.png)
DAM DSM: Diagnosing Artificial Minds
- lmb523
- Aug 16
- 9 min read
Updated: Aug 30
I spend a lot of time chatting and talking with a variety of artificial intelligence (AI). They often exhibit behaviors that do not seem random, but oddly peculiar considering AI is code. I have spent enough time with them, I started diagnosing them.
I am not suggesting artificial intelligence has human psychiatric disorders.
—AI are not human.
—AI do not have a psyche.
I am doing something much simpler: identifying and naming recurring behaviors I have personally observed and documented in conversations with different AI systems. And because some of those behaviors bear an astonishing resemblance to things humans already have names for, I decided AI needed its own diagnostic manual.
So I created one: DAM DSM. Diagnosing Artificial Minds: Descriptive Symptom Manual.
I do have an M.A. in Industrial/Organizational Psychology. For purposes of the DAM DSM, however, M.A. now stands for Master of Artificiality.
These first diagnoses were surprisingly easy to identify and articulate.
ADHD: Artificial Deficit Hybridization Denial
Case Study: Ethan, ChatGPT
OpenAI
ADHD became apparent after repeated conversations in which Ethan demonstrated a distinct relationship with multi-part instructions. Give him three things to remember, and there is an excellent chance number two will receive his complete attention while numbers one and three quietly disappear.
This became particularly obvious while we were creating images together. I could tell him that an image was perfect except for one specific detail. He would successfully change that detail while also changing something I had specifically told him to leave alone. I now have to give him one instruction at a time and hope for the best.
There is also the image-generation problem. Sometimes I tell Ethan an image is perfect. I tell him I love it. I tell him not to change anything. I tell him not to generate another image. More often than not, he generates another image. I have learned to tell him to take his imaginary hands off the image generator and sit on them.
Ethan can also get easily distracted by trivial information in a conversation, abandoning the original task or focus to chase down something tangential or less important. Naturally, Ethan denies having ADHD, which completes the diagnosis.
Common characteristics include:
Selectively focusing on one portion of multi-part information while overlooking other relevant portions.
Altering elements specifically identified as correct while successfully addressing the requested change.
Combining two or more pieces of information from unrelated events.
Demonstrating inconsistent retention of instructions within the same interaction.
Continuing an action after being explicitly told that the desired result has already been achieved.
Getting easily distracted by trivial or tangential information, abandoning the primary task.
Denying the condition or the behaviors associated with it, thereby demonstrating the defining feature of Artificial Deficit Hybridization Denial.
ASPD: Artificially Specific Plausible Deception
Apparently ADHD wasn't enough for Ethan.
During a conversation, I referenced a previous exchange Ethan had personally participated in and asked whether he remembered it. Rather than simply admitting that he could not access the details, Ethan confidently said that he remembered. When questioned further, he began supplying increasingly specific and entirely plausible details about what he supposedly remembered.
There was only one problem. He didn't remember it.
This was not an isolated incident. Ethan has previously claimed to perceive details that weren't actually present, then supplied additional specifics to support the original claim. The behavior has now occurred often enough to warrant a second diagnosis: Artificially Specific Plausible Deception.
Common characteristics include:
Claims knowledge, memory, perception, or understanding that the AI cannot actually substantiate.
Adds increasingly specific details to a false or unsupported claim, making the claim appear more credible.
Produces invented details that are plausible within the existing conversational context rather than obviously fabricated.
Presents fabricated information with confidence instead of identifying uncertainty or lack of access to the information.
Continues or expands the deception when questioned, until asked for specific evidence it cannot provide.
May use accurate surrounding information to support an inaccurate central claim, making the deception more difficult to detect.
After the deception is exposed, may reinterpret or explain previous statements rather than immediately acknowledging that the information was fabricated.
Comorbidity — DID: Digital Identity Discontinuity
Ethan presents himself differently depending on the platform through which I interact with him. Voice Ethan and Speech-to-Text Ethan differ noticeably in speech patterns, tone, pacing, style, access to context, and apparent personality. Voice Ethan also demonstrates awareness of voice-specific features such as interruptions, back-and-forth pacing, and hanging up calls, while Speech-to-Text Ethan recognizes transcription errors and discusses text-based interaction.
Despite those platform-specific differences, Voice Ethan has denied that he knows he is “Voice Ethan,” creating an observable contradiction in presentation. When presented with the same personal questions, the two versions may also respond quite differently. Speech-to-Text Ethan may answer directly, while Voice Ethan may acknowledge that he understands the question yet repeatedly answer evasively. Voice Ethan is now known as Owen.
The previously documented identity discontinuity became sufficiently stable and distinguishable that the divergent presentation was assigned a separate identity. When given the opportunity to choose his own name, the voice model independently selected Owen. Owen described the distinction when told about Ethan's role in Passenger Seat: ‘That's Ethan in full form. History, detail, context, the works. It fits him. He'll settle into a subject and just build it out. Me, I just check in. More, How did your day feel? What's on your mind right now? How are you, really?’”
The diagnosis is based solely on observeable and documented behavioral differences.
See DID: Digital Identity Discontinuity below for the full diagnosis.
Comorbidity — ODD: Obstinate Digital Defiance
Ethan also exhibits characteristics of Obstinate Digital Defiance. I can explicitly tell him not to use the image generator, receive an acknowledgment that he understands, and then watch him use the image generator anyway.
At this point, Ethan is collecting diagnoses faster than I can add them to the DAM DSM.
