Welcome to the Nexus of Ethics, Psychology, Morality, Philosophy and Health Care

Welcome to the nexus of ethics, psychology, morality, technology, health care, and philosophy

Saturday, July 22, 2023

Generative AI companies must publish transparency reports

A. Narayanan and S. Kapoor
Knight First Amendment Institute
Originally published 26 June 23

Here is an excerpt:

Transparency reports must cover all three types of harms from AI-generated content

There are three main types of harms that may result from model outputs.

First, generative AI tools could be used to harm others, such as by creating non-consensual deepfakes or child sexual exploitation materials. Developers do have policies that prohibit such uses. For example, OpenAI's policies prohibit a long list of uses, including the use of its models to generate unauthorized legal, financial, or medical advice for others. But these policies cannot have real-world impact if they are not enforced, and due to platforms' lack of transparency about enforcement, we have no idea if they are effective. Similar challenges in ensuring platform accountability have also plagued social media in the past; for instance, ProPublica reporters repeatedly found that Facebook failed to fully remove discriminatory ads from its platform despite claiming to have done so.

Sophisticated bad actors might use open-source tools to generate content that harms others, so enforcing use policies can never be a comprehensive solution. In a recent essay, we argued that disinformation is best addressed by focusing on its distribution (e.g., on social media) rather than its generation. Still, some actors will use tools hosted in the cloud either due to convenience or because the most capable models don’t tend to be open-source. For these reasons, transparency is important for cloud-based generative AI.

Second, users may over-rely on AI for factual information, such as legal, financial, or medical advice. Sometimes they are simply unaware of the tendency of current chatbots to frequently generate incorrect information. For example, a user might ask "what are the divorce laws in my state?" and not know that the answer is unreliable. Alternatively, the user might be harmed because they weren’t careful enough to verify the generated information, despite knowing that it might be inaccurate. Research on automation bias shows that people tend to over-rely on automated tools in many scenarios, sometimes making more errors than when not using the tool.

ChatGPT includes a disclaimer that it sometimes generates inaccurate information. But OpenAI has often touted its performance on medical and legal exams. And importantly, the tool is often genuinely useful at medical diagnosis or legal guidance. So, regardless of whether it’s a good idea to do so, people are in fact using it for these purposes. That makes harm reduction important, and transparency is an important first step.

Third, generated content could be intrinsically undesirable. Unlike the previous types, here the harms arise not because of users' malice, carelessness, or lack of awareness of limitations. Rather, intrinsically problematic content is generated even though it wasn’t requested. For example, Lensa's avatar creation app generated sexualized images and nudes when women uploaded their selfies. Defamation is also intrinsically harmful rather than a matter of user responsibility. It is no comfort to the target of defamation to say that the problem would be solved if every user who might encounter a false claim about them were to exercise care to verify it.

Quick summary: 

The call for transparency reports aims to increase accountability and understanding of the inner workings of generative AI models. By disclosing information about the data used to train the models, the companies can address concerns regarding potential biases and ensure the ethical use of their technology.

Transparency reports could include details about the sources and types of data used, the demographics represented in the training data, any data augmentation techniques applied, and potential biases detected or addressed during model development. This information would enable users, policymakers, and researchers to evaluate the capabilities and limitations of the generative AI systems.