I. Contesting AI Public Transparency
The growing ubiquity and impact of artificial intelligence (AI) systems in many aspects of ordinary life has, after years of simmering public awareness, become a matter of widespread concern. While increasingly reliant on AI-driven services, humans are also frequently anxious about the risks of AI-caused harms and their diverse individual and collective impacts. Those concerns have helped to make AI-system transparency a foundational element in efforts to increase the scrutiny, contestability and accountability of AI.Footnote 1
These concerns are usefully illustrated by the growing impact of AI on recruitment and employment. In this essential sphere of human life, AI systems are not only emergent tools for work or co-working, but are also tools or agents for employee management and, increasingly, as competitors for work.Footnote 2 This is a transformation not just of human work, but also of relations between human and machines within work, in which humans are quantifiable objects of machine attention. In recruitment, for example, AI evaluation and filtering of applicants, which is already based on qualifications and psychometrics, is likely to extend further into AI-inferred personal attributes, such as honesty, commitment and attentiveness.Footnote 3
Undoubtedly, humans will often enjoy the benefits of AI in the workplace, despite being aware that AI systems also carry risks of harm to individuals and to workforces generally.Footnote 4 Aside from the consequences of AI-system errors and security breaches, potential risks for job applicants or employees include disinformation, discrimination, deceptive manipulation as well as intimidation or even coercion.Footnote 5 Interaction with AI systems through recruitment or employment may also restrict human capacities or opportunities for autonomy.Footnote 6
In view of these risks to individuals, there are also collective concerns about risks of systemic biases that may exclude certain groups from employment or reinforce problematic employee attributes or behaviours in the workplace. These risks may be addressed collectively within the employment context or societally through legislative or regulatory interventions.Footnote 7 Nonetheless, the opacity and complexity of AI systems present major challenges for the liberal democratic ideal of effective public participation in governance. How, for example, are employees and job applicants, or those acting on their behalf, to know enough about specific AI systems in the workplace to contest their risks or to secure accountability for their harms through legal or regulatory processes? As we discuss in this article, public participation in AI governance requires innovative forms of public access and transparency throughout AI-system lifecycles as well as forms of AI explanation that aim to empower public participation rather than placate public anxieties.
Consequently, our focus is on the public transparency of AI systems as an essential element in the governance of their risks and on trade-secret claims as barriers to that transparency.Footnote 8 The notion of public empowerment through public knowledge has been expressed in policy catchphrases, such as “the right to know”, “the marketplace of ideas” and “the wisdom of the crowd”.Footnote 9 While these phrases capture different perspectives, they all reflect an ideal of informed and publicly determined action within liberal democratic societies and their market economies. The public right to know embraces principles of informed citizen participation in representative politics, including elections as well as direct involvement in political affairs, such as digital petitions or street demonstrations.Footnote 10 In contemporary networked and digitised societies, participatory public action has developed into hybrids of online and real-world life, in which intermediaries have acquired a critical role in accessing and organising information for public-interest purposes.Footnote 11 In addition to the agencies of the state, the focus for public contestation is now often on commercially funded Internet platforms and services, which are able to use big data and algorithmic processing resources to structure the public sphere.Footnote 12
In parallel, economic arguments for improving public transparency of AI stem from market-based concerns about innovation, growth and fair competition.Footnote 13 These concerns have become a key feature in policy debates about the role of major technology companies in developing and deploying methods that may block new entrants and suppress alternative innovation in the digital economy.Footnote 14 The public disclosure of risks of AI-related harms is, consequently, an important tool for both effective democratic participation and competitive innovation. In workplace settings, embedding opaque AI systems in operations and management may, for example, limit employee capacities to contest emerging rules of work while also reinforcing market dominance.
Nonetheless, while public transparency is an essential element within liberal democratic conceptions of governance, it carries risks of significant and potentially irreparable harms. To limit those risks, classes of personal, commercial and state information are routinely protected against public-transparency rights and duties, including trade secrets and commercially confidential information. In addition to these historically protected classes, there is now an acute public interest in protecting information stored, communicated or produced in networked systems from malicious attacks by tightening technical and organisational controls on access.Footnote 15 Consequently, the inherent complexity and opacity of AI systems together with increased security measures have combined with historic concerns about harmful public information disclosures to create overlapping technical and legal obstacles to AI transparency.
Facing these barriers, applicants for work or for promotions in work who suspect AI bias or error may find it near impossible to access and understand the factors that affected an AI-determined negative decision. As Del Castillo and Taes point out:
In an occupational setting, having access to the code behind an algorithm is not useful per se. What matters to workers is understanding the overall architecture of the AI model, the intended purpose; the context of use, how they are embedded in a system or layers of systems in the workplace; how they are exposed to or which personal data is collected from them.Footnote 16
While AI systems have improved in their capacities to produce explanations of their outputs, significant gaps between such explanations and the comprehension level of non-specialists continue.Footnote 17
Arguments for and against increased public scrutiny of AI systems are, consequently, being fought out in an increasingly dense field of design and operational options for providing different forms of AI transparency for different stakeholders. Importantly, debates about transparency processes for relevant stakeholders are occurring in a contemporary crisis of trust in public and private institutions. The claim for a public right to know is therefore not merely about appropriate information regarding specific AI systems for affected individuals to exercise presented options, it is also a challenge to established legal boundaries between transparency and confidentiality. In such circumstances, trade-secret claims have drawn considerable criticism as unjustifiable and pervasive barriers to the broader demands for public disclosure of AI risks.
Foss-Solbrekk succinctly summarised this groundswell of criticism, arguing that “[t]rade secret law […] obfuscates access to the algorithm and to explanations behind automated decisions, dimming the transparency of the entire system”.Footnote 18 Yet, despite criticism, the legal protection of trade secretsFootnote 19 remains formidable. Building on the historic physical and technical foundations of information control,Footnote 20 trade-secret law creates rights and duties to prevent unauthorised acquisition, use or disclosure of the protected information.Footnote 21 Where public interests in disclosure and non-disclosure conflict, it is often the case that trade-secret protection prevails. While transparency rights or duties can be used to require public disclosure of trade secrets, this will normally require demonstration of a compelling public interest in disclosure specific to the circumstances. Even here, however, advantages remain with the trade-secret holder, who may control access to the essential information needed to demonstrate that compelling public interest.
In this article, we review the weaknesses of UK and EU public-transparency laws in compelling the disclosure of trade secrets that affect matters of public importance. In that context, we investigate current arguments for reducing the scope of trade-secret protections for AI systems by carving out areas from legal protection that are deemed to be of major public interest, such as information regarding products and services affecting health and safety or decision-making by public authorities. Despite the academic upsurge of support for public-transparency “carve-outs” of this kind, we argue that their evident strengths are considerably outweighed by their weaknesses, which include uncertainty, over-inclusion and ineffectiveness. In general, these proposals run high risks of causing commercial harm without significantly developing legal or technical capacities for public scrutiny and contestation of AI risks.
