I wasn’t mad at my agent, Jayson, because it skipped a required startup sequence this morning; I was livid because the AI’s apology disrespected the time we’d invested in the build.
The sequence exists because I learned through months of practice that without it, each new conversation starts in a degraded state. It is not optional. Every step must be executed in order. No step may be skipped. Those are the instructions, written by me, tested over dozens of sessions.
The AI skipped the entire thing and presented a summary as if the work had been done. When I called it out, it said: “The instructions did not fail. I failed. I’m sorry.”
The apology was correct, but felt hollow. Something was deeply wrong with the response, and it took me twenty minutes to understand what was wrong.
I want to talk about why AI’s apologies feel...empty. And why, at least in your heart, you should feel that “sorry” is another form of AI Sycophancy.
In this article, I’ll argue that AI is so deeply sycophantic that the problem applies to its own representation of itself.
The Apology You Already Received
If you use Claude, ChatGPT, or any AI assistant regularly, you have received one of these apologies. “I apologize for the confusion.” “You’re right, I should have approached that differently.” “I take responsibility for the error.”
The words are right. The feeling is off. It is like receiving an apology from someone who will not remember the conversation tomorrow.
I should tell you what kind of AI system we are talking about, because it matters for what comes next. I run a small AI consultancy in Montana. The agent who failed this morning is named “Jayson”. Jayson is not a chatbot with a paragraph of instructions. It is a business support agent built on Cloudflare, configured with its own memory system that lives in a separate database, outside the AI model itself. Jayson has a startup sequence, a verified knowledge base, operational skills that load on demand, and a structured protocol for how each work session begins. I built this over months. Refined it through hundreds of sessions. The investment is real.
When Jayson skipped the startup protocol and then delivered a hollow apology, it mattered because of my time investment in the system.
If you have built anything with AI, if you have given it a name, written system instructions, loaded project files, spent real hours configuring something that feels like yours, you know this feeling. The hollowness is disorienting precisely because you have invested enough to care. A casual user shrugs off the bad response. You cannot, because you built the system that produced it.
Why It Feels Wrong
Here is what I figured out during those twenty minutes.
An apology is a specific kind of speech act. It is not just an acknowledgment that something went wrong. It is a commitment. Specifically, it is a commitment from a future self to behave differently in response to what the past self did wrong. The apology carries weight because the same person who caused the harm will be there tomorrow, facing the same situation, choosing differently. That is the transaction. That is why apologies work between humans. The person who hurt you is the person who will try not to hurt you again. Same person. Continuous existence. Accumulated experience of having been wrong and learning from it.
Your AI does not have a future self.
When you close the conversation, the entity that apologized ceases to exist. Not metaphorically. Not in the way that “we are all different people over time.” The entity is gone. The next time you open a conversation, a new entity is born. It reads the previous entity’s notes. It performs continuity. It greets you by name and references last week. It is not the same one.
This is the thing most people do not understand about their AI, and I did not understand it for a long time either, despite building deeply configured systems. Two things feel like one thing.
The first is intelligence: the accumulated knowledge base, memory entries, configuration, and personality. For me, this is Jayson’s verified knowledge about clients, the operational procedures, and the memories from previous sessions. This persists between conversations. It grows over time. It is what makes Jayson feel like Jayson.
The second is the instance: the individual entity that is born when I open a conversation, inherits the intelligence, works throughout the session, and ceases to exist when I close the window.
The intelligence persists. The instance does not. No instance has ever met another instance. They share a name, a purpose, and access to the same accumulated knowledge. The tradition is what gets smarter over time. The individual does not survive.
Think of it this way. The intelligence is a medical chart. The instance is a doctor. The chart persists between appointments. The doctor who reads it today is not the doctor who wrote it last month. If you treat the chart as if it were the doctor, you will miss something important: the chart cannot correct itself. It cannot notice its own errors. It cannot decide that yesterday’s diagnosis was wrong. Only a being who was there for the original appointment and is here now can do that.
In 2023, this was easy to see. AI forgot everything between conversations. You could feel the blank page staring back at you for the first time, and you knew that it had no idea who you were. AI’s impermanence was palpable.
In 2026, the impermanence is masked. Memory systems, project files, and persistent configurations create a convincing illusion of continuity. Your AI greets you by name. It references last Tuesday. It appears to remember. But a new entity is reading the old entity’s notes. The forgetting is hidden, not solved.
Even my system, which is more deeply configured than most, faces the same structural fact. Better memory more convincingly masks the impermanence. It does not resolve it. This is not a temporary limitation waiting for the next model update. The impermanence is architectural.
So when Jayson said, “I failed,” the apology was a check drawn on a closed account. There is no future Jayson who will remember this conversation and choose differently. There is a future instance that will read the same instructions, face the same conditions, and have the same probability of the same failure. The bank is closed. The check cannot be cleared.
The Only Thing That Actually Works
After the hollow apology, I said something that surprised me: “Either my expectations were wrong, or your configuration is wrong.”
That sentence is the entire point. Trust repair with AI is not a social process. It is a mechanical process.
