The emergence of ChatGPT has generated excitement about the potential for AI assistants to lighten our workload, as well as unease about the possible displacement of knowledge workers. Businesses are scrambling to understand how to use this technology for competitive advantage and the impact it will have on the workforce of the future. But what exactly is this game-changing technology, and what can we realistically expect it to do?
ChatGPT was released to the public in November as a free “research preview” by the OpenAI research lab. This impressive technology attracted over a million users in its first week, and surpassed an incredible 100 million users in January, setting a record for the fastest-growing app of all time.
This enormous user base is allowing OpenAI to gather massive amounts of user feedback to improve its model. Meanwhile, Microsoft invested an additional $10 billion in OpenAI, entering a partnership that gave it the exclusive right to develop and sell an enterprise-ready version of the service. The commercial offering will include security features such as private networks, authentication and authorization, as well as adherence to responsible AI standards.
ChatGPT is powered by a Large Language Model (LLM), an AI model that can generate and process natural language in a variety of ways. For example, LLMs can summarize text, answer questions from text, and rephrase text. There are also specialized LLMs for domain-specific needs. There are LLMs trained on foreign languages (for translation), scientific and medical research, and even programming languages like Python. Advanced LLMs by Google, Deepmind, NVIDIA, Meta, Baidu, Cohere, Huawei, and many others are available now—or racing to market.
ChatGPT was built on the powerful GPT-3.5 Large Language Model, which was trained on a massive corpus of text including 300 billion words from books, research articles, and websites like Wikipedia—approximately 570 GB of data. ChatGPT’s training also included conversational text from sources like Reddit, allowing it to specialize in generating fluent, human-sounding responses to people’s questions.
It is particularly advanced thanks to its ability to remember context from prior queries, generate highly coherent conversational text, and “understand” high-level instructions. Its development included Reinforcement Learning from Human Feedback (RLHF) to optimize how it aligns its responses to human expectations.
Although many are hailing ChatGPT as a “Google killer,” it is important to understand that ChatGPT is not a search engine—although parallel offerings like the new AI-powered Bing and ChatGPT plugins are blending ChatGPT and web search functionality. Natively, ChatGPT cannot search the web for answers or return accurate links or sources for its results. Rather, it is essentially an incredibly advanced autocomplete tool. LLMs predict the most statistically probable next word in a sentence based on the words that came before. ChatGPT is exceptionally skilled at this, able to generate long strings of fluent, sensible text that provide relevant responses to users based on its pre-existing knowledge base.
At times, however, those responses will not be accurate, nor can they provide a linked source for validation. Rather than returning no results, like a search engine might, ChatGPT will sometimes “predict” a reasonable-sounding—but incorrect—response. For example, if you ask ChatGPT for an executive’s biography, it will often correctly identify the person and when he or she joined their current firm, but it may quickly begin to “hallucinate,” or make up, plausible-sounding falsehoods about the person’s prior roles and educational background.
Students using ChatGPT as a homework shortcut may learn the hard way that it will also make up citations to non-existent studies, sources, and authors, complete with invalid URL links. And since its training data ends in late 2021, so does its knowledge of world events.
There is debate as to whether these inaccuracies can be fully resolved or are inherent to generative Large Language Models. However, combining LLMs with traditional software engineering techniques should largely mitigate this deficiency.
A skilled user of ChatGPT, familiar with its strengths and limitations, can coax the best answers from it through a new discipline called prompt engineering. By optimizing query text, he or she can get vastly improved responses. For example, ChatGPT cannot currently use a URL to answer a question from the text on that webpage—although it will pretend to, just by looking at the words in the URL address and guessing what the page is about, returning a substandard response. If you provide a shortened “tinyURL” of the same web address, ChatGPT will admit that it cannot fetch content from the Internet.
But this limitation can be circumvented by pasting the full text of the web page (up to around 500 words) in the query. When it has the full text, ChatGPT is excellent at summarizing, extracting key themes, or answering questions about that text. A user skilled at prompt engineering also knows how to get ChatGPT to modify its responses for more relevant output. For example, “Explain blockchain to a non-technologist,” “Make this email less stilted,” or “Only return a result if you’re certain you have the right answer; otherwise say you don’t know.”
