In the short time since artificial intelligence has become mainstream, its ability to do the previously unimaginable is already evident. But along with this stunning potential comes the possibility that AIs will be unpredictable, aggressive, and even dangerous. This possibility prompted Google CEO Sundar Pichai tell Employees believe that responsible development of artificial intelligence was the company’s top priority in 2024. We have already seen tech giants like Meta, Apple and Microsoft. Subscribe the US government’s efforts to promote responsible artificial intelligence practices. UK also investment in creating tools to regulate AI, like many others, from European Union To World Health Organization and beyond.
Increased attention to AI’s unique ability to behave in unexpected ways is already influencing how AI products are perceived, marketed and accepted. Firms no longer market their products solely using traditional business success criteria such as speed, scalability and accuracy. They increasingly talk about their products in terms of their behaviorwhich ultimately reflects them values. Today, the strength of products ranging from self-driving cars to smart home appliances is how well they embody certain values such as safety, dignity, fairness, harmlessness and usefulness.
In fact, as AI becomes embedded in more aspects of everyday life, the values on which its decisions and behavior are based become critical product characteristics. As a result, ensuring that AI results at all stages of use reflect specific values is not a cosmetic concern for companies: the alignment of values that drives the behavior of AI products will significantly impact market acceptance, ultimately market share, and ultimately company survival. Instilling the right values and demonstrating the right behavior will increasingly become a source of differentiation and competitive advantage.
But how do companies update their AI efforts to ensure their products and services behave as their creators intended? To help address this challenge, we have divided the most important transformation challenges into four categories, based on our recent work in Harvard Business Review. We also provide an overview of frameworks, practices, and tools that leaders can use to answer the question: How do we properly define the values of AI?
1. Define your values, write them into the program and make sure that your partners share them too.
The first task is to determine whose values must be taken into account. Given the scale of AI’s potential impact on society, companies will have to consider a more diverse group of stakeholders than usual. This extends beyond employees and customers to include civil society organizations, politicians, activists, industry associations and others. The preferences of each of these stakeholders need to be understood and balanced.
One approach is to introduce principles based on established moral theories or concepts developed by credible global institutions such as UNESCO. For example, the principles of Claude’s Anthropic model are taken from the UN Universal Declaration of Human Rights. BMWMeanwhile, AI values are based on EU requirements for trustworthy AI.
Another approach is to formulate your own values from scratch, often by creating a team of specialists (technologists, ethicists and human rights experts). For example, artificial intelligence research laboratory DeepMind prompted feedback based on philosopher John Rawls’s idea of the “veil of ignorance,” in which people propose rules for a community without knowing how the rules will affect them individually. What’s striking about DeepMind’s findings is that they focus on how AI can help those most disadvantaged by making it easier to get user support.
Determining the right values is a dynamic and complex process that must also accommodate changing regulations across jurisdictions. But once these values are clearly defined, companies will also need to build them into the program to explicitly constrain AI behavior. Companies like Nvidia and OpenAI are developing mechanisms to incorporate formal generative AI constraints into their programs to ensure they don’t cross red lines by making incorrect queries or producing inappropriate content. OpenAI actually distinguished its GPT-4 model due to its improved performance, touting it as 82% less likely than previous model. to respond to inappropriate requests such as hate speech or malware code.
It is important to note that alignment with values requires the further step of attracting partners. This is especially important (and difficult) for products built using third-party models, due to restrictions on how much companies can customize them. Only the developers of the original models know what data was used to train them. Before entering into new partnerships, AI developers may need to establish processes to identify the value of external AI models and data, similar to how companies evaluate the sustainability of potential partners. As underlying models evolve, companies may have to change the models they rely on, further strengthening value-based AI due diligence as a source of competitive advantage.
2. Evaluate the trade-offs
Companies are increasingly trying to balance often competing values.. For example, companies offering products to help older people or educate children must consider not only safety, but also dignity and agency. When should AI not help older users to build their confidence and respect for their dignity? When should it help ensure a child has a positive learning experience?
