Since last November, the world has become a different place, even if we’re not yet able to fully grasp the implications. ChatGPT, the most popular solution of “generative artificial intelligence”, is fundamentally changing how we generate and develop information, insight, and knowledge in a context manner. We are leap-frogging from searching and reproducing information available on the Internet to a new era of creating content through machine learning at a marginal cost. Algorithms produce articles, reviews, images, videos and, conversations that have the feel and persuasiveness of human statements without any further correction or moderation by humans. Any user can generate a substantive essay of 25,000 words in seconds through the “prompts” made by just a few keywords, sounds or images. Additional tools are available to translate, refine, and “customize” the output – resulting in seemingly individual comments from users that improve search engine results and product rankings, just to name two very limited examples.
The tsunami of the Information Society
Within five months, ChatGPT has become the world’s fastest scaling application – a veritable tsunami sweeping our information society. The obvious usage as a “smart typewriter assistant” could be persuasive enough, all the more so in light of the dramatic shortage of knowledge-based workers in many sectors and geographies. But ChatGPT obviously goes far beyond that. It generates insight and contextual information in a non-deterministic manner through smart, open-ended prompts.
Of course, Chat GPT-4 does not question the uniqueness of human behavior, nor does it raise the question of “singularity”. The system neither thinks or expresses emotions, nor does it provide definitive clarity on right or wrong, true or false. In effect, it only calculates the likely sequence and continuation of text in human language. The results are artificial texts of impressive quality, even if in many cases the content may still contain apparent errors. But as the underlying database grows, the output will improve and obvious errors will become less visible. As a result, the boundaries and distinctiveness of human and machine generated texts will blur and disappear over time – and trust in artificial texts will increase: “I guess it should be right – well, it has to be right!”.
How do we seize the opportunities, try things out, and accelerate innovation and overall digitalization? What kind of regulation do we need to prevent abuse and manipulation? How do we avoid putting concerns before opportunities?
Data protection law – stumbling block to innovation?
The Italian data protection authority has banned the operation of ChatGPT nationwide. Personal data in training data, the lack of a legal basis, lack of transparency, and problems involving data accuracy and age control were seen to stand in the way of using this technology. This drumbeat of data protection is possibly not more than a short-sighted reflex – as demonstrated by the authority itself, which lifted the ban after OpenAI made some short-term adjustments in ChatGPT. Nevertheless, questions concerning profiling, processing sensitive data, and indirect data collection remain an issue. It remains to be seen how other data protection authorities will act going forward. In many cases, processing personal data remains irrelevant, so that a complete ban on ChatGPT would most likely not withstand judicial scrutiny.
In fact, ChatGPT reveals that data protection law needs to be revisited. The principle that processing personal data is prohibited unless the controller can argue a legal basis forces the controller to typically resort to the reason of “legitimate interest”. By contrast, “informed consent” is no longer feasible for AI-based data processing, given that processing at the level of the individual data record takes place in a “black box”. The requirement of “data minimization” has become obsolete as a criterion for corrective action, if we are willing to accept that the GDPR applies in a technology-neutral, innovation-oriented manner. Data processing in third countries remains a problem – the effort required for impact assessments and additional protective measures is considerable. That said, data protection authorities should examine the functioning of generative AI in the context of certifications and, if necessary, confirm the same through certifications and conformity seals. In return, data protection authorities must finally provide reliable guidance for the pseudonymization and anonymization of personal data.
The end of copyright
Copyright protects only the expression of a human intellectual creation. The prompt “Create a poem in the style of T.S. Eliot” lacks originality from a copyright perspective, whereas the artificial text produced by ChatGPT lacks the human creative act, as long as it does not directly integrate identical text elements from an authentic poem. As a result, the outcome is not protected by most copyright laws and, hence, would need to be considered as unrestricted or “in the public domain” once it is has been published. In addition, the more artificial texts flow into the training data, the smaller the remaining substrate of human creative activity will be. Public domain texts will increase rapidly without being clearly distinguishable from human works. In software developed by ChatGPT, special tools might be helpful – such as in open-source software – to identify “snippets” of human-developed, copyrighted code as a matter of quality control. In the case of artificial texts, however, this will be increasingly difficult if the output of ChatGPT does not contain identical text elements copied from human works.
