The field of generative AI has made remarkable progress in recent years. This is driven by breakthrough innovations such as Generative Adversarial Networks (GANs) in 2014, Transformer Networks in 2017, and Reinforcement Learning with Human Feedback (RLHF) in 2017. As these algorithms have advanced, continued improvements in hardware acceleration have unlocked unprecedented capabilities to enable the training of large-scale language models (LLMs) on huge natural language datasets. . But the real turning point came with the release of OpenAI’s game-changing ChatGPT. This important announcement, and the subsequent announcement of GPT3.5 and DALL-E 2 models on Microsoft Azure, Google Bard, GPT4, and Microsoft CoPilot, ignited a wildfire of rapidly intensifying interest in the generative AI landscape. formed exponentially. .
Today, companies are actively exploring the potential of generative AI, identifying use cases, conducting proof-of-value exercises, conducting pilots, and integrating the technology into their operational processes.
There are four major areas where generative AI is applied within the enterprise.
- Enterprise Virtual Assistant and Knowledge Search: This includes use cases such as enterprise document question-and-answer, intelligent search, and summarization. The aim is to improve customer experience, increase employee productivity and accelerate research and development.
- Content generation for digital commerce and media: Generative AI develops digital marketing accelerators for adaptive storytelling, generates dynamic digital avatars with natural voice interactions, generates scripts and images for advertising, and creates conversational interfaces for digital commerce. , and to perform content generation across various media channels.
- Accelerate the software development lifecycle: Generative AI is considered an innovative technology that accelerates the software development process by providing code generation assistance, transformation of legacy code for migration purposes, documentation and test case generation, and software incident management assistance. I’m here.
- Generating synthetic data: Many analytics firms predict that the majority of digital data will be generated by AI. Most of that data is used to train AI models. This entails using Generating AI to create synthetic data when real data is not available or accessible due to regulatory constraints, or when real data lacks the required variation. . Generative AI facilitates data anonymization to address privacy and security requirements, and also enables simulation of environments such as digital twins and metaverses.
Some of the biggest challenges of generative AI models such as GPT and PaLM are data privacy and protection. There are many other concerns about generative AI models that violate common norms. The main concerns companies have are:
Copyright infringement: Some models are trained on large amounts of internet data that contain code and utilize this data or code to generate responses that are identical to the original data. This may violate the copyright of the original.
Data privacy: Internet data can contain personal information, which can remain in the model’s response if not masked. We still have a long way to go before all personal information is filtered. Additionally, other aspects such as private business websites, browsing history, and chat history may also be used to train generative AI models. Interestingly, this information is difficult to erase or forget once trained.
Ethical or Responsible Use of AI: Lack of explainability, use of inappropriate or dangerous language, response bias, and wrong responses with high confidence are areas of concern for many companies.
There are solutions to the above concerns, some of which are easy to implement, while others take longer. Given the potential benefits of generative AI, many companies are beginning to build solutions.
We recommend the following principles that provide an approach to addressing these risks and concerns from an application architecture perspective.
Information security
Restricting direct employee access to interfaces such as ChatGPT is critical to protecting sensitive information. Instead, organizations should provide access to dedicated business applications that utilize generative AI models within a robust information security framework. This can be achieved by leveraging enterprise-grade frameworks provided by cloud service providers or by implementing an in-house solution within a secure enterprise network. An information security framework should cover all aspects of data exchange with generative AI models, including prompts, custom training and fine-tuning data, trained model instances, and generated responses. Encryption, access controls, and data retention controls should be implemented to ensure the highest level of protection for these data elements.
purpose limit
General-purpose generative AI models, such as GPT and Bard, are trained on vast information corpora spanning nearly every area of human knowledge. This makes it possible to elicit responses from these models that are irrelevant to business purposes and potentially harmful, whether intentional or unintentional. Exposure of customers and employees to such actions may adversely affect business performance, credibility and goodwill, and may even result in legal liability.
To mitigate these risks, companies must develop specialized, individual applications built on generative AI models and limited to specific business purposes. Each request to your application should be programmatically evaluated to see if it is relevant to its intended purpose before providing a response. This filtering process can be achieved using a custom trained classification model.
Implementing this filtering mechanism ensures that the application is only used for its intended purpose and establishes the necessary guardrails to prevent unintended use.
Custom training and fine-tuning
Many generative AI use cases in the enterprise can leverage knowledge available within the enterprise itself. This requires custom he training or fine-tuning, and provides a higher level of control over the data used to train the model, providing greater transparency. Additionally, by saving document embeddings, you can create a transparency framework on top of the enterprise document corpus used for custom training. This framework makes it possible to track the responses to the source documents used for training, providing transparency and explainability to the system’s responses. The result is increased reliability. Custom training and fine-tuning also help achieve fairness by allowing you to control the data used to train your model.
Response management
Response moderation is necessary to detect and remove harmful elements in generated responses, such as irrelevant, inappropriate, plagiarized, or copyrighted content. Use cases that rely heavily on pre-trained knowledge of the model, which comes from training done outside the enterprise, are more likely to result in harmful content. Therefore, the implementation and rigor of response moderation should vary depending on the specific use case. Tailoring responses programmatically can be accomplished using a combination of available services that detect various types of harmful content and an ensemble of custom-trained classification models.
Conclusion
These broad principles mitigate many of the risks associated with generative AI in enterprise environments, but their applicability depends on the use case. For example, virtual assistants used by employees are highly adaptable to custom his training, and thus relatively low risk when implemented in parallel with other principles. Marketing tagline generators and code generators, on the other hand, rely heavily on the pre-trained knowledge of the model, which makes them more prone to generating plagiarized or copyrighted content.
Given the applicability of these principles, companies should conduct a risk assessment for each individual use case and determine a roadmap for adapting generative AI based on each use case’s specific risk profile. I have.