One approach to this balancing act is to segment the market according to values. A company like DuckDuckGo does this by focusing on a small search market that cares more about privacy than algorithmic accuracy, allowing the company to position itself as a differentiated option for Internet users.
Managers will have to make detailed decisions about whether certain content generated or recommended by AI is harmful. To guide these decisions, organizations need to establish clear processes and communication channels with stakeholders early on to ensure ongoing feedback, alignment and learning. One way to handle such efforts is to create an AI overseer within the company with real independence and authority.
3. Provide feedback to people
Maintaining the values of an AI product, including removing biases, requires extensive feedback from people about AI behavior, data that will need to be managed through new processes. The AI research community has developed various tools to ensure that trained models accurately reflect human preferences in their responses. One of the fundamental approaches used by GPT-3 involves “supervised fine-tuning” (SFT), in which models are given carefully selected answers to key questions. Building on this, more sophisticated techniques such as Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO) have made it possible to fine-tune AI behavior in a more iterative feedback loop based on human evaluations of the model’s output . .
What all of these fine-tuning methodologies have in common is the need for real feedback from people to “nudge” the models to become more aligned with their respective values. But who provides feedback and how? In the early stages, engineers can provide feedback while testing AI results. Another practice is to create “red teams,” which act as adversaries and test the AI by pushing it toward undesirable behavior to find out why it might fail. These are often internal teams, but external communities can also be used.
In some cases, companies may ask users or consumers themselves to provide valuable feedback. For example, social media and online gaming companies have implemented content moderation and quality management processes, as well as escalation protocols based on user reports of suspicious activity. Reports are then reviewed by moderators, who follow detailed instructions when deciding whether to remove content.
4. Prepare for surprises
As AI systems become larger and more powerful, they may also exhibit more unexpected behavior. This behavior will increase as AI models perform tasks for which they were not explicitly programmed, and endless versions of the AI product will be created depending on how each user interacts with it. The challenge for companies will be to ensure that all these versions are consistent.
AI itself can help reduce this risk. Some companies are already using one AI model to challenge another through adversarial learning. More recently, out-of-distribution detection (OOD) tools have been used to help AI deal with things it has not encountered before. robot playing chess who grabbed a child’s hand because he mistook it for a chess piece is a classic example of what can happen. OOD tools help AI “know what it doesn’t know” and refrain from acting in situations it is not trained to handle.
While the problem cannot be completely eradicated, the risk associated with unpredictable behavior can be actively managed. The pharmaceutical sector faces a similar challenge when patients and doctors report side effects not detected in clinical trials, often leading to approved drugs being removed from the market. When it comes to artificial intelligence products, companies must do the same to identify unexpected behavior after release. Companies may need to create specific AI incident databases, like those developed by the OECD and the Partnership on AI, to document how their AI products are developing.
Conclusion
As AI becomes more ubiquitous, companies’ values (how to define, design and protect them) become increasingly important as they ultimately determine the behavior of AI products. For executives, navigating a fast-changing, value-based marketplace where unpredictable AI behavior can determine adoption and acceptance of their products can be challenging. But addressing these issues now and delivering reliable products that align with your values will lay the foundation for creating a long-term competitive advantage.
***
Read others Luck columns by Francois Candelon.
Francois Candelon is a managing director and senior partner at Boston Consulting Group and global director of the BCG Henderson Institute (BHI).
Jacob Abernethy is an assistant professor at Georgia Tech and co-founder of water testing company BlueConduit.
Theodoros Eugeniou is a professor at INSEAD, an advisor to the BCG Henderson Institute, a member of the OECD Network of AI Experts, a former World Economic Forum AI Partner, and co-founder and chief innovation officer of Tremau.
Abhishek Gupta is the Director of Responsible AI at the Boston Consulting Group, a fellow at the BCG Henderson Institute, and the founder and principal investigator of the Montreal Institute for AI Ethics.
Yves Lostanlin has held senior management positions and advised CEOs of numerous companies, including AI Redefine and Element AI.
Some of the companies featured in this column are former or current BCG clients.