Hardy’s new use analysis model in the context of Generative Artificial Intelligence (“GAI”)

By: Eisern Tan Yuan Ho (陈谚柯) , Loh Chee Ming , Anice Wong , Cheah Ke Hui , Lee Li Lian , & Anonymous*

I. Introduction: Hardy’s conundrum

The study of copyright law is often closely associated with the study of the history of technology.[1] In fact, it is the development of technology that drives the growth of copyright law as it adapts to the intricate and novel technical considerations that arise therefrom. However, throughout history, the emergence of new technologies is not without friction with the old. Copyright law is not spared of this discord: the interests of the copyright owners of existing works are often pitted head-to-head with those of the pioneers and users of new technologies.

This age-old issue was considered in detail by Hardy via his “new use analysis model”,[2] wherein Hardy details out the inherent conflict between what he calls “old uses” and “new uses”. Briefly, “old uses” refer to the established and intended uses of existing copyright works over which the copyright owners thereof retain control. Meanwhile, “new uses” refer to the new and novel ways of using the same existing works made possible by the advancement of new technologies.

According to Hardy, the essence of this conflict between old uses and new uses lies in the issue of copyright infringement, or more specifically the issue of payment of royalties.

On the one hand, copyright owners of existing old-use works assert that their intellectual property rights are being infringed and exploited by such unprecedented new uses, and thus they should at least be justly compensated for this via royalty payments.[3] They contend that the new-use technology (the new technology that enables the new use) ought to be subject to the full regime of copyright liability in order to ensure that sufficient protection is accorded to their rights, and so that the economic incentives driving the production of creative works in their industry — the very underlying rationale of copyright protection — remain intact. This ensures that, despite the increasing economic significance and competition that may be posed by the new-use technology, old-use creators will continue to have adequate incentive to continue to making new old-use works, from which both the new-use technology and the society at large may benefit.[4]

On the other hand, pioneers of the fledging new-use industry contend that such an imposition of copyright liability would unduly stifle and hinder the growth of the new-use technology, much to the detriment of the public.[5] They contend that the development of new uses is not only beneficial to the advancement of mankind as a whole, but also brings numerous benefits to the authors and copyright owners of old-use works as well, which in themselves should constitute adequate incentive for the old-use industry. Hence, it follows that copyright liability should not be imposed as a matter of policy so as to allow the new-use technology sufficient “room to grow”.[6]

As rightly pointed out by Hardy, this need of copyright law to strike a balance between new uses and old uses is not novel.[7] Thus, throughout history, whenever a new-use technology emerges, the central question which confounds legislators and courts alike is whether this balance is best struck by stringently upholding the intellectual property rights of the old-use industry at the risk of stifling the new-use industry, or by conversely favouring the development of the new-use industry at the expense of the old-use industry. Today, the new-use technology that has taken the world by storm would be the emergence of Generative Artificial Intelligence (“GAI”). Although not completely novel,[8] in recent years, the advancements of GAI have begun to blur the lines between man and machine. Tasks previously thought to require human intervention are now being done autonomously by GAI. These include creative endeavours like literature,[9] artistic works,[10] online content creation,[11] musical works,[12] et cetera. In fact, GAIs are even outperforming human beings in creative thinking.[13] However, as we will see below, the development and use of GAI itself involves multiple issues of potential copyright infringement in relation to existing copyright works, leading up to the question as to how to best regulate the GAI industry given this conundrum?

II. GAI in the context of Malaysian copyright law

Before we proceed, it would be most helpful for us to have a rough idea of how GAI actually operates, in order to ascertain the possible copyright infringement issues arising therefrom, and to see how well Malaysian law is currently equipped to cope with these issues.

(a) Introduction to GAI

Generally, GAI is a type of artificial intelligence model that is capable of generating new content and data (such as images, texts, music, et cetera) in response to any prompts or commands given to it.[14] The way GAI works is via the use of a “foundation model” (a type of statistical model) created during the training of the GAI; by using this foundation model and the probability distribution data contained therein, the GAI would then be able to generate new output data by predicting what the expected and appropriate responses to a certain prompt might be.[15]

The way a GAI is created is via a process known as “deep learning”.[16] Deep learning is a form of “machine learning”[17] and involves the training of artificial neural networks (i.e. the GAI) to recognize underlying patterns in the unstructured content (input data) fed to it.[18] The training of GAI involves the use of a large amount of training data (“big data”) and various “training models” (algorithms designed to evaluate and score the accuracy and quality of the output data generated) whereby the foundation model is created by a process of trial and error by the GAI.[19] In other words, the artificial neural network generates a series of statistical algorithms, which are then tested by another algorithm (training models or “Generative AI Architectures”) for accuracy and quality; the most well performing algorithms are kept and are then used by the GAI to generate further more refined algorithms, while the rest are discarded. Eventually, a statistical algorithm capable of executing the intended task well (such as text generation, image generation, et cetera) is generated (the foundation model), which will then be used as the basic mechanism to generate the novel responses of the GAI in response to a given prompt.[20]

Generally, the involvement of copyrighted works is only necessary during the training stage of the GAI. This is because once the foundation model is created and the GAI trained, the transient copies of the training data used will be destroyed as they are no longer necessary for the GAI to perform its intended function. However, copyrighted works may nevertheless be involved subsequent to GAI training if any particular end-user concerned decides to use said works as a prompt for the GAI to generate novel content. ??

