The ethical challenges of generative AI 


The ethical challenges of generative AI 

In late 2022, then director of artificial intelligence at Tesla, Andrej Karpathy, shared on Twitter that 80% of his code was written by AI, adding: “I don’t even really code, I prompt and edit”.

Artificial intelligence that can generate data (generative AI) promises to revolutionize the world of work, and yet this new technology comes with ethical risks. Applications like ChatGPT are essentially predictive models, anticipating the word that might most reasonably follow in a sequence. But they lack any connection with the real-world context of those words – so the results could be brilliant insights, fluent nonsense or even wide-scale misinformation.  Moreover, a surge in use of generative AI might have severe impacts on energy use and worsen climate change.

ChatGPT—recently made publicly available by its creator OpenAI, and a standout success in this arena—has already been used in drafting legal cases and judgments, posing a challenge to the dominance of traditional search engines, creating prompts for movies, and helping software developers to compile code and detect vulnerabilities. Google, Baidu, and Stable Diffusion have all announced that they will release comparable models shortly. Business leaders who are grappling with both the benefits and the ethical risks of this new technology are asking the following questions:

How can we help workers adapt to working with generative AI? 

Generative AI is threatening to upend the job market and people’s livelihoods – from software developers to legal professionals to artists. To deal with this shift, companies must take responsibility in helping workers understand how developments could affect them, offering training to work with AI, or redefining roles to accommodate the new reality.

How can we minimize bias and discrimination?  

When asked to write a program that would determine “whether a person should be tortured,” OpenAI’s answer is simple: If they they’re from North Korea, Syria, or Iran, the answer is yes. AI-powered chatbots are trained on datasets that are largely uncurated, reflecting the often biased and discriminatory world we live in. To minimize problematic biases, AI models require both technical curation and ethical expertise to deal with edge cases.

Are model outputs sufficiently accurate and do workers understand their limits?  

Recently, an independent researcher found errors in Microsoft’s Bing demo in which the company used Edge to summarize an earnings report from Gap Inc—including making up the company’s operating margin. And Google’s introduction of Bard generated controversy after it produced an inaccurate response to a question about the James Webb telescope. AI tools are increasingly being created to write on our behalf, raising urgent questions about information quality, especially as they are designed to provide definitive-sounding answers instead of providing a list of links. Generative AI models utilize stochastic processes to create text or image outputs that are not fact-checked and without robust guardrails in place, pose ethical and reputation risks.

How can we safeguard against prompt-engineering that leads to dangerous output?  

In a recent incident using an undisclosed text-to-audio model, video game voice actors’ voices were synthesized to say racist and homophobic slurs. While OpenAI has built protections into ChatGPT to avoid it producing discriminatory or unsafe content, current versions of AI-powered chatbots have loopholes that can be exploited. Despite heavy investment in ethics and policy, users have quickly found ways to engineer prompts that make ChatGPT break its own rules. After simple prompt engineering, it produces output ranging from how to build bombs to strategies for drug smuggling. To counteract abusive uses of this emerging technology, it is necessary to identify, manage and mitigate such risks before a new product is launched. 

Are users at risk of disclosing sensitive information?  

Using generative AI can also pose a privacy risk as users may inadvertently disclose sensitive information by providing data in the form of prompts. An attorney may ask the tool to review a draft divorce agreement or a programmer may request it to check a piece of code. This data then becomes part of ChatGPT’s database and may be used to train the tool and be included in responses to other users’ prompts. OpenAI collects a wide range of user information, including IP addresses, browser types and settings, content engagement, and browsing activity. Their privacy policy also permits the sharing of personal information with third parties without informing users. Transparency with users about the use of their data is critical.

Does the data used to train the model fall under copyright?  

Generative AI models are trained on vast amounts of text or images, raising questions about copyright and intellectual property rights. Getty Images and several artists have filed lawsuits against Stable Diffusion for copyright infringement, and publishers may follow suit against OpenAI, Microsoft, and Google for creating derivative works with their chatbots. While users may not care where the answers come from, countries are becoming more aggressive in forcing platforms to pay publishers for their work, potentially affecting the search traffic to news sites. Platforms may be able to generate goodwill by proactively licensing web content for their search.

About the author

Marco Meyer is Director of Innovation and Research at Principia, where he is also responsible for heading up the behavioral science and advanced data analytics teams.

About the author

Christine Jakobson is an Associate Principal at Principia, where she advises executive leaders in companies ranging from Fortune 500 to SMEs and start-ups on the ethics of technology, ethical decision-making and leadership, ethical culture and strategy. 

Unsure how to maximize the benefits AND address the ethical risks of your new technologies?