How to advance generative AI exploration safely and reliably?

Nowadays, the exploration and implementation of generative AI are inevitably intertwined with security issues. According to a recent data report, 49% of business leaders believe that security risks have become a primary issue, and 38% of business leaders list human errors/human data leaks caused by not understanding how to use GPT tools as the top challenge. .

While these concerns are valid, the benefits to early adopters will far outweigh these potential consequences of hindering exploration

In this article, we will help team members and customers Understand why security should not be treated as an afterthought, but as a prerequisite for integrating AI with business, and explore a series of best practices for working in this area

Start with security strategy

Enterprises have become aware of the emerging security risks and urgency brought by AI applications. According to the statistical report mentioned earlier, 81% of business leaders said that their companies have or are developing user policies built around generative AI. However, due to the rapid development of this technology and the emerging applications, With use cases emerging all the time, policy content must also be constantly updated to address risks and challenges that arise at any time.

In order to minimize security risks while accelerating exploration, it is naturally necessary to set up "guardrails" for testing and learning efforts. In addition, when formulating relevant policies, we should never proceed in isolation. Instead, we should fully solicit the opinions of representatives from various departments within the enterprise and consider how different functional units use/whether they can use generative artificial intelligence to deal with the security risks they face

In short, the exploration of AI technology by various departments should not be roughly prohibited. If you impose an enterprise-wide ban out of sheer fear, you don't have to worry about competitors eating up your market share - you're destroying the Great Wall yourself.

Focus on frontline personnel

To ensure that generative AI can be used safely, we should first provide permissions for ordinary developers. For example, they could be given full access to a private instance of Insight GPT, a large language learning model. On the one hand, this can help discover potential use cases, and at the same time, it can also perform stress testing based on the output to continuously improve the generative AI service

We soon discovered that a member of the warehouse team found a way to improve the efficiency of order delivery. Methods. In this particular case, the member asked for a script to be written in SAP to automate a portion of the workload. Although effective, this attempt can easily lead to accidents if protection is not set up correctly. For example, if a staff member accidentally executes a transaction that does not exist in an order, subsequent automated steps will not be aborted

In the process of promoting civil exploration and limiting risks as much as possible, we need to take the following measures: Review The committee should develop clear guidelines, conduct risk assessments, and increase transparency around AI systems. At the same time, appropriate training should be carried out to educate employees on how to apply AI to work scenarios in a responsible manner, especially clear ways to deal with key issues such as ethics, bias, human supervision and data privacy. In addition, internal forums should be set up to encourage team members to share their findings and lessons learned within the company’s innovator group

Reduce the risk of “hallucinations”

Why generative AI is risky , an important reason is that it occasionally produces "hallucinations". According to the Insight report, a common theme that business leaders are most concerned about is how hallucinations can lead to poor business decisions. However, the risks caused by hallucinations often vary from minor to serious, and the specific effects often vary from scenario to scenario.

Although the GPT tool will inevitably output some results that are inconsistent with objective reality, we soon realized that such wrong answers often belong to confusion at the wording level. For example, in early testing, we asked Insight GPT which song Eddie Van Halen had collaborated with. The correct answer is "Beat It," but its answer is "Thriller." But from another perspective, "Beat It" is indeed a piece on the "Thriller" album, so its answer is not unreasonable.

Doing this ensures that we can effectively manage and supervise AI-generated content to reduce the risk of hallucination issues. At the same time, we also need to strengthen the training and monitoring of AI systems to ensure that the content they generate complies with relevant policies and standards. Only through such measures can we better cope with problems that may arise when dealing with highly subjective workloads

At present, the generative AI industry is still in its infancy. Whoever can figure out responsible and safe application methods while reducing potential threats caused by data leakage, information errors, bias and other risks will be able to establish a clear technical advantages. Enterprises need to ensure that their AI policies continue to keep pace with changes in the industry, and gradually build user trust while maintaining compliance and alleviating hallucination problems.

The above is the detailed content of How to advance generative AI exploration safely and reliably?. For more information, please follow other related articles on the PHP Chinese website!