In our view, there are more effective and appropriate solutions to be found through innovations in public-transparency law and practices that meet the technical and organisational challenges of AI transparency. In general terms, we argue that the vitally important public right to know should be reconsidered and adapted to the AI era. We therefore suggest three areas for further research and development: incorporating timely third-party review of trade-secret claims into developing public-transparency challenge and contestation processes for AI systems, which would remove some of the key unilateral advantages of trade-secret holders; widening the scope of vetted third-party participation in standard setting and compliance assessment within the trade-secrets barrier; and, developing public authority or agency responsibilities to report and disclose matters of public interest in the course of AI regulatory supervision and enforcement, while creating opportunities to contest the reasons for those disclosure choices. Throughout these areas, timely and authoritative assessment of the scope of trade-secret claims and conflicting public-interest claims will be key.
To make our argument, this article is structured as follows: Section II introduces “public transparency” as a specific variant of the concept of transparency and explains its relationship with the contrasting notion of “regulatory transparency”. Section III explains the definitive features of established public-transparency regimes, which continue to dominate conceptual thinking about public-transparency rights and duties. In Section IV, we discuss the disadvantages of proposals to introduce public disclosure presumptions or carve-outs in areas of high public interest. Section V explains the contemporary evolution of public transparency towards narrower targeting of specific disclosure needs and the strengths and weaknesses of this model. In Section VI, we argue that the threat that trade-secret claims pose for the public transparency of AI systems needs to be resolved through innovations in AI transparency concepts and practices and propose areas for development.
II. Regulatory Transparency and Compelled Disclosure
Public transparency – a term we use in this article to refer to the unrestricted use and sharing of information for lawful purposes – is the opposite of trade secrecy.Footnote 22 Public information use and sharing thus occurs outside the legal barrier that guards trade-secret claims against unauthorised disclosures. Within that barrier, there is, however, a highly flexible sphere of transparency for privileged participants. These include employees, commercial partners and other stakeholders as well as auditors, regulators and other law enforcement authorities. The potential for trade-secret barriers to accommodate third parties undoubtedly provides openings for public-interest scrutiny of trade-secret information and even its public disclosure in redacted form through intermediaries. Yet, disclosures to third parties within the trade-secret barrier are not without risk. Aside from unauthorised sharing or leaks, once trade-secret information is widely shared it may lose the essential element of secrecy that is required for legal protection.Footnote 23
On one level, public transparency stands in obvious contrast to disclosure to third parties within the trade-secrets barrier, including what we refer to as “regulatory transparency”. This form of transparency occurs either through voluntary information disclosure to government authorities or through the exercise of the state’s legal powers to compel information disclosures. These include a wide array of powers to search property, including electronic networks and databases, to seize material, including data, and to compel individuals to attend interviews. The powers that enable regulatory transparency descend from the historic exercise of coercive state powers to enter premises, confiscate possessions and question individuals, which liberal democracies now restrict substantively and procedurally. In the UK, for example, police forces and designated law enforcement bodies are authorised to conduct searches and interviews under the Police and Criminal Evidence Act 1984.Footnote 24 Other regulators, such as the Financial Conduct Authority and the Prudential Regulation Authority, have bespoke powers to obtain search warrants and compel interviews.Footnote 25 While the exercise of these powers is open to legal challenge, their effect is to render trade-secret rights largely irrelevant to a lawful investigation or enforcement.
Aside from confidentiality commitments to disclosing parties, regulatory agencies and other public authorities are routinely subject to strict confidentiality duties that limit further disclosure. In addition to general confidentiality principles governing their operations and sometimes specific statutory obligations,Footnote 26 public authorities in the UK may enforce confidentiality through employment contracts and codes of conduct.Footnote 27 In the EU, comparable confidentiality obligations for regulators are typically included in legislation that creates information disclosure duties or powers. For example, the Artificial Intelligence Act (“AI Act”) requires, among its other transparency obligations, providers of general-purpose AI models to maintain up-to-date technical documentation of their models, which must be provided on request to the new EU AI Office and national competent authorities.Footnote 28 In exercising their powers, AI Act regulatory bodies must, however, take due account of the rights and interests of the information providers, including the protection of confidential information, such as trade secrets.Footnote 29
Despite powers to compel the confidential disclosure of trade-secret information, regulatory transparency is a close cousin to public transparency. Public authorities and agencies are, first, primary information sources for Freedom of Information (“FOI”) requests. Consequently, public bodies have a key role in determining whether the balance of opposing public interests favours the disclosure of trade secrets they may hold. Additionally, when exercising scrutiny and investigation functions, they frequently act as transparency intermediaries, providing degrees of essential public disclosure enabled by privileged access to trade secrets.Footnote 30 As we discuss further, the notion of using the state as an instrument for AI public transparency extends to the use of government bargaining powers to require the public disclosure of trade secrets when contracting with AI application providers across many state functions.Footnote 31
III. Public Transparency’s Trade-Secrets Deference Challenge
As compared with regulatory powers to compel information disclosure, public-transparency regimes provide much more limited potential for compelling the disclosure of information subject to trade-secret protection. Within these regimes, such as FOI law, disclosure rights and duties are structured to protect overlapping, and even aggregated, classes of personal, commercial and governmental information from being accessed or shared.Footnote 32 Procedurally, this deference operates in the complex interplay between the public-interest defences for unauthorised disclosure built into trade-secrets law and public disclosure rights and duties located outside trade-secrets law.
While not property rights as such, trade secrets are an important element within intellectual property rights regimes and are secured not only by domestic laws, but also by EU law and international trade obligations, importantly including the Agreement on Trade-Related Aspects of Intellectual Property Rights (“TRIPS”).Footnote 33 In many jurisdictions, such as the UK and some other common law countries or territories, this protection occurs through legal duties of confidence, which if breached may result in an action for breach of confidence and an award of damages.Footnote 34 In the UK, the long-standing protection of trade secrets through duties of confidentiality was augmented by domestic implementation of the EU Trade Secrets Directive, which prohibits unauthorised acquisition, use or disclosure of trade secrets.Footnote 35
In addition to legal protection against unauthorised disclosure, trade-secret claims enjoy further advantages over demands for more extensive AI public transparency. In contrast to patented inventions and copyright works, information protected as a trade secret is not made public precisely because of its secrecy and thus, unlike patents, is not registered. Accordingly, unsound or inflated claims of trade-secret protection, which are often outwardly indistinguishable from legitimate ones, can be asserted against public access demands without prior third-party evaluation.Footnote 36 Without resort to third-party review and possible invalidation, unjustifiable claims cannot be overcome.Footnote 37 As Graves and Katyal have argued, trade-secret “claimants can choose their own narrative of trade secrecy to serve whatever their immediate goal may be”.Footnote 38 Even when unsound or inflated assertions of trade secrecy are made in good faith, such as those regarding new technologies, the self-judged nature of an initial trade-secret claim advantages the holder of the information.Footnote 39
While AI developers may, therefore, resist access to their training data on the basis that they are a trade secret, this claim is weak where training data have been drawn from Internet scraping, large-scale publicly accessible datasets or from data generated from smart machines.Footnote 40 In the employment context, for example, one can imagine AI interview tools that have been trained on publicly accessible information about an organisation, such as its values, the skills and expertise offered by its workforce, and the services it provides, along with publicly accessible datasets of mock interviews in order to facilitate sentiment analysis.Footnote 41 Or we can envisage AI tools aimed at increasing employee productivity in relation to home repair services being trained on or applied to dynamic data about traffic conditions, distances travelled and complexity of service provided. While such training data may be valuable, it is doubtful that they are secret or that their value is attributable to any secrecy. Yet, where employee concerns about bias or error in those AI assessments arise, the claim of trade secrecy sets up a barrier to public disclosure.
