This week, the White House made an exciting announcement regarding the management of risks posed by artificial intelligence (A.I.). They have successfully obtained “voluntary commitments” from seven leading A.I. companies, including Amazon, Anthropic, Google, Inflection, Meta, Microsoft, and OpenAI. This achievement is significant due to the fact that these companies are fierce competitors with slightly different approaches to A.I. research and development.
One example is Meta, which is incredibly eager to provide its A.I. models to developers. They have made many of their models open-source, allowing anyone to utilize their code. On the other hand, companies like Anthropic have adopted a more cautious strategy, releasing their technology in a more limited manner.
But what do these commitments actually entail? And how likely are they to bring about significant changes in the operations of A.I. companies, considering that they are not legally binding?
Given the potential impact of A.I. regulation, it is essential to examine the details of these commitments and assess their potential effects.
Commitment 1: The companies pledge to conduct both internal and external security testing of their A.I. systems prior to release.
These A.I. companies already perform security testing, commonly referred to as “red-teaming,” on their models before making them available. This commitment, while not entirely new, lacks specific details about the type of testing required and who will be performing it.
According to the White House, independent experts will participate in the testing of A.I. models, focusing on risks such as biosecurity, cybersecurity, and societal impacts. It is beneficial to have A.I. companies publicly commit to ongoing testing and encourage transparency throughout the process. Government and military officials are better suited to evaluate certain A.I. risks, such as the potential use of A.I. models for bioweapon development. Ideally, the A.I. industry should establish a standardized set of safety tests, such as the “autonomous replication” tests conducted by the Alignment Research Center. It would also be advantageous for the government to provide funding for these expensive tests, which demand skilled engineers. Currently, companies often fund and oversee safety tests, raising concerns about conflicts of interest.
Commitment 2: The companies commit to sharing information on managing A.I. risks with industry peers, governments, civil society, and academia.
This commitment is somewhat ambiguous. Several of these companies already publish information about their A.I. models, typically through academic papers or corporate blog posts. Some, like OpenAI and Anthropic, also release system cards detailing the safety measures implemented.
However, companies occasionally withhold information citing safety concerns, such as OpenAI’s decision not to disclose data amounts and model sizes for its latest A.I. model, GPT-4. They expressed concerns about competition and safety, but it is also information that tech companies prefer to keep away from rivals.
Under these new commitments, will companies be obligated to disclose such information? Could it accelerate the A.I. arms race? The White House likely aims to encourage information sharing among A.I. companies, specifically regarding the risks associated with their models. Nevertheless, sharing such information poses additional risks. For instance, if Google’s A.I. team prevents the misuse of a model for engineering bioweapons, should they share that information externally? Doing so may provide bad actors with ideas on using less guarded models for similar tasks.
Commitment 3: The companies commit to investing in cybersecurity and insider-threat safeguards for protecting proprietary and unreleased model weights.
This commitment is straightforward and uncontroversial within the A.I. community. “Model weights” refer to the mathematical instructions that enable A.I. models to function. These weights are valuable and need protection to prevent foreign governments or rival corporations from stealing them. A.I. companies have a vested interest in tightly controlling access to the model weights. There have been cases of model weights being leaked, such as Meta’s LLaMA language model weights getting leaked shortly after their public release. Considering the risk of further leaks and the potential interest other nations may have in stealing U.S. companies’ A.I. technology, requesting increased investment in security measures is a logical move.
Commitment 4: The companies commit to facilitating third-party discovery and reporting of vulnerabilities in their A.I. systems.
The meaning of this commitment is somewhat unclear. A.I. companies often discover vulnerabilities in their models after release, mainly due to users attempting malicious activities or circumventing guardrails in ways not anticipated by the companies.
The White House’s commitment calls for companies to establish a “robust reporting mechanism” for such vulnerabilities, but the specifics are yet to be defined. Will it resemble an in-app feedback button like those used by Facebook and Twitter, allowing users to report rule-breaking posts? Or will it be a bug bounty program similar to OpenAI’s initiative, rewarding users who identify flaws in their systems? Further details are required to fully grasp this commitment.
Commitment 5: The companies commit to developing technical mechanisms to ensure users can distinguish A.I.-generated content, such as implementing watermarking systems.
This is an intriguing proposal, but it leaves significant room for interpretation. A.I. companies have faced challenges in creating tools that allow users to determine whether content is A.I.-generated. Technical limitations hinder the accurate detection of A.I. outputs, which becomes problematic when individuals pass off A.I.-generated work as their own, as teachers often encounter. Many of the currently promoted tools claiming A.I. detection capabilities lack reliable accuracy. While the problem may not have a perfect solution, it is reassuring that companies are committing to addressing it.
Commitment 6: The companies commit to publicly reporting the capabilities, limitations, and appropriate/inappropriate use of their A.I. systems.
This commitment sounds sensible but lacks specificity. How frequently must companies report on system capabilities and limitations? How detailed must the information be? Additionally, since many A.I. companies have been surprised by their systems’ capabilities after their release, accurately describing them in advance becomes challenging.
Commitment 7: The companies commit to prioritizing research on societal risks, such as avoiding bias, discrimination, and protecting privacy.
Committing to prioritize research is a vague commitment. Nevertheless, it will likely be welcomed by the A.I. ethics community, which advocates for addressing near-term harms like bias and discrimination over concerns related to doomsday scenarios, as emphasized by A.I. safety experts. The A.I. research community is divided into two factions, each believing the other focuses on preventing the wrong types of harms.
Commitment 8: The companies commit to developing and deploying advanced A.I. systems to tackle significant societal challenges.
The notion of using advanced A.I. to tackle society’s greatest challenges is widely accepted. The White House specifically mentions cancer prevention and mitigating climate change as areas where A.I. companies should concentrate their efforts. They’ll enjoy immense support…
Denial of responsibility! VigourTimes is an automatic aggregator of Global media. In each content, the hyperlink to the primary source is specified. All trademarks belong to their rightful owners, and all materials to their authors. For any complaint, please reach us at – [email protected]. We will take necessary action within 24 hours.
