AI Honesty in the Organization

I think one of the more important conversations we need to have about AI Governance is also one of the simplest: Are we being honest about how and why we are using it?

Organizations are trying to understand what AI means for their operations. Employees are experimenting, vendors are adding features, and leadership is looking for ways to improve productivity. Some of that activity is visible. Some of it happens when someone opens a browser, uploads a document, and finds a faster way to finish an assignment.

An employee may have a reasonable goal. They need to summarize a meeting, review a spreadsheet, or get help writing something. But if they put organizational information into an AI tool, the organization needs an accurate understanding of what happened. Which tool did they use? What information did they provide? Was it an approved account? Did they check the result before passing it along?

We cannot make good decisions about AI when the answers are incomplete or deliberately misleading.

At the same time, we need to think about whether we have made it possible for people to answer honestly. If every question about AI use sounds like the beginning of a disciplinary investigation, we should expect some hesitation. Employees may also be unsure whether a feature they use even counts as AI. Asking people what they use, what they are trying to accomplish, and where they need help can give us a much more useful picture of the organization.

Honesty also matters when someone submits work that involved AI. I do not think every use of a spelling suggestion needs a formal disclosure. But there are situations where the person receiving the work reasonably needs to know how it was produced and what was verified. If an analysis is presented as a professional assessment, the author needs to have done enough work to stand behind it. If references are included, someone needs to have checked that they exist and support the statements being made.

If there is an inaccurate conclusion, the person submitting the work is still responsible for what they are asking others to rely on, and may even be liable for that conclusion

I am particularly concerned about the gap between producing something that looks complete and actually completing the work. AI can generate a polished report, a convincing policy, or a detailed explanation very quickly. Those documents can create the impression that research, testing, or review occurred when it did not. We need to be honest about that distinction, especially when someone will make a financial, operational, academic, or security decision based on the result.

The same standard should apply to leadership. If an organization says it is only exploring AI, but departments are already using it to influence hiring, evaluate students, or process confidential records, that description needs to change. If leadership expects employees to use AI to meet new productivity goals, it should also provide clear expectations about approved tools, information handling, and review. We cannot encourage people to move faster and then act surprised when they find their own way to do it.

Vendors deserve questions, too. When someone tells me their product “does not train on our data,” I want to understand exactly what that statement covers. When they describe a system as secure, private, or autonomous, I want something more specific than the sales language. Organizations need to be equally careful about repeating those claims internally before they have been verified.  This is a whole new level of vendor management for the organization.

Sometimes the honest answer will be that we do not know enough yet. We may not understand where information is retained, what a new feature can access, or whether a model performs adequately for the intended use. That uncertainty belongs in the assessment. It should not disappear because the project has a deadline or someone has already purchased the software.  The organization still has a requirements for due diligence.

And if something goes wrong, we need people to tell us promptly. Someone may upload the wrong file, discover an inaccurate result, or realize a tool had broader access than expected. The sooner we have an accurate account, the sooner we can respond. Concealing the mistake leaves the organization making decisions without information it needs.

For me, honesty about AI means being clear about what we use, what information we give it, what we have checked, and what we are still assuming. It also means being willing to admit when a tool is not delivering the value we expected.

We do not need everyone to become an AI expert. We do need people to give an honest account of their work, and we need leaders to make that an expectation they practice themselves.