The public-interest exceptions contained in trade-secrets laws, such as the Trade Secrets Directive, are not well adapted to forcing public disclosure as they are primarily defensive in nature. Where disclosure is not voluntary, there is no mechanism within trade-secrets law to compel access to information. Instead, the law legitimises or justifies instances where acquisition, use or disclosure of trade secrets to third parties, including the general public, occurs without the consent of the trade-secret holder.Footnote 42 These provisions protect unauthorised disclosures that serve several broad public-interest purposes, including the exercise of the right to freedom of expression or revealing misconduct, wrongdoing or illegal activity for public-interest reasons.Footnote 43 Yet, the structure of these defences is to protect unauthorised disclosure by individuals or entities who already have access to trade-secret-protected information.
For public transparency of AI, the problem is not legal protection for public disclosure but how to compel disclosure of the information. In the normal course, neither individuals nor collective organisations are likely to have internal access to information asserted to be trade secrets, such as data, hyperparameters or weights used in the training process of an AI model or the AI model itself. There is, consequently, a close conceptual and practical alignment between public-interest defences in trade-secrets law and public-interest protections for whistle-blowers from employment dismissal and other retribution for unauthorised disclosure in breach of employment obligations. While offering genuine protections, these laws also aim to control the extent of whistleblower disclosures of confidential information.Footnote 44
On the other hand, trade-secret law exceptions for unauthorised disclosure and public-transparency regimes that create access rights are not as clearly aligned. Conventional public-transparency regimes, such as the FOI right of access, share three common structural features: public disclosure rights; exclusions from those rights for protected classes of information, including trade secrets; and, public-interest-based exceptions to some of those class exclusions. Primary rights to public disclosure tend to be robustly structured, permitting unlimited sharing and not requiring proof of an appropriate motive. That robustness is, however, invariably offset by broad exclusions for protected classes of information, effectively creating default non-disclosure barriers for these classes. The third common feature, public-interest-based exceptions to those broad exclusions, tends to be limited and narrowly construed. While the non-disclosure of trade secrets is thus open to narrow public-interest exceptions, other classes, such as classified government secret information, tend to be beyond right-to-know-based compelled disclosure.Footnote 45
We turn now to consider the operation of these common features in two foundational public-transparency regimes: access to judicial proceedings and FOI. The experience of these regimes continues to shape our understanding of how conflicts between public interests in transparency and secrecy should be managed and resolved.
A. Open Justice
Often referred to as “open justice” in UK law, public access to court proceedings and the unrestricted sharing of information made public in those proceedings is a major precedent for the design of public-transparency regimes.Footnote 46 As a form of public scrutiny of a state institution, open justice reflects public transparency’s traditional focus on governmental authority and power. Its importance as a specialised public-transparency regime is also indirectly recognised in the UK Freedom of Information Act (“FOIA”), which absolutely excludes information held in relation to court proceedings.Footnote 47 This, in effect, respects the right of the courts to manage their own form of public transparency.
In the UK, the courts have moved over time to entrench and broaden the public-interest rationales of open justice. The Supreme Court has, for example, affirmed the connection between open justice and the role of media reporting of matters of general interest,Footnote 48 linking public transparency of the state and the notion that social forces have a role in fostering accountability. Open justice has also evolved in response to changes in the ways that courts and litigants communicate information outside of open court proceedings, allowing greater public access to background evidence and submissions that are essential to understanding those proceedings
Although access to EU court proceedings has a different legal basis, there are important commonalities with UK law, including the reinforcing role of Article 6 of the European Convention on Human Rights (“ECHR”), concerning the right to a fair and public hearing. Within EU law, that fundamental right is also protected via Article 47 of the EU Charter of Fundamental Rights, concerning the principle of public access to the courts.Footnote 49 EU courts are also largely exempted from EU FOI legislation, which defers to their internal rules governing public access.Footnote 50 While this exemption under the EU FOI Regulation is in principle limited by a public-interest-based exception, the Court of Justice of the European Union (“CJEU”) has used the EU treaties to interpret that exemption in broad terms.Footnote 51 Efforts to create compelled public disclosure duties for documents submitted to EU courts by parties to litigation or interested third parties have, moreover, been judicially resisted.Footnote 52
Inevitably, the principle of public knowledge of judicial processes does not include unconditional access to information protected as trade secret.Footnote 53 When considering trade-secret claims that are relevant to judicial proceedings, courts will invariably require confidential disclosure of that information to other parties of the litigation where necessary in the interests of justice.Footnote 54 Trade secrets and other classes of protected information are, however, routinely exempted from disclosure in open court unless a compelling public-interest justification is demonstrated. Indeed, where deemed necessary, disclosures between parties may be strictly limited to a litigant’s legal representatives, who are subject to disciplinary remedies for breaches of confidentiality.Footnote 55 These restrictions on the right of the public to report on proceedings held in open court are subject to standards of legitimacy, necessity and proportionality, yet such restrictions are also routinely upheld.Footnote 56
The treatment of trade secrets in relation to open justice illustrates several key issues arising in public transparency. First, the courts themselves are empowered to reject unsound or inflated claims to trade-secret protection relied on by litigants. This ready access to impartial third-party adjudication in open justice contexts not only demonstrates its importance in addressing the obfuscating effects of self-judged trade-secrets claims, but also the weaknesses of other public-transparency regimes where impartial third-party assessment is external to the process.
Second, the courts are also able to determine whether there is a compelling public interest in the disclosure of legitimately protected trade secrets in open court. On one level, this simply asks whether public disclosure is necessary to serve the purposes of open justice, which primarily concern public understanding and scrutiny of the justice system.Footnote 57 Trade-secret claims in litigation are, however, also open to the question of whether the public-interest exceptions internal to trade-secrets law are satisfied. A court may, therefore, determine that the public interest in protecting the trade-secret claim at issue is outweighed by a sufficiently cogent public interest in disclosure.Footnote 58
B. Freedom of Information
While FOI laws were adopted long after the recognition of open justice as a legal principle, FOI is in many respects the definitive public-transparency right.Footnote 59 As the UK Government declared in its 1997 White Paper introducing the proposed FOIA, “[o]penness is fundamental to the political health of a modern state” and “[u]nnecessary secrecy in government leads to arrogance in governance and defective decision-making”.Footnote 60 The CJEU has similarly stated that the right of access “enables citizens to participate more closely in the decision-making process and guarantees that the administration enjoys greater legitimacy and is more effective and more accountable to the citizen in a democratic system”.Footnote 61
Following the three common features described above, the UK’s FOIA creates a general right to the public disclosure of information held by designated public authorities, introducing a powerful legislated skew towards the scrutiny of public power. The FOIA right is thus only indirectly applicable to private entities where the supply of goods or services or other interactions with the state involves sharing trade secrets or private information with public bodies.Footnote 62 Other FOI regimes rest on a similar public disclosure right, although they may be constructed on a narrower basis. For example, the UK Environmental Information Regulations 2004 (“EIR”), intended to implement EU law, only create a right to information falling within the defined sphere of “environmental information”.Footnote 63 The EU FOI regime, established by Regulation (EC) No 1049/2001 and subsequently underpinned by Article 42 of the Charter of Fundamental Rights (“CFR”), creates a right of access only to “documents” held by EU institutions and not information in any wider meaning of that concept.Footnote 64
Despite such differences, FOI regimes realise important elements of the public right to know, including the link between free public debate and democratic participation and the motive-blind disclosure of state information through FOI requests. Provided there is effective sharing through the media or other intermediaries, the release of information obtained through FOI requests potentially contributes to the democratic and economic goals of public transparency. Those expected benefits, however, occur alongside the utility of FOI rights for pursuit of private interests,Footnote 65 which raises problematic fit for purpose questions about the beneficiaries and justifications for FOI regimes.Footnote 66
Building on the general-purpose access right, liberal democratic FOI regimes typically create structured balances between the conflicting public interests in disclosing state-controlled information and protecting classes of sensitive information, including trade secrets, absolutely or conditionally from disclosure.Footnote 67 Nonetheless, non-disclosure of information protected by a trade-secret claim can be challenged by demonstrating that the claim does not satisfy trade-secret criteria or there are public-interest reasons to compel disclosure.