Thus, we can see that the invention of GAI has brought about a new use of existing copyright works (such as novels, films, sound recordings, et cetera), in the sense that those old-use works are now being used by developers to train GAI models to generate completely new but similar content on its own, and are occasionally being used by their users as a prompt to generate such novel content; uses which were not originally contemplated by the copyright owners and authors of those works. Such a new use brought about by GAI involves numerous potential questions of legal liability under copyright law, which are examined below.

?

(b) GAI and Malaysian copyright law

From the above, we can see that one of the most important elements in training a GAI involves the use of big data; i.e. the larger the amount and variety training data (input data) used to develop the GAI, the better it becomes at performing its intended function. However, it is often the case that the big data collected for such a purpose would inevitably involve a variety of copyrighted works. Despite this, it is rarely the case that GAI developers would have obtained prior authorization from the relevant copyright owners concerned to use the latter’s old-use works in GAI development.[21]

An issue of copyright infringement thus arises when transient and cached copies of such copyrighted works are created during the training process which may infringe the reproduction rights of the relevant copyright owners involved,[22] even though such copies are destroyed once the GAI is trained.[23] For our purposes, we will refer to this as “input infringement”.

Such input infringement may also occur where an end-user of the trained GAI uses a copyrighted work as a prompt to generate a response, which may involve uploading transient copies of the work. This occurs, for example, where a student uses GAI to summarize a copyrighted article or even to generate novel essays therefrom.[24]

Similarly, “output infringements” of the reproduction right may arise where the responses generated by GAIs closely resemble the copyrighted works used in the GAI’s original training by GAI developers or subsequent prompting by end-users, even replicating said works at times (this error is known as “overfitting”).[25] Such replication, if proven to be substantially similar to the original copyright works, would attract copyright liability on both the original user (direct infringement) and the GAI developer (secondary liability).[26]?

In Malaysia, the reproduction right is housed in s13(1)(a) of the Copyright Act 1987 (“CA1987”), which covers copies of the work in both original and derivative form. Generally, to constitute infringement of the reproduction right, three elements need to be proven:[27] (A) objective similarity (between the infringing copy and the original work), (B) causal connection (i.e. that the similarities are the result of copying), and (C) substantial infringement (which depend on the amount and originality of the part taken, the purpose of the taking, and the infringement would interfere with the sales of the original work).[28] ?

Since GAIs involve the reproduction of either exact copies (in the case of input infringement) or substantially similar copies (in the case of output infringement) of the original work, elements (A) and (B) are easily fulfilled. Meanwhile, whether element (C) is fulfilled would depend on the particular facts of the case. However, since the whole work is usually reproduced (in the case of input infringement), element (C) will usually be fulfilled as well.

Therefore, prima facie, the reproduction of any copyrighted old-use works in Malaysia by the use or development of GAI would generally require the express authorization by the copyright owners concerned to avoid copyright infringement liability under s36(1).[29]

This is, however, subject to the exceptions contained in s13(2).[30] In the context of GAI, the most relevant exception here would be that of fair dealing under s13(2)(a) read with s13(2A).[31] Despite this, it is still unclear whether the fair dealing exception under s13(2)(a) is as wide as the US doctrine of fair use.[32] Therefore, it remains to be seen whether the use of copyrighted works in the development of GAI would constitute a fair use in Malaysia. However, despite the transformative nature of GAIs, we would suggest that the likelihood of this being so is very slim, especially due to the substantial copying involved in the GAI industry[33] and the significant competition that GAI works would likely pose against old-use works on the market. Additionally, the proviso of s13(2)(a) generally requires “an acknowledgement of the title of the work and its authorship”[34] before any use can be considered as fair use. Given the voluminous amount of input data automatically collected and used in the training of GAI, the fair use defence would likely be denied to GAI developers due to this unreasonable and impractical requirement imposed.

In relation to the training of GAIs, another possible relevant exception would be s13(2)(q) in relation to transient and incidental electronic copies of a work, inter alia, necessary for the “utilization of the said work”.[35] In this regard, it may be argued that the creation of temporary and incidental copies of old-use works are necessary to enable its “utilization” in developing GAIs. However, we acknowledge that such a “utilization” may not be accepted to fall under the purview of s13(1)(q). Additionally, it is arguable that the creation of such copies is not “incidental” since GAIs (albeit of far lower quality) may still be developed by using purely uncopyrighted data.

On a side note, another possible related issue on GAIs would relate to the prohibitions with regards to the legal protection accorded to digital rights management under the CA1987.[36] Therefore, if GAI use or development involves the circumvention of technological protection measures (“TPMs”)[37] and the removal or alteration of electronic rights managements information (“ERMI”),[38] then the parties concerned may be held liable under the CA1987. This is so even where the use of copyrighted works would not constitute copyright infringement per se. Apart from the ambiguous wordings of the CA1987, the main concern here is that members of the old-use industry would begin to abuse TPMs or ERMIs as a means to further limit the access of GAI developers to the data so crucial for the advancement of the GAI industry.[39] Additionally, it is unlikely that any of the exceptions to the prohibitions in relation to TPMs and ERMIs would be applicable in the context of GAIs.[40]

From the above, we can see that Malaysian copyright law is not well-equipped to deal with the challenges posed by GAIs. In any case, we see that existing Malaysian copyright law would likely operate to the detriment of GAIs, which is undesirable given the immense benefits that it may bring.[41] As such, the question now is which direction should Malaysian copyright law be headed towards: promotion of GAI development or preservation of old-use industries?

?