Outside literary works, the creative author (and their copyright protection) may soon belong to a dwindling minority, particularly in light of a temptation to create texts without indicating ChatGPT as a source or co-contributor to a text. Notably, ancillary rights do not exist to protect artificial texts against adaptation, modification or exploitation. Thinking it through, this could unhinge copyright law to a considerable extent in a rather short period of time– with consequences that are difficult to imagine.
By the same token, verifying natural authorship is becoming increasingly difficult. We will need to integrate a rules-based use of generative AI in schools and teaching, and change the way we teach and perform practical training – with all the practical and cultural challenges for teachers and trainers that this involves. Anything else, even the common play for time, would be negligent and call into question our ability to remain innovative.
Concern about over- and under-regulation
The EU Commission is targeting “high-risk systems” with its planned AI Act. Extending tight regulation on “general purpose AI” could inhibit innovation. The Digital Services Act seeks to prevent illegal content on social media, etc., and hold digital service providers accountable. Individual users could abuse ChatGPT for mass hate speech, fake news, etc., before this appears on relevant platforms. Any regulation should be cautiously developed to address professional content providers, along with further transparency requirements. However, even the best-intended transparency requirements will be exhausted if ChatGPT becomes the norm for preparing texts and other forms of output.
Generative AI and competition law
While the planned regulatory framework that is mentioned above is addressing security concerns for consumers and the liability of providers, it is unclear to what extent it will be capable of preventing risks to competition. With various generative AI model providers growing incredibly fast and leading big tech companies such as Microsoft or Alphabet making their expansion into the AI market a top priority, the question arises whether and to what extent the European and national rules against anticompetitive behavior of so-called “digital gatekeepers” will apply to generative AI providers. At the EU level, the Digital Markets Act (DMA) imposes various obligations and prohibitions on core digital platform services with a gatekeeper function in order to prevent them from leveraging their strong market positions into up- or downstream markets. Generative AI models are data-driven and require access to large amounts of data. Also, they generate high economies of scale. Due to their unrivalled access to good quality data, their processing capacities and financial strength, various existing big tech players may as a result quickly build a position that allows them to tip the market in their favor.
Once identified as gatekeepers, under the DMA rules providers of generative AI models would be subject to obligations, such as guaranteeing interoperability, ensuring access to data provided or generated by business users, refraining from self-preferencing, refraining from using data provided by business users to improve their own services, and refraining from cross-service use of personal end user data.
On a national level, the German Federal Cartel Office has in recent months focused on identifying the first big tech companies that hold – similarly to digital gatekeepers under the DMA – a paramount significance across markets for which under the new rules for digital gatekeepers (in particular § 19a of the Act against Restraints of Competition (ARC)) very strict regulations against anti-competitive behavior (including a prohibition of self-preferencing and access to data) will apply.
Hence, regulators in the EU will put business models of digital gatekeepers that integrate generative AI models in their offerings on neighbored markets under high scrutiny.
Consequences for companies and legal services
ChatGPT provides impressive efficiency gains for text-based workflows. Companies must develop guidelines to manage compliance requirements when setting prompts (do not enter personal data or confidential information), and implement quality control processes with regard to the output. In addition, ChatGPT triggers fundamental questions about workforce management, i.e., in which areas ChatGPT will help alleviate the shortage of skilled workers, replace existing activities and require new skill profiles to be developed.
In legal services, we will see a dynamic shift away from the traditional “hand-made approach” towards increased prefabrication and automation of legal advice. However, this will not replace the lawyer’s assessment of the individual case or their professional liability (!). ChatGPT will also help alleviate the dramatic decline of qualified jurists in the judiciary and public administration – if used in a proper, rules-based manner. In that context, the real challenge over time will be to be mindful of the “automation bias” (“this should be right/it has to be right!”) – both for the recipients and creators of legal advice and documents. This is clearly an area where the legal profession will need to develop clear rules and quality control standards going forward.