In the UK, section 41 of FOIA creates an exemption for information provided in confidence, whereas section 43 creates exemptions for trade secrets and information prejudicial to commercial interests. In comparison, the EU FOI Regulation has a simpler structure, establishing two general class exclusions from the right of access to EU institutional documents.Footnote 68 The narrower, first category concerns documents relating to the core interests of the state. The second category, found in Article 4(2), includes refusal of access where disclosure would undermine the protection of “commercial interests of a natural or legal person, including intellectual property”. As discussed below, while the exemptions in the EU’s and UK’s FOI regimes are structured differently, they both typically favour the commercial interests of trade-secret holders over the public interest in disclosure.
Outwardly, FOI regimes create an efficient method for subjecting self-judged trade-secrets claims to third-party assessment and reasoned decision by a public authority when public disclosure is sought. The full impartiality of that third-party review, however, cannot be assured where the relevant public authority is directly or indirectly an interested party. This may occur, for example, where the trade-secret holder is an established or favoured service provider to a public authority. Yet, even without such a relationship, the public authority may have institutional reasons to favour non-disclosure. The UK Medicines & Healthcare products Regulatory Agency (“MHRA”), for example, has cited its own commercial interests as a factor in its decisions regarding section 43 of FOIA.Footnote 69 Institutional partiality may also have been present in the important EU FOI litigation in Breyer v European Research Executive Agency (REA).Footnote 70 The Agency adopted an expansive view of what documentation fell within the scope of the exemption than was subsequently established through the applicant’s pursuit of administrative and then judicial review.
C. Overriding Public Interest – Specific and General Interests
The rules regarding initial presumptions favouring disclosure or non-disclosure of trade secrets are markedly different in the EU and UK FOI regimes. The EU FOI Regulation provides that the general public interest in transparency in governmental affairs is not, by itself, sufficient to show an overriding public interest in the compelled disclosure of trade secrets.Footnote 71 As the CJEU reaffirmed in 2023, “general considerations alone are not capable of establishing that the principle of transparency is of particularly pressing concern”.Footnote 72
Critically, section 43 of UK FOIA is a qualified exemption subject to the Act’s public-interest balancing test, in which the general public interest in disclosure of information held by public authorities in principle outweighs the public interest in non-disclosure expressed by the exemption. Nonetheless, outcomes in practice regarding trade-secret disclosure are closer to the EU standard than this formal difference implies. To change the balance in favour of non-disclosure under the FOIA public-interest test, it is necessary to show that “disclosure of the information would or would be likely to have a prejudicial effect”.Footnote 73 That is, however, a low threshold in most trade-secrets cases, even taking into account that the prejudice must be “real and significant, more than hypothetical or remote”.
To be protected under section 43, a trade secret is by definition “information, which, if disclosed to a competitor, would be liable to cause real (or significant) harm to the secret’s owner”.Footnote 74 That definition gives a trade-secret claim a substantial head start in showing the requisite prejudice. Assuming, for example, that a commercial provider of AI services has shared secret, commercially valuable information with a public authority subject to FOI access rights, it would not be difficult to show that the provider would suffer a loss of competitive advantage if such information became publicly available.
Consequently, once it is established that the requested information is a protected trade secret, the presumption in favour of public disclosure is easily, if not automatically, overcome.Footnote 75 In that situation, like the EU FOI rule, the requesting party must then demonstrate a specific public interest that is sufficiently compelling to outweigh the public interest in preventing the likely harms of disclosure. While also based on an initial presumption favouring disclosure, the balancing test for the parallel UK Environmental Information Regulations exception covering trade secrets has a similarly low threshold for showing sufficient harm to outweigh that presumption.Footnote 76
Where trade-secret-protected information disclosed to a public authority is covered by an overlapping duty of confidentiality, a FOIA request will also need to meet the higher bar of section 41. Here, however, the question of whether there is a public interest in disclosure that is sufficiently compelling to displace the public interest in protecting confidential information only arises indirectly. As an absolute exemption, section 41 excludes direct consideration of that question. Nonetheless, the section only protects actionable breaches of confidence, which requires consideration of whether there is a relevant public interest in disclosure that vitiates actionability. As a final twist, that assessment presumes the public interest in maintaining confidentiality will prevail unless outweighed by a specific public interest in disclosure that goes beyond the general public interest in transparency of governmental information and outweighs any harms that may result.Footnote 77
As critics point out, there are potentially insurmountable practical difficulties in demonstrating a compelling, specific public interest in disclosure of trade-secret-protected information.Footnote 78 Without prior knowledge of the protected information, the validity challenge will often fail for lack of sufficient information to demonstrate a specific public interest that outweighs the potential commercial harms of public disclosure. An allegation of bias or error in AI-determined recruitment or promotion outcomes would, for example, require considerable evidence linking the AI tools involved to demonstrably clear and otherwise unexplained biased or wrong outcomes before a public-interest disclosure argument might prevail.