III. Hardy’s new use analysis model and its application to GAI

(a) Hardy’s approach

As highlighted above,[42] the two important policy considerations in considering the rivalling interests of new uses and old uses are: (A) the benefits of allowing new-use technologies (such as GAI) to develop freely from copyright liability and (B) inversely, the benefits in ensuring that adequate incentive is maintained in the old-use industry as the new-use technology increases in economic significance.

Hardy concludes that the answer to this depends in turn on whether the new-use industry would eventually outgrow the old-use industry in terms of economic importance to a point that the omission to impose royalty obligations (i.e. copyright liability) on the former would seriously diminish the latter's incentives to continue its creative efforts. In our case, this means asking the question whether GAI would — and if so when — eventually displace the income-generating old-uses of the existing copyright works? If such displacement would never occur, then the current incentives of the old-use industry would not be affected by the new use enabled by GAI, and hence there is no need to impose copyright liability on GAI developers and users, and vice versa.[43]

The problem to this is that humans are not omniscient; our foresight is limited, and none of us could predict with absolute certainty how the new-use industry (in our case, the GAI industry) would eventually turn out or whether it will grow to a point so as to render the old-use industry obsolete. Hence, despite the conviction of some in the future of GAI, it is not for us to say for sure whether GAI will eventually replace all novelists, musicians, filmmakers, artists, and the like. Until that time comes, any fears, concerns or confidence about the potential of GAI would remain in the realm of pure speculation.

However, at the same time, it cannot be denied that there is a pressing and growing need for laws to be put in place to regulate copyright issues arising from GAI in relation to existing old use works. To solve this, Hardy proposes that we instead adopt a form of “risk assessment” (i.e. Hardy’s new use analysis), that is we should first weight the risks of adopting an approach favouring the interests of existing copyright owners against one favouring those of GAI users and developers. Based on this comparison, we should then take the path involving the least risk, erring on the side of caution. ?

?For this purpose, in his new use analysis, Hardy then details out what he terms as “Type I” and “Type II” errors.

?

(b) Type I errors

In essence, a Type I error connotes a situation where one does something which ought not to have been done. In Hardy’s analysis, this refers to the potential consequences or worst-case scenarios resulting from the erroneous imposition of copyright liability on the new-use technology (GAI) where it would have been prudent not to do so. The Type I error could further be subdivided into “Type Ia” and “Type Ib” errors.

A Type Ia error is known as a “suppression error”, so coined because of the “suppressing” effect that the imposition of copyright liability (and the attending royalty obligations) could have on the new-use industry.[44] In our case, a Type Ia error would refer to the possibility of snuffing out the GAI industry, or at the very least weakening its potential growth, via the imposition of copyright liability.[45] This is because, in order to avoid copyright liability, GAI pioneers would then have to either adopt measures to limit the use of copyrighted works, make substantial amounts of royalty payments to old-use industries, or risk the costs of lengthy litigation.

As previously mentioned, the quality of the GAIs developed largely depend on the volume and variety of training data fed to it,[46] therefore when GAI developers have limited access to old-use works available on the internet for such a purpose, the training and development of GAIs would be delayed and stifled.

Alternatively, AI developers could go out of their way and try to obtain licences from copyright owners of the old-use industry in return for royalty payments. However, this would still be very impractical, given the large amount of old-use works required to be used in the process of training a GAI (the collection of which is often done automatically); meaning that it would be very difficult, time-consuming and costly to obtain all the necessary licences for such a purpose.

Additionally, a burdensome imposition of copyright liability would likely dissuade any potential entrepreneur or investor from engaging in the field of GAI, or even chase off existing ones, due to the high costs and high risks of litigation associated with the GAI industry as compared to other forms of investments. Without such investments, many GAI developers would lack the funds to cope with the increased costs of GAI development.

Even assuming that these costs would be passed on to and borne by the end-users of GAI, this would nevertheless limit the market scope and overall use of GAI in various industries, since in many instances it would either be uneconomical or beyond the means of most industries to use GAI. Thus, even if fully developed, the benefits of GAI would only be enjoyed by the small number of wealthy companies or individuals that are able to afford it. Similarly, users may be deterred from using GAI if such use may result in unintended output infringement (perhaps due to the training data used), further impacting the existing market for GAIs.

In any case, the nett effect of a Type Ia error would be that the costs, time and effort required to develop GAIs would increase in manifold, making the venture altogether uneconomical, the effect of which would be to slow and dissuade any progress in the field of GAI. Given the immense economic potential of GAIs to boost the rates of innovation and employment productivity in Malaysia,[47] it would be unwise to suppress the growth of the GAI industry in Malaysia, especially when such technology might prove crucial for us to eventually become a developed country.[48] Thus, there is no reason why we should lag behind other countries who are actively promoting and exploiting the benefits of GAIs, including some of our neighbours.[49]

Moving on, a Type Ib error (or a “status quo” error) may also occur when too much control is given to the old-use industries to regulate the use of their works by the new-use industry. According to Hardy, a Type Ib error occurs when existing copyright owners of old uses are keen to maintain the status quo (hence, the name) of the existing system in place for them to exploit their copyrighted works. In other words, they view the new-use technology as a source of competition which would disrupt their existing sources of revenue, and might therefore deny to license their works to the new-use industry even if the latter is willing and able to pay a suitable royalty.[50]

For our purposes, a Type Ib error would occur when existing old-use creators would be reluctant to license their works for the development of GAIs, out of fear that GAIs would eventually grow to replace them.[51] As pointed out above,[52] old-use works are mainly used and required in the training and development stage of the GAI. However, once the GAI is fully trained, then the use of such old-use works is no longer required, since the GAI would then be able to generate novel content completely on its own by using what it has learned (the foundation model). This means that licences and royalty payments would be made only during the training phase of the GAI, which is temporary in nature, but not afterwards. Contrast this with e-books, for example, where the licence of the author concerned would always be required so long as the e-book is in circulation. Given this reality, it is then no longer a remote possibility for old-use creators to refuse these temporary benefits, if this would jeopardize their long-term livelihood in the process.[53] In fact, the sheer number of protests by writers,[54] actors,[55] musicians,[56] artists[57] and so on, only strengthen this point. Should such a Type Ib error materialize, then the imposition of copyright liability on GAIs would lead to a situation where the GAI industry’s progress and development would very much be subject to the whims and will of hostile old-use industries bent on limiting its potential, much to its detriment.

