From Open-Source to Secrecy – The Implications for Protecting AI

OpenAI, the AI research company behind the chatbot service ChatGPT, made headlines last month when they decided to withhold details of how their latest language model, GPT-4, actually works.

While the company has historically been forthcoming with such details in relation to their earlier models, the technical report that accompanied the release of GPT-4 stated that it contains “no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar”.  In addition to highlighting the need to maintain a competitive advantage, the report notes that the safety implications of large-scale models like GPT-4 were a consideration for this decision.

The developer community has long championed an “open-source” principle, whereby source code is publicly accessible for others to use or modify, since it encourages collaboration and enables developers to build upon each other’s work.  However, with companies such as OpenAI (whose founding ethos was to be open-source, as their name suggests), beginning to keep the workings of their models proprietary, what does this say about the future of the open-source principle and the implications this will no doubt have on protecting AI innovations?

What approach are other companies taking?

The companies active in the field of AI have established different stances on whether, and how, to protect their AI inventions.

Google, for instance, have discussed their ongoing commitment to open-source AI, claiming that its open-source software strategy encompasses the entire “idea-to-production” lifecycle of AI models.  Google is also extremely active in filing AI-related patent applications, suggesting that it may seek to prevent others from using at least some of its AI inventions – possibly even including some of the core self-attention methods that paved the way for systems like ChatGPT.

Elsewhere, IBM announced a seismic shift in its IP strategy earlier this year: after 29 years as the most prolific patent-filer in the United States, the company announced that it will now be taking a more selective approach to patenting, while they focus more on achieving advancements in specialised areas like AI.  While IBM may not be focussed on quantifying their success through patent numbers going forward, their Senior Vice President and Director of Research Darío Gil has emphasised that “you simply can’t do innovation at scale without a thoughtful, proactive IP strategy to protect parts of what you have created”.

It therefore seems that most companies active in the field of AI are beginning to look carefully at what they consider to be the right balance between proprietary and open innovation.  It may be that we will see a greater shift in the AI industry away from open-sourcing and toward secrecy, or maybe a combination of secrecy and patents.

A shift to trade secrets?

A shift in the AI industry from open-sourcing may provide companies options of using patenting for their innovations or keeping the innovation secret.  While a patent provides a monopoly right to prevent others from using an invention, this is in exchange for the invention being disclosed to the public.  In contrast, a trade secret recognises the value of secret information in supporting business competitiveness.  As such, a shift in the AI industry from open-sourcing could lead to a heavier reliance on trade secrets as a means of protecting AI innovations, as there is no requirement to disclose the information to the public in exchange for protection.

However, a trade secret is not an exclusive right, which means that competitors are free to create independently a product protected as a trade secret.  There is also a risk that competitors may reverse-engineer a product and figure out the trade secret.  If it is impossible to reverse-engineer the product, the use of trade secrets as the backbone of a defensive competitive strategy can be relatively effective.  However, it is becoming less viable to keep AI solutions hidden as technological advances in the field are making it easier to reverse-engineer training datasets, or even the underlying AI models themselves. And of course there is always the risk that developers moving between employers may deliberately or inadvertently transport some confidential knowledge to a competitor.

There are therefore drawbacks associated with a company employing a non-disclosure protection strategy.  It could be, for instance, that such a strategy will enable a competitor to obtain patents for technology which the company are already using in their products and protecting by way of a trade secret.  In theory, it would be possible for the competitor to sue the company for infringement of these patents.

If this were to happen, the company would not be able to retaliate without having their own patents for the technology.  For example, they would not be able to counter-sue for infringement or negotiate with the competitor to avoid litigation through a cross-licensing agreement.  For these reasons, there would be no deterrent to prevent a competitor from initiating an infringement action against the company.

At least in some countries, the company may be able rely on a ‘prior user rights’ defence to continue using their own technology, which the competitor has now patented.  However, prior user rights would apply only to the technology that the company used before the priority date of the competitor’s patent.  This means that there may be some claims of the competitor’s patent in respect of which the prior user rights do not apply.  Also, any improvements that the company makes to the technology protected by prior user rights may still infringe the competitor’s patent, which could prevent them from keeping up with their competitors and place them at a commercial disadvantage.

Given the drawbacks of relying largely or solely on trade secrets, it is difficult to envisage that many companies active in the field of AI will move to such a trade-secret-heavy approach, despite the possible shift in the AI industry from open-sourcing to secrecy.  Patents will still likely play a prominent role in the protection strategy of most companies active in the field of AI.

Possible implications of a push for transparency

Notably, there is upcoming legislation that will actually affect the ability of commercial AI models to remain hidden.  The European Union, for example, is working on a proposal to regulate AI, which will likely require the inner workings of AI products used in certain industries to be made transparent in order to ensure they meet European regulatory requirements.  The question is whether this might work against the use of trade secrets or, conversely, whether the biggest obstacle for algorithmic transparency is the concept of trade secrets.

It is currently the intention that the disclosure of information will be carried out in compliance with relevant legislation in the field, including Directive 2016/943 on the protection of trade secrets, and that when public authorities and notified bodies need to be given access to confidential information or source code to examine compliance with substantial obligations, they will be placed under binding confidentiality obligations.  It could be inferred from this that the upcoming EU legislation may elevate trade secrecy above transparency obligations for AI systems by the promise of compliance with confidentiality standards.

However, Directive 2016/943 allows “the application of Union or national rules requiring trade secret holders to disclose, for reasons of public interest, information, including trade secrets, to the public or to administrative or judicial authorities for the performance of the duties of those authorities”.  Thus, it appears that “public interest” may well be a legitimate reason for AI system transparency and, according to the EU’s current proposal to regulate AI, this could include health, safety, consumer protection and the protection of other fundamental rights (‘responsible innovation’) when high-risk AI technology is developed.

So, it does appear that the upcoming legislation is likely to limit the options for companies in the AI space to keep the details of their AI systems secret.  This will no doubt make trade secrets look even less appealing, and encourage companies to patent their AI inventions, or to make a nuanced set of decisions to use a combination of both trade secrets and patents.

Final thoughts

While there may be a shift in the AI industry away from open-sourcing and toward secrecy, companies still risk facing difficulties relying on trade secrets to protect their AI innovations.  Securing patent protection for AI innovations remains key to preparing for the challenges and opportunities that the ever-evolving AI landscape will no doubt bring.