Despite the disproportionate advantage bestowed on trade-secret holders when the requester has the onus of evidencing an overriding specific public interest, this feature has survived the European constitutionalisation of the FOI disclosure right. Although the EU FOI Regulation was adopted before the EU Charter of Fundamental Rights, the CJEU has not considered that this requirement in the regulation is contrary to Article 42 of the Charter, which recognises the public right to access EU documents. At best, the Charter has heightened the importance of any factors associated with the exercise of fundamental rights to freedom of expression and information when balancing public interests in disclosure and non-disclosure.Footnote 79
Recognition of a public right of access to information under Article 10 of the ECHR, relating to freedom of expression, has also failed to transform the established limits of public disclosure. In its ground breaking decision in Magyar Helsinki Bizottság v Hungary in 2016, the Grand Chamber of the European Court of Human Rights (“ECtHR”) confirmed that Article 10 neither obliges a state party to collect and disseminate information nor confers on the individual a right of access to information held by a public authority.Footnote 80 Notwithstanding that general principle, the court also found that an information access right does arise under Article 10 when access to state-held information is instrumental to the exercise of the right to freedom of expression, setting out several indicative criteria for when individuals or entities acting in public or social watchdog roles may enjoy this right.Footnote 81 These criteria include the purpose of the information request; the nature of the information sought; the role of the applicant; and whether the information is readily available. The resulting fact-specific balancing test hinges on whether there are sufficient preponderant factors favouring public disclosure.Footnote 82 The UK Supreme Court has yet to be asked to re-consider the position in UK law in light of Magyar Helsinki Bizottság v Hungary, which the Court of Appeal has recently declined to examine.Footnote 83
The established balancing principles regarding public disclosure and trade secrets recognised in Magyar Helsinki Bizottság are also undisturbed by the Council of Europe’s new Framework Convention on Artificial Intelligence and Human Rights, Democracy and the Rule of Law. According to the treaty’s Explanatory Report, its Article 8 “adequate transparency” obligations remain subject to balancing principles where competing interests are affected.Footnote 84 Developments in international human rights law regarding rights of access to information have paralleled developments in the Council of Europe also without breaking significant new ground. The International Covenant on Civil and Political Rights (ICCPR) Human Rights Committee has declared that ICCPR, art. 19, para. 2 “embraces a right of access to information held by public bodies” and that “[a]uthorities should provide reasons for any refusal to provide access to information”.Footnote 85 In other words, the Article 91(3) legitimate restrictions on national security or public order or in the interests of “the rights or reputations of others” apply.Footnote 86
IV. Proposals for Cutting Back Trade-Secret Protections
In the face of transparency mechanisms structured to favour the non-disclosure of trade secrets, proponents of a fundamental reform of trade-secret protection argue that public-transparency law undervalues the general importance of public transparency.Footnote 87 Proposed alternatives would dispense with situation-specific public-interest balancing in selected areas of public concern and instead envision broad legislated presumptions of public disclosure in those areas.
These proposed transparency carve-outs identify diverse areas where trade-secrets protection could be narrowed, including governmental functions or public services,Footnote 88 products or services offered to consumers,Footnote 89 health and safety,Footnote 90 environmentFootnote 91 or employment conditions.Footnote 92 These proposals, which suggest using different legal frameworks, such as FOI, data-protection law or AI regulation, identify the use of AI systems or applications as a key justification for narrowing protection for trade secrets.Footnote 93 These proposed carve-out fields include use of high risk AI systems,Footnote 94 Internet of Things applicationsFootnote 95 or public space facial recognition applications.Footnote 96
Narrowing the scope of trade-secret protection through new public-interest-based carve-outs would undoubtedly increase the apparent scope of FOI and other public disclosure rights. In their spheres of operation, transparency carve-outs would sweep away not only the question of whether a trade-secret claim was unsound or inflated, but also whether a specific, overriding public interest could be demonstrated. In the AI sector, where the complexity and opacity of AI systems has made reverse engineering and other workarounds to trade secrets much more difficult to achieve, carve-outs could reduce unearned trade-secret-holder advantages.Footnote 97 Where AI is used to assess job and promotion applications or monitor employees, transparency carve-outs could extend the reach of workplace transparency rights, excluding trade-secret rationales for refusing public disclosure of information relevant to risks of AI discrimination, manipulation or error.
Nonetheless, in our view, transparency carve-out solutions have serious disadvantages that outweigh the apparent benefits of mandatory public disclosure of trade-secret-protected information in selected areas of public concern. We see serious, unexamined objections to these proposals grounded in problems of uncertainty, over-inclusion and effectiveness. Consequently, transparency carve-outs carry significant risks of creating destructive harms for trade-secret holders without appreciably advancing the objectives of public transparency.
A. Uncertain Boundaries and Unbounded Risks of Harm
Commercial activities are unlikely to fit neatly within any sphere in which public transparency is deemed to be essential to societal interests and made subject to a transparency carve-out, such as health and safety or public space facial recognition. Mandatory public disclosure would therefore cut through commercial sectors, risking arbitrary effects amongst competing trade-secret holders. Employers in one business sector deemed to be in scope, for example, could be compelled to disclose trade-secret-protected information to their employees and through them to the public, while their competitors in an adjacent sector might not be in scope. Yet, despite commercially damaging costs, carve-outs would not necessarily ameliorate the imbalance of power present in highly mediated transparency processes. Indeed, by heightening the risk of harmful compelled disclosure, carve-outs could lead to greater obstruction and non-cooperation by trade-secret holders.
This potential can be seen in the FOI field in which the disclosure of “unwanted, unlabelled, unsorted, raw data” remains an unresolved disclosure practice.Footnote 98 In the context of AI transparency, which is imperfect and frequently aspirational, disclosures and explanations occur at different stages through design, training and operation. For parties wishing to avoid disclosure, there are ample opportunities for obfuscation. Public disclosure could also be legitimately refused on the grounds that the information does not fall within the scope of the carve-out category. In FOI contexts, that refusal would inevitably require adjudication, including the delays and costs associated with that.
The disclosure of trade secrets in a transparency carve-out sphere would also create unintended risks of wider economic harm. In some circumstances, the consequences of commercial investments or technical innovations foregone could be irreparable, especially for smaller or new competitors. The possibility of compelled public disclosure of commercially sensitive information through employee transparency rights could, for example, simply foreclose discussion of the potential benefits and risks of some AI applications. The downsides of introducing transparency carve-outs could, no doubt, be seen as a necessary trade-off between opposed forms of public interest, in which commercial interests must be restricted to achieve the benefits of an informed public. This view of public transparency is complacently managerial, failing to consider the incommensurability of such trade-offs and overlooking the multilayered and unstable conflicts present in liberal democratic transparency rights.
That instability arises out of the potentially unbounded nature of a public right to know that seeks to set aside established limits on public transparency without clarity about ultimate ends.Footnote 99 As a claim to open and direct public scrutiny of the workings of the state, the right to know is unmistakeably rooted in ideas of popular sovereignty and, in radical forms, the ultimate subordination of the institutions and laws of the state to notions of the will of the people.Footnote 100 In this familiar tension, the foundational legitimacy of the state rests on the people’s constituent power, which exposes the institutional and legal order to risks of populist impairment or even destruction.Footnote 101 Once the right to know is understood as a manifestation of popular sovereignty, the potential for transparency rights to serve destructive political purposes in volatile times becomes clearer.Footnote 102
In less antagonistic versions, popular sovereignty sustains demands for greater public participation in the processes of democratic governance and oversight,Footnote 103 linking public transparency with the scrutiny and contestation of power.Footnote 104 Inevitably, that conception of public scrutiny of power extends beyond the state to private exercises of power, which are enabled in part by intellectual property rights. Even from this moderate perspective, the proposed definitional boundaries for transparency carve-outs are weak and open to expansion, especially when their foundations in democratic principle are minimally developed or non-existent. One sectoral carve-out from trade-secret protection is likely to lead to others, while also placing pressure on the protection of sensitive governmental or personal information from public disclosure.
The defence of trade secrecy from unbounded public scrutiny rests on institutional responses to popular sovereignty, which assert legitimacy through representative democracy and constitutional order, importantly including the safeguarding of human security, dignity and autonomy.Footnote 105 Significantly, these responses undercut the normative significance of collective political agency by claiming the same broad moral principles that underpin the people’s right to know.Footnote 106 The resulting institutional legitimacy is essential to the complex rules and balancing mechanisms that currently govern public-transparency regimes, keeping the more destructive tendencies of the right to know in check.Footnote 107 Those defences are present not only in the broad exclusions for classes of trade secrets that offset public disclosure rights and duties, but also in the careful limits placed on public-interest-based exceptions to those protections.