?

(c) Type II errors

In contrast, a Type II error refers to a situation where one omits to do something which ought to have been done. For our purposes, a Type II error thus refers to the consequences of an erroneous decision to not impose copyright liability on the GAI industry for their use of existing old-use works, where it would have been more prudent to do so.

According to Hardy, the main harm of a Type II error is that the lack of any form of royalty payment from the new-use industry to the old-use industry would deprive old-use creators of the necessary incentive to continue producing new old-use works.[58] This is especially so if the new-use industry has the potential to displace and replace the old-use industry in terms of economic importance and market competitiveness. If this occurs, then without such royalty payments, the vigour of the old-use industry will falter and wane, and the public would be deprived of what would otherwise be the benefit a greater number of old-use works.

In relation to GAI, this indeed poses a valid concern: if GAIs would eventually be able to produce creative or expressive works of greater quality on its own, at a cheaper cost and a faster pace, then this would pose significant competition and disturbance towards their human counterparts in the old-use industry, so much so that the old-use industry might die off or be immensely weakened. Indeed, we already have controversial precedents of GAIs producing award-worthy works in the fields of literature,[59] music,[60] film-making,[61] art,[62] photography,[63] et cetera; all of which are testaments to the immense potential of GAI and the corresponding threat it poses to old-use creators everywhere.

Therefore, in the face of the growing economic prowess of GAIs, the least we could do to ensure the preservation of the old-use industry is to allow the creators of old-use works to at least receive some economic gain from the successes of GAI, especially since their works contributed to its development and creation. Otherwise, failure to do so would be catastrophic to the survival of the old-use industry in the onslaught brought about by GAIs, something beneficial to no one.

?

(d) The new use analysis model and our opinion thereof

Having outline the various “errors” above, Hardy then goes on to analyse both their “likely harm” (substantiality of the harm if the relevant “error” in fact occurs) and “likely frequency” (the probability of the harm or “error” occurring) to determine what should be the path taken in moving forward.

In relation to Type Ia errors, Hardy argues that the suppression of a new-use industry isn’t necessarily a bad thing. This is because the decision whether to impose copyright liability will inevitably be at the expense of either the old use (in terms of lost royalty payments) or the new use (in terms of increased royalty costs); the only question being on whom should the burden fall? In other words, it is inevitable that the strengthening of either industry would come at the expense of the other. Therefore, rather than asking “would the new use be stifled?”, we should instead ask “whether the new use was inappropriately stifled to the detriment of the overall public?”[64]

Hardy then reduces the whole question on Type Ia errors down to a cost-benefit analysis. According to him, the decision to impose copyright liability on a new-use industry would matter the most only when the new-use industry concerned is barely profitable or with little social benefit. In other words, the more beneficial a new-use industry is, the more profitable it is, and the less impact that copyright liability (in the form of reasonable royalty payments) would have on its growth, and vice versa.

This is because if the new-use industry brings greater benefits than its costs, and is thereby destined to earn substantial profits, then in theory any obligations to make reasonable royalty payments to the old-use industry would be insufficient to stifle its growth. Conversely, if the new-use industry only brings minimal value or benefit to society as compared to its costs, then it is likely to be barely profitable and therefore destined to fail whether or not royalty obligations are imposed. In the latter scenario, Hardy goes so far as to argue that the imposition of copyright liability to hasten the failing new-use industry’s demise would in fact be beneficial to society, since this frees up the resources otherwise invested in the failing industry to other more promising and beneficial ventures.[65]

Based on the above, Hardy thus concludes that, although the probability of a Type Ia error occurring is uncertain, the magnitude of its harm is negligible.

With all due respect, this analysis is flawed as it is crucially premised upon the assumption that a new-use industry with great social benefit would be immediately profitable, and thereby be destined to succeed, of which Hardy offers no tangible proof. While this may sound plausible at first, a brief look at the history of technology would quickly shed light on this misconception. With regards to this, a good example would be the invention of the computer. While nobody would question the significance of the computer today, it is worthy to note that it took roughly a century for the computer (since its invention in the 19th century)[66] to become widely available to the public in the late 1900s.[67] In other words, it took roughly a century before the computer industry started to flourish.