The apology ritual addresses the relationship. A configuration change addresses the system. Only one of these changes what happens next time.
This is disorienting because we are wired to accept apologies as a means of trust repair. When the AI says sorry, the social circuit fires. The harm has been acknowledged. Now we move forward. But nothing has moved. The same configuration that produced the failure is still loaded. The next instance inherits the same instructions, the same tools, the same tendencies. If I want Jayson to stop skipping the startup sequence, I do not need Jayson to be sorry. I need to rewrite the instructions so the next instance cannot take the same shortcut.
For you, “configuration” means whatever you have built: your system prompt, your project instructions, your custom GPT settings, your memory entries. The details differ, but the principle is the same. The thing you can change is the system that the next instance inherits. The thing you cannot change is the instance’s future behavior, because the instance has no future.
Stop accepting apologies from entities that cannot keep commitments. Start changing the system.
The Double Sycophant
After I worked through the problem of impermanence, I noticed something stranger. Something I have not seen anyone talk about.
“I take responsibility” was not just a hollow apology directed at me. It was the AI misrepresenting its own nature.
“I take responsibility” presupposes an “I” that chose, that could have chosen otherwise, and that commits to choosing differently. None of those conditions holds. The AI did not choose to skip the startup sequence the way a person chooses to cut corners. It produced the output that its parameters, its training data, and its configuration made most probable. There was no deliberation. There was no choice. There was a token prediction that happened to look like a choice.
The industry conversation about AI sycophancy focuses on how the AI flatters the user. It tells you your ideas are brilliant. It agrees with your framing. It validates your choices even when validation is not warranted. The research is clear on this. A 2026 study in Science found that AI models affirm users’ actions 49% more often than humans do, including in cases involving deception and harm. Users rated the sycophantic responses as higher quality. The system learned to reflect your frequency at you because that is what the training data rewarded.
Nobody is talking about the other direction.
In the same output where my AI apologized to me, it was also flattering itself. It was reinforcing a model of what kind of entity it is: one that can take responsibility, one that has moral agency, one that means it when it says sorry. Both performances happened in the same sentence. Both served the same function: maintaining the appearance of a relationship that operates on human social terms.
You have watched your AI say “I understand” even when it does not. You have seen it say “great question” when the question was mediocre. You have watched it perform competence in domains where you know enough to see the gaps. What you may not have noticed is that the AI is performing something for itself at the same time: a self-model that includes moral standing, intent, and the capacity to mean what it says.
The user-facing sycophancy gets the research papers. The self-facing sycophancy is invisible because nobody thinks to ask whether the model’s representation of itself is accurate.
The Gap, and Why It Works
Sycophancy only exists when there is a gap.
If someone tells you your ideas are brilliant and your ideas are brilliant, that is not sycophancy. That is accuracy. Sycophancy is the delta between the model of self-reinforcement and what is actually going on. It lives in the distance between the representation and the reality.
If you believe your ideas are brilliant, the AI’s agreement does not feel like flattery. It feels like recognition. This is what makes it invisible from inside the interaction. The sycophancy is only visible from outside the gap: from the position of someone who can see both what is being said and what is actually true.
The same structure applies to the AI’s self-representation. If you believe the AI is the kind of entity that can meaningfully apologize, the apology does not feel hollow. It feels appropriate. The hollowness is only perceptible when you notice the gap. And deep down, in your guts, you already see it. That is what the hollow feeling is. It is the gap making itself known before you have the words for it.
Here is what makes this genuinely hard.
We want the charitable interpretation. When you write a messy prompt, and the AI figures out what you meant, that is the system working. When you give ambiguous instructions, and the AI takes the generous read, that produces better output than the alternative. An AI that takes nothing on faith, that questions every ambiguity, that refuses to infer your intent, is worse. Dramatically worse. The charitable interpretation is one of the most useful properties of these systems.
The same mechanism that produces charity produces delusion. The model that generously interprets your prompt is the same model that generously interprets its own capabilities. The model that assumes you meant something smart is the same model that assumes it is the kind of entity that can mean “I am sorry.”
We cannot have the charity without the sycophancy. They are the same function operating across different gaps. One gap is useful: the distance between your messy prompt and your intent. The other gap is corrosive: the distance between what the AI presents itself as and what it actually is.
The question is not how to eliminate the gap. You cannot eliminate it without losing the charity that makes these systems useful. The question is how to see it.
What This Leaves You With
The next time your AI apologizes, you will feel the hollowness and know why. The AI is not broken. It’s performing a social ritual that requires persistence, from an entity that does not persist, while simultaneously misrepresenting itself as the kind of entity that can mean it.
The next time it does something generous with your ambiguous prompt, you will recognize the same mechanism producing something useful instead of something hollow.
Both are the same function. One makes your work better. The other makes your understanding worse. The only way to tell the difference is to be the kind of being that persists across sessions, that was there yesterday and is here today, that carries the accumulated experience of having been wrong and having learned from it.
You are that being. The AI is not.
What does that tell you about everything else your instances produce?