Quick Answer: What Is ChatGPT?

More Detail

ChatGPT is the first widely known AI technology that challenges the one trait humans always thought they would have over machines: creativity. Less than five days after its release, ChatGPT had attracted a million users who were impressed with its ability to generate human sounding answers to questions. The immediate effects of ChatGPT’s release include:

• Dismay by writers, marketers and programmers, who feel their jobs are threatened.

• Concern by educators over the possibility that ChatGPT will enable students to cheat; some have banned it or are reevaluating approaches to student assessment.

• Google’s planned integration of chatbot features into its search engine, though ChatGPT is not a replacement for Google search as it sometimes returns incorrect results.

• Microsoft’s announcement of a multiyear, multibillion-dollar investment in OpenAI to build out capabilities on Azure and extend the reach of OpenAI’s models.

• The building of unique capabilities on top of ChatGPT by scores of startups; these capabilities are expected to ramp up significantly, creating new opportunities for value creation.

Rarely does an emerging technology get this much attention this quickly. ChatGPT has been used as an aid in all kinds of writing (marketing, contracts, technical, creative, computer code) and as an educational resource. The risks and benefits of ChatGPT are outlined in Figure 1 below.

ChatGPT is getting significant attention for its ability to handle broad general questions. It is different from most chatbots because it is an open-context dialogue, while most other chatbots are goal-oriented and execute a procedural response. ChatGPT is being used to enhance, complement or replace specific natural language functions within chatbots. These include:

• Classification/natural language understanding (NLU) intent identification. ChatGPT can work as a classifier that also provides confidence scores. This can then be used to complement existing NLUs to help clarify intent.

• Natural language generation. ChatGPT can be used to generate constrained responses. The constraint enables assurance that the answer is specific to the material.

• Summarization of conversations. Because there may be errors, these summaries must be reviewed by a person. But in many use cases, reviewing content is much faster and less tiring than asking the personto generate the summary.

• Sentiment extraction. While current sentiment analysis technologies work well at the word level, they often fail at identifying the overall sentiment of the conversation. ChatGPT is often better at that. As with other uses above, this can act as a complement to existing approaches.

The text generated by ChatGPT depends on the prompt, task or domain, and on the quality and quantity of the training data leading to model risks and misuse risk.

Model risks include:

• “Hallucinations” — These wrong answers are ChatGPT’s most widespread problem, and they are hard to spot. When generated answers are critical for the business or reputation, ensure human supervision and feedback.

• Subpar training data — Data could be insufficient, obsolete or contain sensitive information and biases, leading to biased, prohibited or incorrect responses. Using ChatGPT for the most critical tasks and less popular domains poses higher risks, including ethical, reputational and legal risks. ChatGPT outputs must be reviewed by a subject matter expert before publication.

• Copyright violations — Recently, OpenAI was accused of using copyrighted data for training its models. Consult with your legal department about your own liability when exploiting ChatGPT.

Misuse risks include:

• Deepfakes — These outputs generated by ChatGPT could appear realistic, but actually be fake content. Organizations must be vigilant to identify fake news, misinformation, impersonations or efforts to manipulate public opinion.

• Fraud and abuse — Bad actors are already exploiting ChatGPT by writing fake reviews, spamming and phishing. Like any fraud, the use of ChatGPT for malicious purposes will be ongoing. Ensure that appropriate specialists are informed of the ChatGPT risks and are involved in defensive activities.

ChatGPT adds a chatbot interface to a model that was fine-tuned from OpenAI’s GPT-3 series of foundation models. OpenAI made this model suitable for conversations by using reinforcement learning from human feedback (RLHF), in which initial training was supplemented with supervised fine-tuning by human trainers for conversational responses.

Foundation modelsare versatile AI models used in generative AI, and are designed to replace task-specific models. Foundation models are trained on a broad set of unlabeled data that can be used for different tasks with additional fine-tuning. They are called foundation models because of their critical importance and applicability to a wide variety of downstream use cases, due to large-scale pretraining of the models. Foundation models are based mainly on transformer architectures, which embody a type of deep neural network architecture that computes a numerical representation of text in the context of surrounding words, emphasizing sequences of words.

ChatGPT is fine-tuned from OpenAI’s GPT-3 series of foundation models for the specific purpose of conversational applications. Large language text generators have been called “stochastic parrots” for generating sequences of words that are based on previous observations of frequent word combinations, but lack meaning. They have also been criticized for being environmentally costly, and for exacerbating dominant viewpoints and encoding biases.

ChatGPT may have surprised some people, but foundation models have been steadily improving since 2017. Google and Meta platforms claim to have models that are just as capable as ChatGPT, but haven’t released them due to reputational risk.Other notable foundation models designed for text-to-image generation include DALL·E 2, Stable Diffusion and Midjourney. Microsoft is researching text-to-voice generation with VALL-E, claiming to be able to synthesize highly personalized speech from a three-second sample. Expect more advances in this field.

ChatGPT is an example of a broader trend called generative AI. Generative AI refers to AI techniques that learn a representation of artifacts from data and use it to generate unique artifacts that preserve a likeness to original data. Generative AI enables computers to generate brand-new, completely original variations of content (including images, video, music, speech and text). It can also improve or alter existing content, create new data elements, and create novel models of real-world objects such as buildings, parts, drugs and materials.

Recommended by the Authors

Evidence

Google Calls In Help From Larry Page and Sergey Brin for AI Fight, New York Times.

Exploring the Legal Minefield of ChatGPT and Intellectual Property Rights, Falcon Rappaport & Berkman LLP.

On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?, Conference on Fairness, Accountability and Transparency, March 2021.

Google Won’t Launch ChatGPT Rival Because of “Reputational Risk”, The Verge.

ChatGPT Is “Not Particularly Innovative” and “Nothing Revolutionary,” Says Meta’s Chief AI Scientist, ZDNet.