Advancements in AI have sparked intense discussions about its impact on various aspects of society. While AI systems can streamline tasks, uncover scientific breakthroughs, and reduce bureaucracy, concerns have arisen regarding job automation, AI weaponization, and biases within the models. The fear of rapid AI development is the primary obstacle to fully harnessing its benefits. The public is rightfully concerned about the speed of development of AI, and that fear actually represents the largest obstacle to reaping AI’s multiple benefits.
To alleviate AI-related concerns, organizations can promote transparency regarding their system attributes and processes, instilling confidence among customers, regulators, and the public. However, in the competitive landscape of the ongoing AI race, companies often hesitate to disclose extensive information to protect trade secrets and avoid legal repercussions. As such, striking a balance between transparency and safeguarding trade secrets poses a significant challenge for AI companies.
To establish trust and drive adoption, organizations can initiate the trust-building process by disclosing key information about their AI systems. Here are five attributes that can be disclosed without compromising trade secrets or inviting legal risks. These attributes focus on building public trust rather than divulging technical creation details and serve as a foundational set of low-risk disclosures that any organization can implement promptly.
1- Training Data Timeframe & Timeliness
The timeliness of the model training data is one of the largest limiting factors to how well a model may function for various tasks. For example, a healthcare model trained solely on pre-2020 data would lack information about Covid-19, limiting its usability. Additionally, how often a model is updated with new information will also shape what expectations users should have for it. Disclosing data collection and updating time frames can offer essential context and limitations without revealing your “secret sauce” or creating substantial legal risk.
2- Training Data General Distributions
While the exact datasets used for a model may be confidential, it is generally safe to disclose high-level descriptive statistics to highlight potential limitations. For instance, while state of the art Large Language Models (LLMs) are capable of machine translation, if the model lacks sufficient data for a particular language, its performance in translating that language cannot be relied upon. Similarly, indicating whether the training data covers specific domains like healthcare or law helps convey potential limitations when using the model in those domains.
3- Model Known Uses & Limitations
While prominent models garner attention for their extensive capabilities, disclosing the specific tasks evaluated or implemented by partners can provide insights into their expected functionality. Along similar lines, disclosing which use cases have been attempted that have had issues and limitations is equally important. The disclosure of these doesn’t necessarily have to be in depth, but addressing the public demand for documented successful and unsuccessful use cases is important to establish trust.
4- Monitoring Priorities & Feedback Processes
With great power comes great responsibility. Organizations providing powerful AI systems should prove to the public that they are responsible stewards. In turn, the public should use their power as consumers to reward the most responsible organizations. That can only be done if we know which orgs are being the most responsible. While technology providers track user behavior, it is crucial to disclose the specific types of abuses, misuses, or erroneous outputs they monitor. This builds confidence in their responsible practices and invites the public to contribute ideas for identifying new exploits that should be addressed.
5- Model Release Criteria
To avoid excessive regulation that hampers innovation, regulators are looking at regulating the people and the processes involved with AI development, instead of regulating the structure and outputs of AI models themselves. To prepare for these regulatory requirements, organizations should document a clear and structured process for “go/no-go” decisions when deploying new AI models and use cases. Sharing an outline of this process with the public provides stakeholders with visibility into the protective measures implemented to mitigate risks and reduce harm. It also enables informed discussions on potential additional steps to be taken.
These attributes are not meant to be comprehensive, or even sufficient disclosures for upcoming regulations. But they are a useful starting set of things to consider for organizations seeking to balance trust, building disclosure against protecting their competitive moats.
Read the full article here