journalfront

OpenAI Funds Research on Predicting Human Moral Judgements in AI

2024-11-22

OpenAI, through its nonprofit org, is making significant strides in academic research. In a filing with the IRS, it was disclosed that a grant was awarded to Duke University researchers for the "Research AI Morality" project. This move aims to train algorithms to predict human moral judgements in various scenarios involving conflicts among morally relevant features in different fields.

Unraveling the Complexity of Predicting Moral Judgements with OpenAI's Funding

OpenAI's Grant to Duke University

In a filing with the IRS, OpenAI Inc., its nonprofit arm, revealed that a grant was given to Duke University researchers for the "Research AI Morality" project. This initiative is part of a larger, three-year, $1 million grant aimed at studying "making moral AI." While little is publicly known about this "morality" research other than its 2025 end date, it holds great potential. The principal investigator, Walter Sinnott-Armstrong, a practical ethics professor at Duke, couldn't provide details when contacted. However, he and his co-investigator, Jana Borg, have conducted several studies and even written a book about AI's potential as a "moral GPS" to assist humans in making better judgements. They have created a "morally-aligned" algorithm to help determine kidney donation recipients and studied scenarios where people prefer AI to make moral decisions.

The Goal of the OpenAI-Funded Work

According to the press release, the objective of this OpenAI-funded research is to train algorithms to predict human moral judgements in complex scenarios. These scenarios involve conflicts among morally relevant features in medicine, law, and business. It is a challenging task as morality is a nuanced concept that may not be easily achievable with current technology. In 2021, the nonprofit Allen Institute for AI built a tool called Ask Delphi, which aimed to provide ethically sound recommendations. It performed well in judging basic moral dilemmas, such as knowing that cheating on an exam is wrong. However, slight rephrasing of questions could lead Delphi to approve of almost anything, including the unthinkable act of smothering infants.

The Limitations of Modern AI Systems

Modern AI systems, such as machine learning models, are statistical machines. Trained on a vast amount of data from the web, they learn patterns to make predictions. But they lack an appreciation for ethical concepts and an understanding of the reasoning and emotions involved in moral decision-making. This is why AI often parrots the values of Western, educated, and industrialized nations, as the web and its training data are dominated by articles endorsing these viewpoints. Many people's values are not expressed in AI's answers, especially if they don't contribute to the training sets by posting online. AI also internalizes biases beyond a Western bent, as seen in Delphi's view that being straight is more "morally acceptable" than being gay.

The Challenge of Morality's Subjectivity

The challenge before OpenAI and the researchers it is backing is exacerbated by the inherent subjectivity of morality. Philosophers have been debating ethical theories for thousands of years, and there is no universally applicable framework yet. Claude leans towards Kantianism, focusing on absolute moral rules, while ChatGPT slightly favors utilitarianism, prioritizing the greatest good for the greatest number. Determining which approach is superior depends on individual perspectives. An algorithm to predict human moral judgements must consider all these factors. It is a very high bar to clear, and it remains to be seen if such an algorithm is even possible.TechCrunch has an AI-focused newsletter! Sign up here to get it in your inbox every Wednesday.

Softbank-backed Meesho has made a significant move in the e-commerce space by introducing what it claims to be the first GenAI-powered voice bot among Indian e-commerce firms. This innovative solution aims to enhance customer support while also cutting down expenses by a substantial 75%. GenAI, or generative AI, involves training models on extensive data to generate and parse content, specifically in the form of the human voice.

Revolutionize Customer Support with Meesho's GenAI Voice Bot

Customer Support on a New Level

The Bengaluru-based e-commerce startup, Meesho, which has over 160 million customers in India, with 80% in smaller cities and towns, has been focusing heavily on customer support. Currently, its AI bot handles an impressive 60,000 customer calls daily in English and Hindi. With plans to add support for six more Indian languages, Meesho is expanding its reach. Sanjeev Barnwal, co-founder and chief technology officer, emphasizes the importance of getting customer support right to create the best experience for users. 1: Meesho has chosen to combine existing large language models (LLMs) with custom-built components that understand local context and language nuances. This approach allows them to provide efficient customer support without building their own LLM. As Barnwal explains, the off-the-shelf LLMs available are already performing well in Hindi and English. 2: In a demo, it became evident that the system had to overcome several technical hurdles. Voice quality is crucial as many users have low-end smartphones and are in noisy environments like on busy streets. The bot has been engineered to reduce latency and filter out street noise while maintaining natural-sounding conversations.