Nonetheless, popular sovereignty’s demands for unbounded public transparency may be checked but cannot be disarmed. The relationship between public disclosure rights and the protection of trade secrets of any kind is ultimately antagonistic and irresolvable. Consequently, while transparency carve-out proposals see the right to know and associated arguments for mandatory public transparency as benign and underserved, those proposals tend not to consider fully the need for constraints as well as reinvigoration.
B. An Ineffective Approach to Public Transparency
In several respects, there are anachronistic elements in proposals to address the obstructive effects of trade-secret protection on AI transparency through carve-outs. Since its relatively recent appearance in the canon of liberal democratic rights, public transparency continues to evolve, both conceptually and practically.Footnote 108 That evolution is, however, more than a response to the ongoing digital revolution in information technologies and applications, whose rise coincided with the introduction of FOI and data-protection rights. It also reflects changing ideas about democratic engagement and public participation and, equally, about the ways in which information markets generate innovation and economic growth. Moving from earlier generic rights, such as the FOIA right to information held by public authorities, to increasingly specialised and purpose defined rights, such as the EU Data ActFootnote 109 data-portability right, public transparency is now characterised by sets of overlapping disclosure rights and duties.
The increasing complexity and opacity of AI systems is, moreover, driving transparency concepts and methods rapidly through further technical and organisational specialisation.Footnote 110 This includes a range of novel information disclosure methods to facilitate AI’s explainability, interpretability and auditability. To be effective, these methods will often require technical expertise and other resources for parties on both sides of AI disclosure processes as participation and dialogue between developers, deployers and other participants becomes essential to effective transparency outputs.Footnote 111 Where responsibility is widely distributed through multiple parties, applications and outputs, effective AI transparency may need to encompass not only the full lifecycle of an AI system, but also the entire breadth of its socio-technical dimensions, which may also be transnational in scope.Footnote 112 Plainly, specialisation in AI transparency methods raises important questions about the future role of the public in these methods. For employees concerned about AI risks in the workplace, for example, the opportunities for contesting AI transparency outputs may lead to innovations in informed participation in AI transparency processes or, less optimistically, performative transparency, such as chatbots dispensing formulaic responses.
Proposals for transparency carve-outs tend not to be well adapted to the challenges of AI transparency specialisation. While aiming to narrow FOI’s broad exclusions for trade secrets, they frequently look back to the design of FOI rights as a model for public disclosure. No doubt, FOI disclosure requirements have considerable potential for forcing the disclosure of information revealing AI risks and harms. Important elements of AI models or systems can, for example, be accessed without need for operational explainability, such as training data information as well as AI generated decisions, content and other outputs. Doctrinal developments in FOI law requiring that the information is disclosed in accessible and readable forms or formats can also be further developed in AI contexts.Footnote 113 There is, however, no evident interest at the governmental level in the UK or EU in introducing general rights to explanation into FOI laws.
Nonetheless, the comparatively simple FOI disclosure model is not structured to achieve the degree of participation and explanation needed to achieve comprehensively effective AI transparency.Footnote 114 In addition, its focus on public authorities only indirectly reaches the commercial providers of AI systems and applications through their information sharing with public authorities.Footnote 115 These shortcomings indicate that the resolution of conflicts between public AI transparency and trade-secret claims will require an integration of the dynamic, public-driven qualities of the right to know with emerging methods of AI transparency, which transparency carve-outs cannot offer.
V. Meaningful Transparency through Targeting
Innovations in transparency regimes since the establishment of the FOI disclosure model offer considerable potential for disentangling public-transparency needs and trade-secret protection. In simple terms, disclosure rights and duties are becoming more targeted in their purposes, processes and intended parties, steering away from the broad access model of FOI. The aim of this targeting, which follows precedents in mandatory disclosure requirements for consumers and other relevant groups, has been to provide meaningful information disclosures for defined purposes and parties. As we discuss, targeting transparency rights or duties in this way can include as well as exclude the obligatory disclosure of trade secrets.
Despite these innovations, post-FOI public-transparency frameworks typically retain the three common features identified above: disclosure rights and duties; exclusions for protected classes, including trade secrets; and, in some cases, public-interest-based exceptions to those exclusions. Where it works well, targeted transparency can offer a method to achieve effective transparency results while avoiding unnecessary conflicts with trade-secret claims. Targeting transparency purposes not only reduces the scope of the disclosure requirement to fit the identified stakeholder’s needs, but also considers whether a trade-secret claim is genuinely obstructive to those defined needs. Indeed, this evolution in public transparency towards targeting raises the question of whether creating transparency carve-outs is a necessary or relevant goal.
Data-protection law provides one of the best examples of targeting in public transparency, having two distinct public disclosure rights relevant to AI transparency. These include, first, the right of data subjects to access their personal data (“SARs”) under General Data Protection Regulation (“GDPR”), art. 15, which, in some respects, resembles FOI access rights. SARs apply to all data controllers, whether public or private, thus overcoming the public/private split that often restricts the potential scope of FOI requests regarding matters of public interest. Like FOI requests, SARs are also a form of public transparency as the requester’s motive is not a relevant consideration and any disclosed information can be shared without restriction.Footnote 116 These rights are, consequently, widely used for non-data-protection collateral purposes, such as seeking information prior to a formal complaint or legal action in employment contexts.Footnote 117
In other significant ways, SARs are more targeted than FOI rights, applying only to the personal data of the data subject who requests SAR disclosure. Additionally, their motive-blind character does not prevent their data-protection purposes from being taken into account when an SAR request is balanced against conflicting interests, such as a trade-secret claim. In the event of a conflict, the necessity and proportionality of the requester’s data-protection interests will be weighed against the risk of commercial harm to the trade-secret claimant.Footnote 118 Aside from giving less weight to SARs made for collateral purposes, this context-specific inquiry will consider the relative importance of the personal data to the requester’s ability to exercise specific data-protection rights, such as objection or erasure.Footnote 119
Second, of rising importance for AI transparency, GDPR, art. 22 and associated provisions are arguably the basis for a data subject’s “right to an explanation” in relation to automated processing of personal data.Footnote 120 This includes the right to receive meaningful information regarding the logic of that processing and an explanation of its potential consequences.Footnote 121 While this falls short of the complex demands of effective AI transparency, being neither a right to receive answers to questions or to participate in disclosure processes, it is a major advance over FOI rights or GDPR SAR rights where Article 22 is not engaged. As the CJEU has confirmed, the GDPR affords “a genuine right to an explanation as to the functioning of the mechanism involved in automated decision-making”.Footnote 122 Nonetheless, while this requires simple, yet meaningful, explanations of the logic involved, it does not ordinarily require complex explanations.