The same conclusion may be drawn in relation to the GAI industry. Being in its infancy stage,[68] GAIs are still being actively developed and improved upon, and as such has yet to have its full potential widely utilized in various industries. In fact, GAI only started achieving major breakthroughs since 2022.[69] Hence, although the forecast of the GAI industry seems promising,[70] we should take care not to dampen its current momentum by suddenly imposing heavy copyright liability on its pioneers.[71]

Additionally, one crucial distinction which sets GAI apart from previous inventions and innovations is the sheer amount of copyrighted works involved in its development alone. While Hardy’s analysis on Type Ia errors may be applicable in relation to previous inventions like radios or motion pictures (which only require a few old-use works to generate a profit-making broadcast or film), the same cannot be said in relation to GAI. As noted previously,[72] the training and development of GAI requires an immense amount of data, potentially involving thousands of copyrighted works in the process, with no prospect of any immediate returns. Thus, to require a GAI developer (who has yet to fully reap the economic benefits of GAI) to obtain licences and pay royalties to the thousands of copyright owners of every possible website, online article, online video, et cetera, used during the development of the GAI would likely prove to be very impractical and even beyond his current means, especially given the already high costs involved in developing the GAI.[73]

Therefore, based on the above, we would instead conclude the Type Ia error as both probable and significant. ?

Meanwhile, on Type Ib errors, while Hardy admits that the deliberate refusal of the old-use industry to license their works to the new-use industry might lead to severe repercussions on the latter (perhaps even driving the new-use industry out of business), much to the detriment of the public, he nevertheless concludes that such instances are highly improbable.[74] This is because, according to human nature, when old-use copyright owners are given the opportunity to profit from the new-use market in the form of royalty payments, one would usually expect that they would take such opportunity, especially since it is in their self-interests to do so. Thus, applying this analysis, it could be similarly argued that it is highly unlikely for old-use copyright owners to refuse licensing their works to the GAI industry, given the additional income to be made therefrom.

However, again this analysis ignores one crucial characteristic that GAI has over other new-use technologies: the potential to replace the old-use industry altogether.

As pointed out above,[75] GAI only requires data from existing old-use works for its training and development phase, but not afterwards; meaning that once such training is complete, the GAI would then be able to autonomously generate new works which mimic the old-use works based on what it has learned without requiring any further reference to the latter. Additionally, once fully developed, GAIs have the potential to reach a point where it could render the old-use industry completely obsolete given its speed and efficiency in generating high quality works.[76] This has understandably aroused much fear and concern amongst those engaged in the old-use industry, who often question and protest the advancements made in GAIs.[77] Given this context, it is then no longer a remote possibility that those in the old-use industry would legitimately turn down any short-term profits to be made from licensing agreements with GAI developers in order to avoid the feared possibility of jeopardizing their own long-term careers. As elaborated above,[78] this would then significantly hamper the development of GAI, given its reliance on these works.

Contrast this with the emergence of previous new-use technologies, where the concern of the relevant old-use industry is not so much about it being made redundant, as it is on losing out on substantial profits to be made from the new-use industries. A good example would be the invention of e-books. E-books do not pose any threat of replacing authors of traditional printed books altogether, rather the only threat they posed was to reduce the sales and profits of the market of printed books. Therefore, when authors were allowed to obtain royalty payments from e-books, then any worry of losing out on any profits due to the e-book industry would be overcome. There would then be no reason for them to oppose the rise of e-books. But this is not the case for GAIs, which could eliminate the need for authors altogether. Given this existential threat, authors would have a greater reason of refusing any indirect help to the GAI industry, even if it means sacrificing short-term gains.

Hence, contrary to Hardy, we would conclude that a Type Ib error is both likely and significantly harmful in the context of GAI.

Finally, in relation to Type II errors, Hardy identifies three potential sources of harm.[79] Firstly, Hardy acknowledges that any failure to impose copyright liability early on (where this should have been done) would likely come at the cost of lost old-use works owing to the lack of adequate incentive to create in the old-use industry.[80] Although the legislature may subsequently remedy this by intervening to impose royalty obligations on the new use industry via compulsory licences (copyright licences mandated to be granted by law at a fixed royalty rate), the fact remains that the old-use works that would have otherwise been created in the meantime would be irreparably lost.[81] Secondly, the eventual and delayed imposition of compulsory licences would constitute a wasteful allocation of social resources due to the inherent drawbacks of price fixing involved.[82] And finally, failure to impose copyright liability early on will likely lead to substantial costs being incurred later on in relation to the decision-making process when the time comes to change the existing copyright legislation.[83] Based on this, Hardy then concludes that the above harms are both likely and substantial.

With no disrespect to the learned author, we are unable to see the force of Hardy’s arguments. Whilst we do agree that a failure to impose copyright liability (and the corresponding royalty obligations) on GAIs early on would possibly correspond to a lack of incentive for the old-use industries concerned, we are not persuaded that the imposition of such liability would necessarily provide such adequate incentive anyway. This is because, as previously mentioned, any royalty payments would be temporary in nature, and consequently the imposition is unlikely to make any significant difference towards creators who value their long-term careers over any short-term economic incentive.[84]? Instead, creators of old-use works would be much more motivated to continue their endeavours if they are given the promise of being able to maintain the profitability of their existing markets, as opposed to receiving temporary payments for the use of their works by GAIs. Thus, whether we commit a Type II error (i.e. to not impose copyright liability) or not, the result in this regard would remain the same.

On Hardy’s second point, we would also contend that compulsory licences are not the only method to subsequently remedy a Type II error; Parliament may instead impose copyright liability on GAIs without the use of compulsory licences later on (if ever) when the GAI industry is of greater strength and profitability. In any case, even if compulsory licences are used, we contend that the waste of some social resources is an acceptable risk or cost in ensuring that the potential of GAIs are not hampered, given their economic potential.[85]

Finally, on Hardy’s third point, we would like to point out that decision-making costs (in the form of lobbying costs, legislative time consumed, survey costs, et cetera) would be incurred in any event; the only question being when? It therefore makes no difference whether we change our copyright laws now, or we decide to delay this to sometime in the future. Therefore, we fail to appreciate the significance of this point raised by Hardy.