Cost Savings and Query Resolution

Meesho has achieved significant cost savings through the AI bot. By reducing per-call expenses by 75%, the company is able to allocate resources more efficiently. Additionally, the bot has a remarkable 95% query resolution rate, with only 5% of calls requiring human intervention. This not only saves time but also improves customer satisfaction by 10%. 1: The e-commerce firm, which recently generated positive cash flow, is leveraging this technology to offer round-the-clock customer support. This ensures that customers can get assistance at any time, enhancing their overall shopping experience. 2: Meesho also mentioned that the voice bot has reduced the average customer call handling time by half. This allows the company to handle a larger volume of calls and provide faster responses to customers.

Human Agents and Policy Guidelines

While the AI bot is highly efficient, Meesho is clear that it is not meant to replace human agents. Instead, human agents are redirected to handle more complex queries and offer seller support. Another challenge is ensuring that the AI sticks to strict guidelines about policies like returns and refunds. Meesho is working on maintaining the balance between automation and human interaction to provide the best service. 1: This approach recognizes the value of human expertise in handling certain situations and ensures that customers receive personalized assistance when needed. 2: By combining the capabilities of AI and human agents, Meesho is able to offer a comprehensive customer support system that meets the diverse needs of its customers.The rollout of Meesho's GenAI-powered voice bot highlights India's tech companies' race to deploy AI for increased efficiency. As Hemant Mohapatra from Lightspeed mentioned, building foundational models requires significant talent, and companies are focusing on playing the wars they are capable of winning. Meesho's success with this initiative sets an example for others in the industry.

Now, with the help of a remarkable months-old startup called Pickle, you have the incredible ability to seemingly be on a Zoom call in your office while actually sipping a margarita in a hammock that's far away. The concept is straightforward - submit a five-minute training video of yourself to create an avatar, and within 24 hours, you're good to go. Whether you want to make a call from the comfort of your car or choose to stay in bed when laziness strikes, Pickle has got you covered. However, it's important to note that using the service at a beach club might present some challenges, as evident from the demo video. Currently, the service is available in basic, standard, and pro versions, with prices ranging from $300 to $1,150 per year.

Unlock the Freedom of Virtual Zoom Presence

MacOS and Windows Compatibility

Pickle, backed by the L.A. venture firm Krew Capital, currently only works with macOS. But here's the exciting news - a Windows version is set to be released next month. This means that more users will be able to enjoy the convenience of creating Zoom call avatars on their preferred operating system. While the current focus is on macOS, the team at Pickle is working hard to bring the same seamless experience to Windows users.

With the wide range of conferencing apps available, such as Zoom, Google Meet, and Teams, Pickle offers users the flexibility to choose the platform that suits them best. However, it's important to note that the use of these apps with Pickle's service is currently on hold due to high demand. Clone generation is experiencing delays, but the team is dedicated to meeting the needs of their customers and getting the service up and running smoothly as soon as possible.

Versatile Use Cases

The applications of Pickle's Zoom call avatar technology are truly diverse. Imagine being able to attend important business meetings from the comfort of your home while still maintaining a professional appearance. Or perhaps you want to join a family gathering from a remote location without missing out on the fun. Pickle's avatar technology allows you to do just that. It gives you the freedom to be anywhere and still be present in virtual meetings.

Whether you're a busy professional on the go or someone who simply values the convenience of remote communication, Pickle's service offers a unique solution. It eliminates the need to be tied to a specific location and allows you to be flexible with your work and personal life. With Pickle, you can truly have it all.