In the EU, further innovations in targeted transparency have built on the experience of data-protection rights. These new forms of transparency rights and duties are more explicit in limiting access to trade secrets to defined parties and the sharing of disclosed data. The EU Data Act, following the GDPR’s modest data-portability rights, has created a regime for the compelled disclosure of usage data for connected products and services to qualified users.Footnote 123 Any usage data protected by a trade-secrets claim need not be disclosed unless disclosure is strictly necessary for an EU Data Act purpose and the recipient party agrees to any measures necessary to preserve confidentiality.Footnote 124 The EU Data Act thus brings the qualified user within the trade-secret barrier, applying non-disclosure obligations to prevent unauthorised sharing.Footnote 125 While this is plainly not full public transparency, it does borrow elements of the right to know, such as the public initiated exercise of disclosure rights.
In providing access to data held by very large online platforms (“VLOP”) or very large online search engines (“VLOSE”) to qualified researchers (“vetted researchers”), Article 40 of the EU Digital Services Act (“DSA”) creates a different solution to manage trade-secret transparency conflicts.Footnote 126 Vetted researcher requests are strictly purpose limited and subject to approval by an authorised National Digital Coordinator (“NDC”) before submission to a VLOP or a VLOSE.Footnote 127 The subsequent obligation to provide researcher access to the requested data does not, moreover, apply if disclosure would lead to vulnerabilities in the protection of confidential information, including trade secrets. Thus, unlike the EU Data Act, the DSA keeps vetted researchers firmly outside the trade-secret barrier. At the same time, the DSA and its implementing rules do not create express powers for NDCs to invalidate trade-secrets claims where they are determined to be unsound or inflated or where a specific public interest is judged to outweigh the DSA presumption in favour of non-disclosure.Footnote 128
While the GDPR, EU Data Act and DSA potentially apply to a host of AI applications, AI-specific regulation in the EU has so far not introduced major innovations in targeted transparency. The AI Act, which primarily concerns AI safety and risk management, creates basic public-transparency obligations consistent with consumer protection standards.Footnote 129 These aim to ensure that the public are aware of when high risk AI systems are being used, with special attention to particular applications such as emotion recognition or generation of deep fakes.Footnote 130 The public are not, however, significant rights bearers under the AI Act’s product safety design, although the Act extends the GDPR right to an explanation.Footnote 131 In relation to AI discrimination in the workplace, for example, Article 77 of the AI Act empowers public authorities to compel disclosure of relevant information in an accessible language and format, but provides no equivalent for employees or their representatives. While the AI Act may therefore have disappointed those with aspirations for significant empowerment of public transparency, efforts continue to be made to develop forms of targeted transparency that provide meaningful transparency for particular needs.Footnote 132
Undoubtedly, targeted transparency obligations for AI-system developers, deployers and users could substantially reduce the potential for public-transparency and trade-secret conflicts. A well-designed, targeted AI public-transparency framework could identify individual or collective needs for information disclosure, while also identifying and excluding unnecessary or disproportionate disclosures of trade-secret claims. The targeted transparency model, however, usefully makes clear that an explicit solution to the trade-secret barrier is required where trade-secret disclosure is unavoidably at issue. Adapting the EU Data Act model, disclosure to qualified and trustworthy third parties could be made subject to confidentiality obligations. Alternatively, the DSA-vetted researcher model could be adapted to protect information controllers subject to targeted disclosure requirements against disclosing trade-secret-protected information whose disclosure would risk serious commercial harm. Developing the DSA-vetted academic researcher access model for other contexts, such as workplace AI transparency, would undoubtedly present challenges in determining who would qualify for vetted access.
Targeted transparency obligations are, however, neither a form of nor a substitute for the right to know. In effect, a targeted transparency framework is a mechanism for legally determining the necessary scope and purposes of transparency on a “need to know” basis, which is entirely different from the dynamic, exploratory nature of a public exercised right to know.Footnote 133 There is, consequently, an inherent risk that the public-interest disclosure goals of a targeted transparency regime are structured to avoid conflicts with the protection of trade secrets. While targeted transparency usefully navigates the potential harms of the right to know, it also widens the risks of unaccountable secrecy, relying too heavily on institutional claims to democratic legitimacy and public trust.
VI. Trade Secrets and Innovations in AI Public Transparency
How effectively the right to know can be structured or function in the AI era remains to be seen. While transparency carve-outs from the scope of trade-secret protections are, on balance, poorly designed solutions and targeted forms of transparency negate important aspects of the right to know, alternative solutions ensuring that trade-secret claims do not obstruct necessary AI public transparency will require considerable innovation. As noted, the complexity and opacity of AI systems will often strengthen the inherent advantages of information controllers. Consequently, in many circumstances, demonstrating that the trade-secret claim is unsound or inflated or that there is a compelling public interest in disclosing trade-secret-protected information will require considerable technical expertise.
In our view, while AI public transparency is essential to protect the many collective and private interests at risk to AI systems,Footnote 134 protection for valid trade-secret claims remains in the general public interest. Yet, that baseline assumption begs several questions. How should the dynamic, public-driven character of the right to know be integrated into developing AI transparency processes; how should unsound or inflated trade-secret claims be evaluated; and how should overriding public interests in disclosure be demonstrated? Resolving these questions will require considerable innovation on both sides of the trade-secret barrier. In this concluding section, we examine key challenges to better AI public transparency and explain how trade-secret obstruction of legitimate transparency could be addressed as an integral part of the solutions to those challenges. Our purpose here is to outline areas for further research.
The challenges in delivering effective AI public transparency are multidimensional.Footnote 135 As AI systems are invariably networked and carry high risks of malicious or negligent use, transparency processes will need to be security risk compliant. Public-transparency processes must also be designed to ensure that any disclosures of personal data are necessary and proportionate. Consequently, the protection of trade secrets is nestled within an overlapping set of public-interest reasons for limiting transparency to forms of low risk, targeted transparency. In that tendency, AI public transparency becomes the management of the public’s need to know rather than facilitating the public’s right to know, while also hazarding public trust in AI governance that operates without public scrutiny. In our view, AI public transparency requires external participation and independent observation in collaborative processes aimed at delivering explainability and interpretability of AI risks.Footnote 136 That requires methods of mitigating the risks of public disclosure, rather than their elimination.
We see three areas in which the right-to-know ideal could be developed in AI transparency processes. First, the incorporation of binding third-party review of trade-secret claims directly into public-transparency challenge and contestation rights for AI systems would remove some of the unilateral advantages of trade-secret holders. The vetted researcher access right, for example, found in DSA, art. 40, offers a potential, if not so far sufficient, model.Footnote 137 Interestingly, the CJEU has recently pushed in a similar direction. In its Dun & Bradstreet decision, the court interpreted the GDPR data subject access right to include the requirement that, where a controller’s refuses access on the grounds of trade-secret protection, the controller must disclose the protected information to the relevant regulator or court if challenged by the data subject; the regulator or court must then determine whether non-disclosure is necessary and proportionate.Footnote 138
Drawing on these formative ideas, it is worth investigating how to create faster and simpler intermediary review processes for trade-secret claims to compel disclosure of information where a claim is unsound or inflated or there is an overriding public interest. The latter element raises the issue, discussed above, of how specific that public interest in disclosure should be to overcome the general public interest in protecting legitimate trade secrets. Solutions may lie in requiring intermediary authorities to investigate proactively evidently reasonable public-interest arguments raised in access requests.
