In what's considered one of the most significant parts of the world's first comprehensive artificial intelligence regulation, the European Union has gotten to a preliminary agreement that restricts the use of the ChatGPT model and many other deep learning technologies.
Bloomberg has been able to obtain a document from the EU which outlines basic transparency requirements for developers of general-purpose AI systems, which are powerful models that can be used for a wide range of purposes unless they are made available free and open-source, which is certainly allowed.
As a result of artificial intelligence becoming more widespread, it will have an enormous impact on almost every aspect of our lives. Commercial enterprises can benefit enormously from the use of digital technology, but there are also significant risks associated with it.
Such warnings have even been made by Sam Altman, the developer of the ChatGPT language model, and he also promotes these warnings. It has even been suggested by some scientists that if artificial intelligence develops applications which are aggressive beyond the control of mankind there could be a threat to our existence.
A provisional agreement reached by the European Union (EU) has marked a significant milestone for establishing the world's first comprehensive artificial intelligence (AI) regulation. It limits the operation of cutting-edge AI models, such as ChatGPT, which is one of the most advanced artificial intelligence models available today.
There are several transparency criteria, outlined in a Bloomberg report, directed at developers of general-purpose AI systems that are characterized by their versatility across different applications, as well as their ability to function effectively in a wide range of situations.
A special note needs to be made concerning the fact that these requirements do not apply to free and open-source software models.
Several stipulations must be implemented to comply with these demands, which include establishing an acceptable use policy, maintaining updated information on the model training methodology, submitting a detailed data summary used in the training, and establishing a policy to preserve copyright rights.
To comply with the regulations, models that are determined to present a "systemic risk," which is determined based on the amount of computing power they use during training, are subjected to escalation. Experts highlight the GPT-4 model of OpenAI as the only model that automatically meets this criterion and is capable of exceeding the threshold of ten trillion or septillion operations per second.
Several other models can be designated by the EU's executive arm based on factors such as the size of the dataset, EU business users, as well as the registration of end users.
For highly capable models to be compliant with the AI Act, the European Commission must refine more cohesive and enduring controls while committing to a code of conduct. When models fail to sign the code, they must demonstrate compliance with the AI Act to the commission.
Notably, there is an exception for models posing systemic risks, which is not covered under the exemption for open-source models.
This model entails a number of additional obligations, including the disclosure of energy consumption, adversarial testing either internally or externally, evaluating and mitigating systemic risk, reporting incidents, implementing cybersecurity controls, divulging information that is used for fine-tuning the model, and adhering to energy efficiency standards as needed.
Current AI models have several shortcomings
Current artificial intelligence models have several critical problems that make comprehensive regulation more necessary than ever:
Especially in critical applications such as healthcare or justice, there can be trust issues if there is a lack of transparency and explainability, which can lead to trust issues. Keeping this information safe and secure is very important, and misuse or breaches of this information can have severe consequences, making it crucial to protect it.
There has been a lot of talk about AI systems' reliability and safety, but this is a difficult task since such systems may be subject to errors or manipulation, such as adversarial attacks designed intentionally to mislead AI models. In addition, it is important to make certain AI systems are robust and are capable of working safely, even if faced with unexpected situations or data.
A significant environmental impact of large AI models can be attributed to the need for computing involved in training and operating them. Additionally, as the need for more powerful AI continues to grow, so does the amount of energy required to power them.
Last year, when OpenAI's ChatGPT was released to the public, generative AI became a hot topic in the media. As a result of that release, lawmakers were pushed to rethink their approach to 2021 when the initial EU proposals appeared.
The ability to generate sophisticated and humanlike output from simple queries using vast amounts of data using generative AI tools such as ChatGPT and Stable Diffusion, Google's Bard and Anthropic's Claude has completely surprised AI experts and regulators with these tools.
There are concerns that they might displace jobs, generate discriminatory language or violate privacy, which has sparked criticism.
There is a new era dawning for the digital ethics community, with the announcement of the EU's landmark AI regulations.
The blueprint sets a precedent for responsible AI by navigating the labyrinth of transparency, compliance, and environmental stewardship required in a truly transparent AI world. The digital world is undergoing a profound transformation that has the potential to lead to a technologically advanced future, and society is navigating the uncharted waters of artificial intelligence very carefully as it evolves.