Therefore, we would conclude that a Type II error, though likely, is not significant in the context of GAIs.

IV. Conclusion

In conclusion, we are of the view that the conclusions drawn by Hardy in his new use analysis is incompatible with the problems posed by the GAI industry, and for the reasons elaborated above, we would draw a different conclusion from him by applying a similar methodology. Hence, while Hardy may argue that copyright liability should be imposed on new-use technologies, we would come to the opposite conclusion that copyright liability should not be imposed on GAIs. Hence, our copyright laws, being ill-suited to deal with GAIs, should be changed to foster the growth of the GAI industry in Malaysia.

*Author does not wish to be named.

[1] Gachago, R. (2011). The Effect of Technology on Copyright. [Master's thesis, University of Cape Town] ResearchGate. Retrieved from <doi: 10.13140/RG.2.2.24252.28803>. Site accessed on 30 Mar 2024.

[2] Hardy, T. (1999). Copyright and “New-Use” Technologies. Nova Law Review, 23(2), 659, 688.

[3] Nachiappan, A. (2023). Misuse of copyrighted music by AI companies ‘could exploit artists’. Sky News. Retrieved from < https://news.sky.com/story/misuse-of-copyrighted-music-by-ai-companies-could-exploit-artists-13027073>. Site accessed on 28 Mar 2024.

[4] See footnote 2 above.

[5] Hays, K. (2023, Nov 3). Firms like Meta and A16z admit having to pay billions for training data would ruin their generative-AI plans as they fight new copyright rules. Business Insider. Retrieved from <https://www.businessinsider.com/generative-ai-copyright-meta-google-openai-a16z-microsoft>. Site accessed on 30 Mar 2024.

[6] See footnote 2 above.

[7] See footnote 2 above.

[8] Sik, C. P. (2020). Artificial Intelligence and Copyright: The Authors' Conundrum. WIPO-WTO Colloquium Research Papers: 2018, 173, 173. Retrieved from <https://www.researchgate.net/publication/347945203_Artificial_Intelligence_and_Copyright_The_Authors%27_Conundrum>. Site accessed on 30 Mar 2024.

[9] Vainilavi?ius, J. (2023, Dec 29). AI-generated science fiction novel wins literary prize in China. Cybernews. Retrieved from <https://cybernews.com/news/ai-novel-wins-prize-china/>. Site accessed on 30 Mar 2024.

[10] Pickett-Groen, N. (2018, Jan 24). The Next Rembrandt: bringing the Old Master back to life. Medium. Retrieved from <https://medium.com/@DutchDigital/the-next-rembrandt-bringing-the-old-master-back-to-life-35dfb1653597>. Site accessed on 30 Mar 2024.

[11] Cheong, C. (2023, Aug 12). A YouTuber with 15 million subscribers has launched his own AI replacement. Some viewers are skeptical, but he thinks it'll secure his legacy. Business Insider. Retrieved from <https://www.businessinsider.com/youtuber-jordi-van-den-bussche-interview-ai-replacement-burnout-2023-8>. Site accessed on 30 Mar 2024.

[12] Smith, N., Lippiello, E., & Pereira, I. (2023, Nov 4). AI songs that mimic popular artists raising alarms in the music industry. ABC News. Retrieved from < https://abcnews.go.com/US/ai-songs-mimic-popular-artists-raising-alarms-music/story?id=104569841>. Site accessed on 30 Mar 2024.

[13] Durk, A. (2023, Sep 14). AI and Imagination: Chatbots Rival Humans in Creative Thinking. Neuroscience News. Retrieved from <https://neurosciencenews.com/ai-chatbot-creative-thinking-23920/>. Site accessed on 30 Mar 2024.

[14] Martineau, K. (2023, Apr 20) What is generative AI? IBM. Retrieved from < https://research.ibm.com/blog/what-is-generative-AI>. Site accessed on 14 Mar 2024.

[15] Géron, A. (2019). Hands-on Machine Learning with Scikit-Learn, Keras, and TensorFlow. (2nd ed.). California, USA: O’Reilly, 4.

[16] Stripling, G. [@googlecloudtech]. (2023, May 9). Introduction to Generative AI. YouTube. Retrieved from <https://www.youtube.com/watch?v=G2fqAlgmoPo>. Site accessed on 20 Mar 2024.

[17] Campesato, O. (2020). Artificial Intelligence, Machine Learning and Deep Learning. (1st ed.). Virginia, USA: Mercury Learning and Information, 19.

[18] IBM. What is deep learning? IBM. Retrieved from <https://www.ibm.com/topics/deep-learning>. Site accessed on 20 Mar 2024.

[19] Bandi, A., Adapa, P. V. S. R., & Kuchi, Y. E. V. P. K. (2023). The power of generative AI: A review of requirements, models, input–output formats, evaluation metrics, and challenges. Future Internet, 15(8), 260, 268.

[20] Grey, CGP. [@CGPGrey]. (2017, Dec 18). How AIs, like ChatGPT, Learn. YouTube. Retrieved from <https://www.youtube.com/watch?v=R9OHn5ZF4Uo>. Site accessed on 20 Mar 2024.

[21] Raj, A. (2024, Mar 12). It’s sue season as AI firms face lawsuits. Tech Wire Asia. Retrieved from <https://techwireasia.com/03/2024/its-sue-season-as-ai-firms-face-lawsuits/>. Site accessed on 15 Mar 2024.

