We need smart IP AI laws (2023)

Once a backlog of speculationArtificial IntelligenceNow, with all the hopes and fears of revolutionary technological change, people's hair is on fire. The uncertainty surrounding these intelligent systems and their regulation is huge - in some areas they already exceed human capabilities. Only by making the right decisions to protect or control the technology can we hope for the benefits of artificial intelligence for science, medicine, and a generally better life that will overcome lingering fears of fatalism.

public performanceAI chatbots like OpenAI's ChatGPTThe past year has prompted major warnings. That included a point from New York Senate Majority Leader Chuck Schumer, who said artificial intelligence will "bring massive changes to the workplace, the classroom and our living rooms..."in almost every corner of life", and Russian President Vladimir Putin claimed: "Whoever becomes a leader in this areaBecome the ruler of the world“.This concern also includesWarning of dire consequencesUnlimited AI from an industry leader.

Legislative work to resolve these issues has already begun. On June 14, the European Parliament voted in favor of itNew law on artificial intelligenceFollowing the approval of the European Commission's 771 amendments to the 69-page proposal. The bill requires "engineered" AI systems like ChatGPT to implement a number of safeguards and disclosures, such as systems that "use subliminal technologies outside of human consciousness" or "the vulnerability of certain populations due to their age, whether physical or mental . exploit" disability" and avoiding "foreseeable risks to health, safety, fundamental rights, the environment, democracy and the rule of law".

A pressing global question is whether data used to train AI systems will require the consent of the authors or performers, who also require attribution and compensation for the use of their works.

Some governments have created special copyright exemptions for text and data mining to facilitate their collection and useInformation about training AI. This allows some systems to be trained on online texts, images and other works of others. These exceptions have recently met resistance, particularly from copyright holders and critics in general who want to slow down or degrade the quality of services. In recent months, there has been an explosion in reporting on the dangers of artificial intelligence, fueling further controversy over the technology's potential risks of bias, social manipulation, loss of income and employment, misinformation, fraud and other risks such as "end of humanity".

Current US Copyright LawhearingsThis reflects a common sentiment among authors, artists, and performers: AI training data must adhere to the "3Cs" of consent, acknowledgment, and compensation. Each C has its own practical challenges that contrast with the more favorable exceptions for text and data mining accepted by some countries.

Countries have different and evolving approaches to dealing with intellectual property rights related to training data. The US has faced several lawsuits to determine the extentFair use exceptionCopyrights apply. The 2019 European Union (EU) Directive on Copyright in the Digital Single Market provides exemptions for text and data mining, including mandatory exemptions for research and cultural heritage organisations, while giving copyright holders the right to exploit their projects for commercial services. In 2022, the UK proposed a sweeping exemption for commercial use, but thenin the waiting queueearlier this year. In 2021, Singapore created an exemption for computer-aided data analysis in copyright law, which applies to text and data mining, data analysis and machine learning.Singapore is an exceptionAccess to data is required by law, but cannot be contractually regulated. China issued a statement saying that "content that violates intellectual property rights" will be excluded from the training data. According to an April article by the Stanford DigiChina Project:Described by Helen Toner, Georgetown University Center for Security and Emerging TechnologiesThis is "somewhat opaque, as the copyright status of much of the relevant data—often derived largely from a variety of online sources—is ambiguous." Many countries do not have specific exemptions for text and data mining, but have not yet adopted them in a position. Indian officials say they are unwilling to regulate AI at this time, but like many other countries, India is keen to support a domestic industry.

In formulating laws and regulations, care should be taken to avoid a one-size-fits-all approach where rules that apply to recorded music or art also apply to scientific work and data used for medical research and development.

Previous legislative work on databases has shown that caution is needed. 1990sDistributed sentencesAutomatic granting of rights to information extracted from databases, including statistical data and other non-copyrighted data. An example is a treaty proposed by the World Intellectual Property Organization (WIPO) in 1996. In the US,Diversity-Allianzscholars,libraryAmateur genealogists and public interest groups opposed the proposed treaty. But perhaps more important is the resistance from American companies, such as e.gBloomberg,Dun & BradstreetAndstatistical dataThey argue that database contracts are both unnecessary and burdensome because they increase the licensing overhead of procuring and making available to customers the data they need and, in some cases, create unnecessary monopolies. TheWIPO Database TreatyIt failed at a diplomatic conference in 1996, and subsequent attempts to pass legislation in the US failed, but so did the EU. Implementation of the Directive on the Legal Protection of Databases. Investment in databases has exploded in the United States over the decades, as has investment in databases in the European Union. Attempts to weaken his mandate through court rulings. It's 2005internal evaluation results"The tool showed no output impact on the database."

Pure practicality points to another caveat. The amount of data in large language models can be difficult to understand. The first version of Stable Diffusion, capable of generating images from text, requires training2.3 billion images. GPT-2, an early version of the ChatGPT model, was trained on 40 GB of data. A later version, GPT-3, was trained on 45 TB of data,1,000 times or more. OpenAI, opposite lefttrialIn terms of data usage, the exact size of the data set used for training has not been made publicThe latest version, GPT-4. Even for simple works, it can be difficult to clarify rights for copyrighted works, and for very large works or platforms, it can be difficult to consider searching for metadata and evaluating contracts between creators or performers and publishers. The practical requirements and even the challenge of knowing who owns it are almost impossible. In science, requiring consent to use copyrighted works can give publishers of scientific articles significant leverage to allow companies to use the data, even if most authors are not paid.

Distinguishing who owns what is important. It's one thing to convince copyright holders of popular music recordings to opt out of a database. Excluding copyright holders of popular music recordings from a database is another matter. The situation is different if an important scientific work is not carried out due to a license challenge. When using AI in hospitals and gene therapy, do you really want to exclude relevant information from the training database?

In addition to consent, the other two "c's", credit and compensation, also present their own challenges, as shown by the current high cost of suing for copyright or patent infringement. But one can also imagine datasets and their use in the arts or biomedical research, where a well-managed AI program could facilitate benefit sharing, such as the one proposedOpen source bonusIntroduction of successful biomedical products.

In some cases, the data used to train the AI ​​can be decentralized with a number of safeguards. These include enforcing privacy, avoiding unnecessary monopolistic control, and useThe "Data Space" method.Now let's build on some scientific facts.

All of this poses obvious challenges to any kind of intellectual property rights assigned to education data: while these rights are national in nature, the race to develop AI services is global. AI programs can run anywhere there is electricity and internet access. You don't need a large staff or a special laboratory. Companies operating in countries that impose costly or impractical obligations to collect and use data for AI training will compete with companies operating in more permissive environments.

If anyone thinks differently about the future of artificial intelligence than Vladimir Putin, then this is food for thought.

This is an opinion and analysis article and does not necessarily reflect the views expressed by the authorscientific American.