in Society (2018), https://news.microsoft.com/uploads/2018/02/The-Future-Computed_2.8.18.pdf.
2. 18 U.S.C. § 1836.
3. John Cannan, A (Mostly) Legislative History of the Defend Trade Secrets Act of 2016, 109 L. Libr. J. 363, 365 (2017).
4. Brian T. Yeh, Cong. Res. Serv., R43714, Protection of Trade Secrets: Overview of Current Law and Legislation 13–14 (2016), https://fas.org/sgp/crs/secrecy/R43714.pdf.
5. Id. at 11.
6. Vikram Bhalla et al., Twelve Forces That Will Radically Change How Organizations Work, Bos. Consulting Group (Mar. 27, 2017), https://www.bcg.com/en-us/publications/2017/people-organization-strategy-twelve-forces-radically-change-organizations-work.aspx.
7. Petra Moser, Why Don’t Inventors Patent?, Paper Presented at the Research Symposium on Property Rights Economics and Innovation (Nov. 14, 2008), https://www.researchgate.net/publication/5188211_Why_Don’t_Inventors_Patent.
8. Michael P. Simpson, The Future of Innovation: Trade Secrets, Property Rights, and Protectionism—an Age-Old Tale, 70 Brook. L. Rev. 1121, 1152–55 (2005).
9. See, e.g., Mark A. Lemley, The Surprising Virtues of Treating Trade Secrets as IP Rights, 61 Stan. L. Rev. 311 (2008).
10. Robert M. Sherwood, Intellectual Property and Economic Development (1990).
11. Lemley, supra note 9.
12. Dan L. Burk & Brett H. McDonnell, The Goldilocks Hypothesis: Balancing Intellectual Property Rights at the Boundary of the Firm, 2007 U. Ill. L. Rev. 575, 592.
13. 18 U.S.C. § 1836(b)(3)(A)(i)(I).
14. Id. § 1836(b)(3)(A)(i)(II).
15. Joshua A. Kroll et al., Accountable Algorithms, 165 U. Penn. L. Rev. 633, 657–58 (2017).