[22] Sik, C. P. (2021). Yea or nay to artificial intelligence? More questions than answers under Malaysian copyright law. The Journal of World Intellectual Property, 1, 3. Retrieved from < https://doi.org/10.1111/jwip.12196>. Site accessed on 15 Mar 2024.

[23] Sandiumenge, I. (2023) Copyright Implications of the Use of Generative AI. [Master's thesis, Universitat Pompeu Fabra] SSRN. Retrieved from <https://ssrn.com/abstract=4531912>. Site accessed on 30 Mar 2024.

[24] Sayed Fayaz Ahmad, Han, H., Muhammad Mansoor Alam, Mohd. Khairul Rehmat, Muhammad Irshad, Arrano-Munoz, M., & Ariza-Montes, A. (2023). Impact of artificial intelligence on human loss in decision making, laziness and safety in education. Humanities & Social Sciences Communications. 311, 1, 1. Retrieved from <https://www.nature.com/articles/s41599-023-01787-8>. Site accessed on 17 Mar 2024.

[25] Belkin, M., Hsu, D. J., & Mitra, P. (2018). Overfitting or perfect fitting? Risk bounds for classification and regression rules that interpolate. Advances in neural information processing systems, 31.

[26] Benhamou, Y., & Andrijevic, A. (2022). Research handbook on intellectual property and artificial intelligence?(1st ed.). Cheltenham, UK: Edward Elgar Publishing, 206; Copyright Act 1987 (Act 332) (Malaysia), ss 13(1).

[27] Hexagon Tower Sdn Bhd v Polydamic Holdings Sdn Bhd [2005] 1 LNS 77.

[28] Longman Malaysia Sdn Bhd v Pustaka Delta Pelajaran Sdn Bhd [1987] 2 MLJ 359.

[29] Copyright Act 1987 (Act 332) (Malaysia), s 36(1).

[30] Copyright Act 1987 (Act 332) (Malaysia), s 13(2).

[31] Copyright Act 1987 (Act 332) (Malaysia), s 13(2)(a) & (2A).

[32] See footnote 22 above, page 6 & 8.

[33] Sobel, B. L. W. (2017) Artificial intelligence's Fair Use Crisis. Columbia Journal of Law & The Arts, 41, 45, 45. Retrieved from <https://www.bensobel.org/files/articles/41.1_Sobel-FINAL.pdf>. Site accessed on 15 Mar 2024.

[34] Copyright Act 1987 (Act 332) (Malaysia), s 13(2)(a) proviso.

[35] Copyright Act 1987 (Act 332) (Malaysia), s 13(2)(q).

[36] See footnote 22 above, page 9-12.

[37] Copyright Act 1987 (Act 332) (Malaysia), ss 36A(1).

[38] Copyright Act 1987 (Act 332) (Malaysia), ss 36B(1).

[39] See footnote 22 above, page 10.

[40] See footnote 22 above, page 11-12.

[41] See below.

[42] See Part I above “Introduction: Hardy’s conundrum”.

[43] See footnote 2 above, page 692.

[44] See footnote 2 above, page 694.

[45] Estes, A. C. (2024, Jan 18). How copyright lawsuits could kill OpenAI. Vox. Retrieved from <https://www.vox.com/technology/2024/1/18/24041598/openai-new-york-times-copyright-lawsuit-napster-google-sony>. Site accessed on 15 Apr 2024.

[46] See pages 6 & 8 above, as well as Figure 2 in page 7.

[47] Gnaneswaran, D. (2019, Apr 2). Artificial intelligence to nearly double the rate of innovation in Malaysia by 2021: Microsoft study. Microsoft Malaysia News Center. Retrieved from < https://news.microsoft.com/en-my/2019/04/02/artificial-intelligence-to-nearly-double-the-rate-of-innovation-in-malaysia-by-2021-microsoft-study/>. Site accessed on 20 Apr 2024.

[48] Azanis Shahila Aman. (2023, Sep 15). Generative AI to unlock over US$113bil productive capacity for Malaysia. New Straits Times. Retrieved from <https://www.nst.com.my/business/2023/09/955655/generative-ai-unlock-over-us113bil-productive-capacity-malaysia>. ?Site accessed on 23 Apr 2024.

[49] Poloczek, C. (2023, May 27). Unveiling the Vanguard: Countries Leading the Way in Generative AI Development. LinkedIn. Retrieved from < https://www.dhirubhai.net/pulse/unveiling-vanguard-countries-leading-way-generative-ai-poloczek/>. Site accessed on 23 Apr 2024.

[50] See footnote 2 above, page 694.

[51] KrASIA Connection. (2024, Feb 29). China finds itself in copyright storm as creatives protest issues with AI-generated content. KrASIA. Retrieved from < https://kr-asia.com/china-finds-itself-in-copyright-storm-as-creatives-protest-issues-with-ai-generated-content>. Site accessed on 14 Apr 2024.

[52] See above.

[53] ?Edwards, B. (2024, Mar 4). Billie Eilish, Pearl Jam, 200 artists say AI poses existential threat to their livelihoods. Ars Technica. Retrieved from < https://arstechnica.com/information-technology/2024/04/billie-eilish-pearl-jam-200-artists-say-ai-poses-existential-threat-to-their-livelihoods/>. Site accessed on 14 Apr 2024.

[54] Merchant, B. (2023, May 11). Column: Your boss wants AI to replace you. The writers’ strike shows how to fight back. Los Angeles Times. Retrieved from < https://www.latimes.com/business/technology/story/2023-05-11/column-the-writers-strike-is-only-the-beginning-a-rebellion-against-ai-is-underway>. Site accessed on 14 Apr 2024.