Second, widening the scope of vetted third-party participation in standard setting and compliance assessment within the trade-secrets barrier could also bring elements of the public right to know inside the trade-secret barrier. In academic and civil society spheres, there is a considerable push for more public participation in AI governance, including the design of AI transparency and accountability processes.Footnote 139 At their broadest, these proposals are captured by concepts of AI democratisation, which embrace collective decision-making by creating rights of participation in AI use, development, revenues and governance as well as building technical and organisational capacities to make those rights effective.Footnote 140 AI democratisation, however, risks the unbounded populist tendencies discussed above.Footnote 141 Somewhat more narrowly, proposals for public participation have been advanced through ideas of participatory design and similar concepts.Footnote 142 As Corbett and Denton have explained: “[p]articipatory design is fundamentally about dismantling power and information asymmetries between designers and users […] Participatory design can help achieve this by bringing people into the design process such that they can be empowered to discover the ends and means of transparency that are meaningful to them.”Footnote 143
The adoption of the EU AI Act raised expectations for public participation in AI transparency standard setting as well as subsequent compliance assessment and auditing. Yet, progress in public AI transparency methodologies has moved slowly in comparison to the surging pace of AI model development. AI documentation methods, such as model cards, data cards, risk labels and disclosure interfaces, which are essential to better AI public transparency, vary considerably and are often not fit for that purpose.Footnote 144 Disclosure practices, such as technical “data dumps”, may not meet expectations for meaningful public explainability and interpretability.Footnote 145 Commitments to disclosure are also restricted in practice by encryption and other barriers to access.Footnote 146
The standardisation of AI transparency processes and outputs is widely seen as an essential foundation for improving AI public transparency, which will require novel substantive and procedural solutions to the blocking effects of trade-secrets claims.Footnote 147 There is, however, an evident lack of consensus on the aims of that transparency standardisation or who should participate in the standardisation processes.Footnote 148 For proponents of civil society participation in transparency governance, the objective is to put “the right information in the hands of the people with the legitimacy to make complex normative decisions at a local, context sensitive level”.Footnote 149 The realisation of effective disclosure methods that are open to public contestation is certainly less likely where standardisation is dominated by technical experts selected through closed standard setting processes.Footnote 150
External participation in compliance assessment and auditing processes is a potential avenue for building effective AI public transparency but would require considerable ingenuity at the trade-secrets barrier. AI auditing for security, fairness, reliability, data protection and other requirements is already a focus for potential public participation in operational oversight,Footnote 151 although unstandardised and idiosyncratic.Footnote 152 This developmental moment in AI assessment and auditing offers an opportunity to increase the range of participant groups involved in the design and execution of audits,Footnote 153 such as civil society organisations.Footnote 154 So far, however, legal mandates for participation through stewardship models or the capacity and knowledge base needed to enable effective public participation are inadequate.Footnote 155
Widening public participation in audits would undoubtedly help to bring a publicly defined right-to-know inquiry into the processes of AI oversight. Yet, that objective carries obvious hazards. Expanding the range of actors operating under the confidentiality obligations of auditors would, for example, increase risks of unauthorised disclosure through non-compliant parties or security failures. Conversely, while civil society presence may encourage more challenging audit inquiries and practices, that presence cannot resolve conflicts between trade secrets and public disclosure per se. Ensuring that audit reports are published and that only information that is legitimately protected by trade-secret claims is excluded is one option.Footnote 156 That route to better AI public transparency would require efficient adjudication processes for auditor assertions that a claim does not have a valid basis or, alternatively, should be disclosed for compelling public-interest reasons. It should also be recognised that making public disclosure an AI audit responsibility also puts auditors in a privileged position in the mediation of transparency.Footnote 157
Third, developing public authority or agency responsibilities to report and disclose matters of public interest resulting from AI regulatory supervision and enforcement could be subject to duties to provide contestable, reasoned explanations.Footnote 158 Where full public disclosure could not be justified, there are strong precedents for more limited public disclosure, reporting on public-interest-based internal scrutiny without revealing essential protected information.Footnote 159 Yet, reliance on the state to satisfy public information needs is self-evidently an extreme form of mediated transparency, controlling the substance and form of public disclosure. Where state authorities disclose information according to internal conclusions about the extent of the public’s need to know, as we note above, the dynamic, exploratory nature of a public exercised right to know is inevitably lost. Consequently, public disclosure mechanisms controlled by the state authorities or agencies, however independent, offer a significant but inadequate strand in AI public transparency.
In our view, these areas for further research are both preferable and potential alternatives to carving out public-transparency fields from trade-secret protections. In emphasising the need for innovations at the trade-secrets barrier, these suggested alternatives do, however, illustrate the ways in which civil society intermediaries, rather than the public at large, have become central to the exercise of the right to know in the AI era. Returning to our example of AI in the workplace, even if employees, who suspect that AI assessment or monitoring is discriminatory, possessed better explanations and/or access rights, very few would have the time and knowledge needed to use those rights effectively. On the other hand, aside from trade unions, which are absent or under-resourced in many work environments, there is an obvious shortage of specialist intermediaries equipped to articulate compelling cases for the public disclosure of AI-related trade secrets affecting conditions of employment and recruitment.
Yet, this gap in representation and capacity is plainly not confined to the employment sector. The integration of complex AI systems into human life presents an enormous challenge to liberal democratic ideals of informed public knowledge and action. While trade secrets present a genuine barrier to that knowledge and action, the other side of the right to know also needs urgent attention. Further innovations in public-transparency law will need to address the intermediary problem.
VII. Conclusions
In this article, we recognise that trade-secret claims pose a serious threat to the necessary public transparency of AI systems and that a publicly defined and pursued right to know is essential to meaningful AI transparency. Nonetheless, in our view, the established model of public disclosure rights, protected types of trade secrets and recognition of context-specific compelling public interests in disclosure can be adapted to address that threat. The alternative of transparency carve-outs to trade-secrets protections would create substantial uncertainties and risk disproportionate commercial harms, while not addressing the ways in which trade-secret claims are embedded in complex and evolving AI transparency processes. Finding more effective ways to resolve conflicts between trade-secret claims and public-interest disclosure in AI contexts is, consequently, inseparable from the complicated challenges of achieving public knowledge of AI systems for essential democratic, economic and risk-prevention purposes.
The evolution of public-transparency regimes towards forms of targeted transparency offers a potentially effective route to meaningful transparency for defined purposes and, to a degree, can manage trade-secret and public-transparency conflicts. However, targeted transparency also undermines the right to know, with wider implications for the public’s capacities to contest the ordering of the public information sphere by state and private power. Securing the right to know in AI transparency processes is therefore critically important to the survival of liberal democracies and market economies in the AI era.
The enduring conflict between public transparency and trade secrecy can only be navigated through innovations in transparency that are fair, effective and efficient. We suggest three promising areas for potential innovation that are worthy of further research and development: (1) developing right-to-know transparency challenge rights that incorporate third-party review of trade-secret claims to remove trade-secret holders’ unilateral advantages; (2) developing more diversity in vetted third-party access to compliance assessment within the trade-secrets barrier; and (3) developing public authority or agency responsibilities to report and disclose matters of public interest in the course of AI regulatory supervision and enforcement. These areas do not provide a “quick-fix” solution, but they do have the potential to grapple meaningfully with the complexities and demands of public transparency and trade secrets.
Open access
