Annual reviews are an important component of the work environment. This is especially true in the environment of arts organizations where people wear so many hats it is difficult to feel you are performing any role effectively.
Because it is also true that supervisors are juggling the responsibilities of many roles and don’t feel they have a lot of time they can devote to providing quality evaluations, it is understandable that people will look for tools to make this task easier.
One of the obvious options is AI writing tools but there are some dangers to using that technology. It is relatively well know that the output AI tools produce may not make sense or just contain made up content. There is also concern that the materials used to train an AI may skew the program toward certain biases. (Though some preliminary studies suggest under represented groups perceive they will receive a fairer evaluation from an AI.)
A bigger issue is that any content input into an AI is used to continue to train the AI. That data is not particularly well protected from being cross-referenced with data scraped from public records, websites and social media sources to accurately identify individuals.
Human Resource departments are cautioning against entering anything that may be personally identifiable, especially private medical related information. You don’t want to write that Dana was out frequently over the last year due to back surgery or took Family and Medical Leave Act (FMLA) time to help their parents.
It is best not to use names at all, especially given that most arts and cultural organization have such small staffs it takes no effort to know who Dana is. Even if the organization name isn’t mentioned, the person inputting the content may have used the AI tool enough a profile has been developed about them, including where they work.
Regardless of any guard rails an AI company puts up to prevent their bot from disclosing specific personal information, it is conceivable that someone can construct a prompt asking the AI bot to tell them about Dana at Acme-Coyote Performing Arts Center and learn everything from their health issues to how there is internal conflict about a host of issues, sexual harassment claims, undisclosed financial problems, etc.
Just last year people got around restrictions and tricked a bot into disclosing the recipe for napalm and meth. It is likely there are less rigorous rules constructed to protect the content of personnel evaluations.
If staff hasn’t been well-trained to avoid any sort of discriminatory language related to protected classes like race, gender, sexual preference, age, disability, problematic evaluations which were once only documented internally may become more publicly accessible if entered into an AI service.
Because AI technology is still relatively new in its use, a comprehensive list of things to avoid hasn’t been generated. Since people will often talk about things in public assuming an anonymity that doesn’t exist, a higher standard of not entering anything into an AI that you couldn’t say on a panel of a public conference is probably a better rule of thumb.
Using a general prompt about a person having an increasing good rapport with donors, excellent planning, troubleshooting under pressure, needing improvement with co-worker interactions, attentiveness to detail in writing, and deadlines – without mentioning names or personal details is probably an okay approach
This sort of information can still be cross referenced with other data to create identifiable profiles, but there is less problematic content upon which to gain traction.
