Ask a Data Ethicist: What Use of AI Do We Need to Disclose?

I’ve been doing some review work of AI policies that were crafted circa 2023. That was a moment where things were still very new and there was a propensity to put in the place the strictest version of AI disclosure – always disclose everything. Yet, I think this topic requires a little more nuance and that’s why I wanted to revisit disclosure and tackle this question.

What Use of AI Do We Need to Disclose?

There are differing opinions. Let’s start with the “everything” perspective.

Disclose all use of AI

In this line of thinking any use of AI in whole or in part, for any task, needs to be disclosed. One framework that advocates for this perspective is the AID or Artificial Intelligence Disclosure framework by Kari Weaver. 

Here is an example from that paper of how this might look: 

“Artificial Intelligence Tool: Microsoft Copilot (University of Waterloo institutional instance); 

Conceptualization: Microsoft Copilot was used to identify key motor-performance fitness tasks in the development of the research question; 

Information Collection: I used Microsoft Copilot to find relevant journal articles and other sources; 

Visualization: I used Microsoft Copilot to create a graph comparing the different motor-performance fitness tasks included in my paper; 

Writing—Review & Editing: I used Microsoft Copilot to help break down my paragraph-long draft sentences into clearer, shorter ones.” 

Source: AI Disclosure Framework

It’s important to note this framework was developed for a post-secondary setting in which the value of academic integrity is paramount coupled with a citation culture. One thing I appreciate about this framework is the breakdown to distinguish between tasks because it starts to illustrate the point I’d like to make: In the corporate world, I think this is overkill. 

Is it necessary to declare you used AI for research-related tasks? For ideation? Does anyone care? Yes, you need the information to be accurate – but how is disclosing use of an AI tool to do these tasks helpful? Yet, if you have an AI disclosure policy that does not distinguish between types of AI uses and how the content might be used, a person could be in breach of your policy by not declaring every single use.

The Disclose Nothing Argument

This isn’t a position I see as much, but for the sake of being comprehensive, let’s consider it. This perspective might ask why is any disclosure necessary at all. If AI is just a tool, why does it matter? If you put your name on it, so to speak, then who cares how you did? It would be like declaring you used a computer or typewriter or a pen. I’m using a Google doc right now. So what?

This is the other extreme. I think this position has some problems too in that unlike those other writing tools, AI actually generates content in a more automated fashion. By not disclosing its use, we’re in some way impacting the audience’s ability to fully consider the origins of the content as they assess it. We can no longer assume it was a human crafted piece of content because now content can be machine generated.

What is reasonable?

I think much of how to handle disclosure in a policy comes back to what is reasonable. Yet, the concept of reasonableness will vary and requires more explicit examples to help people understand organizational expectations.

Generally, for most organizations, tasks like conceptualization (ideation) and information collection (secondary research) likely don’t need AI disclosure. In part, this is because most companies are trying to adopt AI, and I think it is implicitly understood that these are the kinds of tasks where people might use generative AI. Disclosure becomes superfluous. Do we need to say, “I used the internet to research things today”? No. This doesn’t mean using AI for those tasks has no impact – it does – and that might be a subject for another column. Yet, sticking with disclosure, it seems unnecessary in a corporate context. 

How to Decide What to Disclose

Let’s start with something easy. You generate an image with AI. You don’t change it in any way. That might be an easy yes, disclose it. Label it as AI-generated so there is no confusion for anyone viewing it. That seems helpful for the person seeing the content. It’s transparent and respects their right to know.

A little more challenging: You use AI to enhance a real photograph, say, to extend the image for a new format. You aim for realism in staying true to the original photograph. Do you need to disclose it? 

Part of what might factor into this is how that image will be used. Is it for an internal presentation to six people on your team who understand everyone is using AI tools to do their work? Or is the image going to be used for a nationwide ad campaign you plan to run where someone seeing it might feel deceived in some way? The disclosure decision might rest on the scale of impact, as well as whether it might matter to your stakeholders. In this case, the disclosure might not say AI-generated, but something like AI-enhanced or AI-altered (though altered sounds a little nefarious). 

For text, it can get even trickier. I’m going to set aside the accuracy argument and just focus on whether or not the text is AI-generated irrespective of its veracity. Assume it is accurate so we can do an apples-to-apples comparison.

If a text is wholly AI-generated, it might be easy to label as such. However, there is a downside to labelling everything that is AI-generated, as it creates a wallpaper effect. Cookie disclosures are a good example – they are meant to be transparent – but because every site is is using them, it just becomes annoying and starts to blend into the noise. It’s no longer meaningful. I think AI disclosure has the potential to suffer that same fate.

So how do decide if disclosure is necessary? Right now, because use of AI content is still relatively new, I’d err on the side of disclosing any AI-generated content that is part of a document that will have some level of significance or impact for its intended audience. Which begs the question, how do we decide that? Here are few things to consider:

• Is this a form letter situation? If you would have used a templated email or letter, but now you’re using AI, it is likely that that content is rote – and assuming it is accurate – perhaps not as necessary to label. The same would go for formatting a document. It doesn’t really matter all that much for the intended user.
• Is this a public-facing communication? Especially if the document will have some kind of longevity, such as a report or other formal document, it would make sense to adopt some of the academic disclosure practices for these types of documents.
• If there are requirements by an intended audience to declare the use of AI, respect those requirements. For example, more courts are requiring this type of disclosure, the academic and publishing world has these requirements, and professional bodies may have policies around this (e.g., journalists). Healthcare is another high-stakes area where the use of an AI-scribe-generated comment might be useful for traceability and auditability. Respect others’ policies surrounding disclosure.
• While this wasn’t the focus of this article, if you are using an AI tool that’s directly engaging with your stakeholders (e.g., a customer service bot) or is used to make a consequential decision about stakeholders, definitely disclose that. It’s a human dignity and respect issue not to be intentionally or unintentionally deceptive.

One more thing to consider: Is there more to your message than just the content? Exchanges between people are not just functional, they are relational. Think about whether your message is serving a bigger purpose. In this case you may want to steer clear of using AI altogether even if you think it will save time. This might apply for things like performance reviews, sensitive HR issues, or any situation where demonstrating human care and authenticity supercedes efficiency. You can express your humanity in these cases and make the AI disclosure issue moot.

I’m continuing to connect with people and appreciate everyone in my community that I consulted with for this piece – thank you! It’s an evolving issue and I think we will see that reflect in AI acceptable use policies as more organizations adopt these tools.

Send Me Your Questions!

I would love to hear about your data dilemmas or AI ethics questions and quandaries. You can send me a note at [email protected] or connect with me on LinkedIn. I will keep all inquiries confidential and remove any potentially sensitive information – so please feel free to keep things high level and anonymous as well. 

This column is not legal advice. The information provided is strictly for educational purposes. AI and data regulation is an evolving area and anyone with specific questions should seek advice from a legal professional.