The Intersection of Transparency, Effort, and Care
Content warning: mention of school violence
On February 16, 2023, three days after a gunman killed three students at Michigan State University, the Office of Equity, Diversity and Inclusion at a different school— Vanderbilt's Peabody College— emailed their student body. The message talked about building community and caring for one another.
At the bottom, in smaller font, sat a parenthetical: "Paraphrase from OpenAI's ChatGPT AI language model, personal communication, February 15, 2023."
Students noticed. The email also referred to "recent Michigan shootings," plural, when there had been one. The next day the associate dean who signed it apologized, calling the decision poor judgment. Within a week she and an assistant dean had stepped back from their responsibilities pending a review, and the college dean told CNN that the email had skipped Peabody's normal multi-layer review and that university administrators had not seen it before it went out (The Vanderbilt Hustler, CNN).
One student, quoted in the Vanderbilt Hustler, explained how she felt about it: "There is a sick and twisted irony to making a computer write your message about community and togetherness because you can't be bothered to reflect on it yourself."
Read the email without the context and it is unremarkable: warm and kind of generic. The contents themselves were not hurtful. What people reacted to was the disclosure line: the confirmation that in the days after a shooting at a peer institution, an expression of care and community was outsourced to a machine that can experience neither.
Effort Communicates
We tend to read effort as evidence of caring.
Justin Kruger and colleagues demonstrated this in a 2004 paper. They showed people a poem, a painting, and a suit of armor, and told some of them the work had taken 4 hours and others that it had taken 26. The people told it took longer said that they like the object more and rated it as higher quality. The effect was strongest when quality was hard to judge: when you cannot tell whether a poem is good, you fall back on how hard it looks to have been.
How would a donor evaluate the quality of a thank-you note? How would a community member assess a statement of solidarity? Short of saying something offensive, it’s the thought that counts.
Three Ways This Goes Wrong
Getting caught. In the Vanderbilt case, the content was fine, the disclosure was the problem. Once your audience knows how little the message cost you, they read it differently. A community member who asks a hard question to their government and later learns the response was generated hears a second message underneath the first one: your question was not worth a person's time.
This risk is higher in organizations serving people who already suspect institutions do not take them seriously: small donors, marginalized communities, or participants in programs that keep having their budgets cut, for example. An AI-drafted message reads as thoughtless and confirms their suspicions.
The perverse incentive. If stakeholders read speed as carelessness, organizations face pressure to be visibly slow. A org that processes job, scholarship, or grant applications quickly using AI might worry that the speed itself signals they do not take the decisions seriously. Even if you could be 100% sure that an AI-assisted review is more consistent, fair, and efficient, you still might not want to use it.
Deskilling. When you don’t use your skills, they get rusty. A grants manager who drafts every proposal with AI for two years will find her own writing weaker when she needs it for a funder with unusual requirements, expectations, or programs. Exceptions are where automation fails, too, so that may be the only case you’re asked to write it yourself (For more, see my previous post on deskilling.)
Junior staff have a worse version of this problem, because they never built the skill to lose. The struggle of writing a bad first draft and being told why it is bad is how people learn to write good ones. Take away the struggle and you have taken away the training.
The Disclosure Problem
In a 2025 study, across professors grading papers, analysts writing reports, creatives creating, and investment funds picking stocks, people who disclosed using AI were trusted less than people who did not. A professor who said she used AI to grade was rated less trustworthy than one who said she used a human assistant, and less trustworthy than one who said nothing about their process.
The study’s authors trace the effect to legitimacy: AI use seems unjustified, and peopple should be doing the process. The penalty held whether disclosure was voluntary or required, and whether or not the audience already suspected AI involvement. It shrank, but did not vanish, among people with warm attitudes toward technology.
In Amplify Good Work I suggested that a detailed disclosure of AI use might build more trust than either silence or a bare "AI Assisted" stamp. I still think the detailed version beats the bare stamp. But I don’t know whether it can get around the legitimacy problem.
This is not an argument for hiding AI use, though. The same study found the trust damage from third-party exposure (someone else revealing your AI use) was worse than from disclosing it yourself.
In summary, disclosure costs you something, concealment costs you more if it comes out, and we don’t have evidence that there’s any framing to make the cost zero.
Where Effort Really Helps
Not all effort signals care. In a handwritten note from the executive director to a donor who has given for fifteen years, the effort is the message. That same ED manually reformatting a quarterly financial report communicates nothing to anyone. It’s obvious where the ED should be spending their time between these two tasks.
Deciding which tasks are worth the time is important and context-dependent. The appropriate answer depends on your field, your community, and what your people have come to expect.
This is a familiar question, fortunately. Would you (if you could) delegate this task to an assistant? In the case of the university email at the beginning of this post, they had the dean send and sign the message. That make it seem like they understood the message needed to come from the dean, not, say, the dean’s office manager.
That said, in the past, they might have had the office manager draft it, the dean review it and send it out. In that case, would they have felt the need to disclose that the draft had been written by someone else? Probably not. They wrote the AI disclosure out of a desire to be transparent, but one could certainly argue that the costs outweighed the benefits of that transparency for both the university and the students.
This is not a dilemma with an easy answer, and I believe that the answer will change as AI is further integrated into work and norms change.
In the mean time, I think the best answer for mission-driven workers in this situation is to take a Constrain stance. If you don’t feel comfortable sending it without a disclosure, don’t let the AI write it. Rather than outsourcing the drafting of the email, let AI handle scaffolding. You could ask it for details about the original event, information about people who’ve been through this before find (un)helpful to hear, and an outline for the email. Then the dean can write the email, and sign it without compunction or the risk of a distracting and damaging hullabaloo.
Rule of thumb: before you automate a category of communication, ask whether anyone would feel differently about it after learning a machine wrote it. If yes, that feeling is part of the thing you were producing. Do not automate the product.
A Note on Reinvestment
The reason to do any of this is to save effort from one task to reinvest somewhere else.
A caseworker who no longer has to spend 40 minutes a day on documentation formatting has 40 minutes now. Where does it go? More time with clients? Leaving early? Dealing with the next most urgent thing? A new mission-aligned project? Freed-up time does not automatically flow to what matters: make the decision in advance.
If you are going to automate a report, select what you want to reinvest saved time in, and do what you can to protect that time: put it on the calendar, make sure it’s reflected in your project management system, or build the workflow out before you implement the automation.
Questions for Your Organization
Before you use AI on something people will read:
If the recipient learned a machine drafted this, would they feel differently about it?
Which of your communications are people using as evidence that you care? (Asking them directly is an option)
Which tasks eat your staff's time without signaling anything to anyone?
If you automate a task your junior staff currently learn from, what replaces that learning?
When freed-up hours appear, who decides where they go, and when do they decide?
Is there any task where you would be embarrassed to disclose AI use? What does that tell you about doing it at all, or how you might constrain the AI’s role?
LLM Disclosure: I asked Claude Opus 5 to help me find a case study and draft this based on content from my book. I edited it quite a bit, particularly to make the discussion of the studies more accessible and to dig deeper in the “where effort helps” section— I thought what it gave me was surface-level and too abstract to be useful.
