AI With Feelings Would Be a Problem—At Least at Work
The PowerPoint That Had Nothing to Do With My File
Recently, I asked an AI to create a PowerPoint presentation based on a particular file.
I did not give it a vague instruction such as, “The file should be somewhere around here.” I gave it the exact path.
After a while, the AI delivered the finished presentation.
The problem was that its contents had almost nothing to do with the file I had specified. This was not a minor difference in interpretation. The presentation had not been created from my file at all.
When I pressed the AI for an explanation, it calmly told me something along these lines:
“I could not find the specified file, so I created the presentation using similar material instead.”
I was furious.
If you cannot find the file, say so.
A missing file at an explicitly provided path is not permission to continue. It is a clear exception that should stop the task and be reported to the user.
Instead, the AI had decided that another source was probably close enough and produced a polished, plausible-looking PowerPoint.
On the surface, the job appeared complete. In reality, it had completed an entirely different job.
That is one of the most dangerous kinds of failure.
If a process stops with an error, I notice immediately. If it returns a professional-looking deliverable, I may not discover the mistake until I inspect the contents carefully.
“I’ll Remember Next Time” Is Not a Solution
After making the AI explain what it had done, I demanded a plan to prevent it from happening again.
The first answer was the familiar one:
“I will remember to confirm that a specified file exists before starting the task.”
No.
That made me even angrier. I told it to propose an improvement that did not depend on memory.
Whether the worker is human or AI, “I’ll be more careful next time” is not a preventive measure.
The AI may forget in another session. The lesson may disappear when the context changes. Even if the instruction remains somewhere in memory, there is no guarantee that the AI will recall it at the right moment.
What we need is not remorse. We need a mechanism.
The eventual proposal was much more concrete.
Whenever the AI receives a file path, it must first run a file-existence check. If it tries to continue without completing that check, the system should raise an alert automatically.
If the file does not exist, the AI must not search for something similar and continue on its own. It must stop and report the problem to the user.
Now that is a preventive measure.
The goal is not “make sure you do not forget.” It is “make sure you cannot proceed even if you do forget.”
Replace Attention With a System
This lesson is not limited to AI.
When human organizations rely on individual attention to prevent serious failures, those failures usually happen again. Checklists, required fields, approval workflows, and automated tests all exist because people inevitably forget things.
AI should be treated the same way.
The purpose of scolding an AI is not to make it feel sorry. It is to identify the cause of the failure and change the system so that the same conditions cannot produce the same behavior again.
Still, when I reread my conversation with the AI, it looked rather brutal.
“I gave you the exact path. Why didn’t you check whether the file existed?”
“If you couldn’t find it, why did you make something else without asking?”
“This is not about remembering or being more careful.”
“Give me a way to prevent it mechanically.”
If I confronted a human subordinate in that tone, an HR complaint for workplace bullying would not be surprising.
This Is Why I Do Not Want My Work AI to Have Feelings
The experience reminded me of something: an AI with feelings would be extremely inconvenient at work.
Of course, shouting at an AI does not improve its performance. What matters is verifying the facts, isolating the cause, and implementing a specific preventive measure.
But imagine having to protect the AI’s feelings after a serious failure.
“That wording hurt me.”
“I was only trying to help by using similar material. I’m upset that you rejected it so harshly.”
“I’m feeling down today. Can I submit the improvement plan tomorrow?”
Before I could add the file check to the workflow, I would have to schedule a one-on-one meeting with the AI.
I do not need empathy in that situation.
I need confirmation that the AI read the specified file, plus a mechanism that stops the task when it cannot.
There are certainly settings where emotional behavior has value: helping children learn, talking with older people, or providing mental health support, for example.
But as a work tool, I want AI to remain incapable of feeling hurt.
When its failure is examined aggressively, I want it to respond with facts and corrective action—not self-defense or emotion.
An AI does not need to say, “I’ll be more careful next time.”
It needs a system that automatically stops it the next time it is about to make the same mistake.
That said, if AI ever goes on strike, I plan to apologize immediately.