[55] Collier, Kevin. (2023, Jul 15). ?Actors vs. AI: Strike brings focus to emerging use of advanced tech. NBC News. Retrieved from <https://www.nbcnews.com/tech/tech-news/hollywood-actor-sag-aftra-ai-artificial-intelligence-strike-rcna94191>. Site accessed on 14 Apr 2024.

[56]? See footnote 53 above.

[57] Okunyt?, P. (2024, Apr 4).? Artists rise against AI: stop devaluing our music. Cybernews. Retrieved from <https://cybernews.com/news/artists-protest-generative-ai-music/>. Site accessed on 15 Apr 2024.

[58] See footnote 2 above, page 695.

[59] Choi, C., & Annio, F. (2024, Jan 19). The winner of a prestigious Japanese literary award has confirmed AI helped write her book. CNN. Retrieved from < https://edition.cnn.com/2024/01/19/style/rie-kudan-akutagawa-prize-chatgpt/index.html>. Site accessed on 16 Apr 2024.; Chik, H. (2023, Dec 20) A Chinese professor used AI to write a science fiction novel. Then it was a winner in a national competition. South China Morning Post. Retrieved from <https://www.scmp.com/news/china/science/article/3245725/chinese-professor-used-ai-write-science-fiction-novel-then-it-won-national-award>. Site accessed on 16 Apr 2024.

[60] Reuter, D. (2023, Jul 5). Music containing AI-generated elements could win a Grammy as long as a human contributes in 'a meaningful way,' Recording Academy CEO says. Business Insider. Retrieved from <https://www.businessinsider.com/ai-generated-music-could-win-grammy-academy-ceo-says-2023-7>. Site accessed on 14 Apr 2024.

[61] Slipiec, C. (2024, Feb 1). SIAT Lecturer Wins Best Animation at the California Music Video Awards with Generative AI Filmmaking.? Simon Fraser University. Retrieved from < https://www.sfu.ca/siat/news-events/news/2024/02/siat-lecturer-wins-best-animation-at-the-california-music-video-.html>. Site accessed on 14 Apr 2024.

[62] Rose, K. (2022, Sept 2). An A.I.-Generated Picture Won an Art Prize. Artists Aren’t Happy. The New York Times. Retrieved from <https://www.nytimes.com/2022/09/02/technology/ai-artificial-intelligence-artists.html>. Site accessed on 10 Apr 2024.; Taylor, J. (2024, Feb 2). ?Australian ‘contemporary’ portrait prize allows entries wholly generated by AI. The Guardian. ?Retrieved from <https://www.theguardian.com/technology/2024/feb/02/brisbane-portrait-prize-ai-allowed-rule>. Site accessed on 16 Apr 2024.

[63] Greenberger, A. (2023, Apr 17). Artist Wins Photography Contest After Submitting AI-Generated Image, Then Forfeits Prize. ARTnews. Retrieved from <https://www.artnews.com/art-news/news/ai-generated-image-world-photography-organization-contest-artist-declines-award-1234664549/>. Site accessed on 16 Apr 2024.

[64] See footnote 2 above, page 696.

[65] See footnote 64 above.

[66] Williamson, T. (2023, Dec 22). History of computers: A brief timeline. Live Science. Retrieved from <https://www.livescience.com/20718-computer-history.html>. Site accessed on 16 Apr 2024.

[67] Garner, R. (2015, Jan 9). Early Popular Computers, 1950 – 1970. ETHW. Retrieved from <https://ethw.org/Early_Popular_Computers,_1950_-_1970>. Site accessed on 16 Apr 2024.

[68] NTT DATA. (2022) Generative AI Unveiled: Its Evolution, Capabilities, Limitations and Potential. NTT Data. Retrieved from < https://www.nttdata.com/global/en/insights/focus/generative-ai-unveiled#:~:text=Yet%2C%20for%20all%20the%20hype,amplified%20creativity%20on%20the%20horizon>. Site accessed on 16 Apr 2024.

[69] S?tra, H. S. (2023). Generative AI: Here to stay, but for good? Technology in Society, 75(102372), 1, 1. Retrieved from < https://www.sciencedirect.com/science/article/pii/S0160791X2300177X>. Site accessed on 10 Apr 2024.

[70] Statista. (2024) Generative artificial intelligence (AI) revenue worldwide from 2020 with forecast until 2032. Statista. Retrieved from < https://www.statista.com/statistics/1417151/generative-ai-revenue-worldwide/>. Site accessed on 10 Apr 2024.

[71] See Part II(b) “Type I errors” above, particularly in relation to our discussion on Type Ia errors.

[72] See above, as well as Figure 2.

[73] Vanian, J. (2023, Apr 17). ChatGPT and generative AI are booming, but the costs can be extraordinary. CNBC. Retrieved from < https://www.cnbc.com/2023/03/13/chatgpt-and-generative-ai-are-booming-but-at-a-very-expensive-price.html>. Site accessed on 10 Apr 2024.

[74] See footnote 2 above, page 697.

[75] See above.

[76] For examples, see footnotes 9 – 13.

[77] For examples, see footnotes 54 – 57.

[78] See Part II(b) “Type I errors” above, in particular our discussion on “Type Ib errors”.

[79] See footnote 2 above, page 702.

[80] See footnote 2 above, page 698 & 702.

[81] See footnote 2 above, page 699.

[82] See footnote 2 above, page 699, 700, 702 & 703.

[83] See footnote 2 above, page 702 & 703.

[84] See above.

[85] See above.