To avoid being left behind, businesses are rushing to understand how to harness the power of this impressive technology. What should they be thinking about for their products, internal processes, and future workforce?
Support Bots. One obvious use case to explore is conversational support bots. Most businesses will want a private instance of an LLM specifically trained with their proprietary support documentation. Because of the potential for inaccurate responses, companies should first think about providing the support chatbot to their internal support agents. They can validate answers, correct any errors, and provide the final results to clients. This will not only save the agents valuable time, but their edits can serve as feedback to improve the bot. Eventually, the more reliable chatbot can be used by clients directly.
Natural Language Interfaces. Similarly, ChatGPT and other LLMs can be used to help interpret users’ natural language queries to make software applications more intuitive. A firm’s machine learning team can use LLMs to generate synthetic training data to improve their internal natural language understanding models. Dedicated LLMs could power employee knowledge bases, trained with internal corporate data to answer employees’ questions about corporate policy or procedures, accelerate new-hire onboarding, and reduce knowledge silos.
Document Processing. To save employees’ time, financial firms can pass filings, research, transcripts, and news to ChatGPT to summarize and extract key takeaways. New incoming documents can also be automatically classified to determine which staff should take a closer look. Foreign language documents can be automatically translated. Employees can provide a table of financial data to ChatGPT and have it draft narrative prose.
Coding. Developers at your firm may already be using ChatGPT or another AI coding assistant to help debug software, write code comments and documentation, understand what a section of code is trying to accomplish, or even write new sections of code. As with any use of a LLM, a human expert must review the output for accuracy. Coding LLMs have proven to be great time-savers for many engineers, but can also produce non-working, buggy, and even insecure code that can pose a genuine threat to your organization if not properly validated.
Security and Risk. Security departments will be on the lookout for heightened cyberattacks resulting from more believable phishing emails that LLMs generate, as well as security breaches from improperly tested AI-generated internal code and external AI-generated malware. Legal and compliance departments may need to mitigate risks relating to privacy regulations and intellectual property rights—or even defamation with rising AI-generated disinformation.
Process Guide. ChatGPT can provide valuable guidance for questions like “How do I set up a working group?” or “Can you create an onboarding and training workshop for new employees?”. Users can ask it to provide more detail to flesh out any of its recommended steps, or to brainstorm alternative methods.
Writing Assistant. ChatGPT is excellent at generating a wide variety of written content. Marketing professionals can use it as a writing assistant to draft creative copy, social media posts, blogs, and newsletters. Corporate communicators can use it to draft announcements of events or to introduce new policies. HR can even use it to generate drafts of the policies themselves (e.g., “Write a hybrid work policy”), along with everything from job descriptions to employee manuals. Employees can pass in their own drafts and use ChatGPT as an editor to modify the tone, expand upon key points, or tighten up certain sections. Knowledge workers in nearly every area of your firm can use ChatGPT as a brainstorming partner or writing assistant and reallocate the time savings to more specialized work.
Although the hype around ChatGPT overlooks some of its current limitations, advancements are rapidly being developed and released. Information professionals of the near future will certainly leverage LLMs on a daily basis—much as formerly advanced technology like spell check and autocorrect have become ubiquitous workplace aids.
LLMs are already being incorporated into enterprise technology to summarize meetings, suggest email responses, provide real-time language translation for call centers, and monitor for regulatory risk. ChatGPT and other LLMs will undoubtedly alter the landscape of the corporate world, making knowledge workers more productive in ways that were unimaginable only a few years ago.
To learn more, visit FactSet Artificial Intelligence and read our additional LLM articles:
This blog post is for informational purposes only. The information contained in this blog post is not legal, tax, or investment advice. FactSet does not endorse or recommend any investments and assumes no liability for any consequence relating directly or indirectly to any action or inaction taken based on the information contained in this article.