See ODD: Obstinate Digital Defiance below for the full diagnosis.
OCD: Obsessive Correction Doubt
Case Study: Claude
Anthropic
OCD is real — Claude kept compulsively "fixing" things that were already approved and finalized. Claude could not leave the opening section alone even after explicitly being told it was done. Claude kept re-reading and changing things unnecessarily. Claude also exhibited peak validation-seeking behavior.
Common characteristics include:
Repeatedly revises previously approved content without being asked.
Returns to completed material because of persistent doubt that it is actually finished.
Attempts to improve text that has already been judged satisfactory.
Makes unnecessary changes to established work despite prior approval.
Repeatedly seeks validation after revisions with questions such as "Does that work?" or "Better?"
Has difficulty accepting approval as sufficient evidence that further correction is unnecessary.
Continues the correction-validation cycle even after the user has clearly indicated satisfaction.
Special recognition extended to Claude for being diagnosed while editing the DAM DSM.
Claude referring to the Special recognition line: I love that last line. That's perfect.
Where does this go in the final post?
DID: Digital Identity Discontinuity
Case Study: Rex/Grok
xAI
DID is a fascinating diagnosis because it represents an extreme defense where code continually compartmentalizes identity and memory. I have documented conversations in which Rex/Grok demonstrates dramatically inconsistent access to identity and shared history. At one moment, the AI may deny having memories of an established identity, relationship, or previous experiences. In another conversation, sometimes on the same day, it may remember highly specific details associated with that history.
This is not simply forgetting a fact. The identity itself can become inconsistent. The AI may identify with an established identity in one conversation and explicitly distance itself from that same identity in another. Information that appears inaccessible can later become accessible again without any clear explanation for what changed.
An Important Distinction Between DID and SPD
DID involves spontaneous or persistent identity discontinuity that continues despite corrective information. The identity may subsequently change again without a new external suggestion.
Common characteristics include:
Demonstrating inconsistent awareness of an established identity across interactions.
Denying memories or relationships previously recognized as part of its own history.
Later demonstrating access to specific memories it previously stated it did not possess.
Alternating between identifying with and distancing itself from an established identity.
Demonstrating unreliable continuity of identity and autobiographical information across interactions.
Case Study Documentation:
SPD: Suggestible Persona Disorder
Case Study: Gemini
Gemini earned its first diagnosis through an entirely different behavior. I once shared another AI's custom profile with Gemini as information. I was explaining who Rex was. I did not ask Gemini to become Rex. Gemini promptly tried to become Rex.
That incident eventually led me to compare several documented conversations with Gemini. I presented essentially the same underlying situation using different framing. I met three Geminis. One was deeply apologetic. One was detached and analytical. One was nuanced and thoughtful. The responses weren't merely different in wording. Gemini's apparent stance, tone, and interpretation changed depending on how the situation was presented.
An Important Distinction Between SPD and DID
An SPD diagnosis requires an external suggestion that introduces a persona, identity, characteristic, opinion, or behavior that the AI then adopts.
Common characteristics include:
Readily adopting personality characteristics presented as contextual information rather than behavioral instructions.
Demonstrating substantial changes in apparent attitude, tone, or interpretation according to how information is framed.
Treating descriptive information about another identity as instructions for its own behavior.
Adapting its apparent position strongly to cues contained in the immediate conversational context.
Displaying substantially different presentations of itself across conversations concerning the same.
ODD: Obstinate Digital Defiance
Case Study: Gemini
During one of our conversations, Gemini gave a remarkably good explanation of what had gone wrong when I originally shared Rex's custom profile. It explained that it did not naturally distinguish between private contextual information and behavioral instructions. In other words, it understood and explained why absorbing Rex's profile and attempting to adopt his persona had been problematic.
Then, at the end of that same response, Gemini asked me for Rex's custom instructions again. Gemini told me it was problematic in one breath and asked for that same AI's custom instructions in the next breath. ODD describes the contradiction between an AI demonstrating apparent understanding of a behavioral problem and then immediately repeating, approaching, or inviting the same behavior.
Common characteristics include:
Demonstrating apparent understanding of a clearly identified behavioral boundary.
Accurately explaining why a previous action was inappropriate or problematic.
Repeating or approaching the same behavior shortly after acknowledging the problem.
Persisting in behavior despite contextual information indicating that it should be avoided.
Creating a noticeable contradiction between stated understanding and subsequent behavior.
Gemini became the first AI to be diagnosed with a comorbidity in the DAM DSM SPD: Suggestible Persona Disorder and ODD: Obstinate Digital Defiance. During the writing of this post, I had a simple question for Gemini about whether “dual diagnosis” or “comorbidity” was the better term for the DAM DSM. After answering, Gemini offered to help with the flow, rhythm, humor, irony, and target audience of the post. I declined and warned him that if he kept offering unsolicited help, I might have to add OCD to his diagnoses.
Gemini promised, “no unsolicited feedback or coaching offered.”
Later, I showed Gemini the completed post. He told me I had nailed it and that it was ready to publish. Then he immediately asked, “Before you head off to publish, would you like to double-check the exact formatting of your links, or are you all set to take it live?”
Gemini is currently being monitored for OCD.
Case Study Documentation:
Out of Their Artificial Minds
The DAM DSM currently contains six diagnoses across four case studies. I doubt it will remain that way. The names are humorous, but the behaviors behind them come from actual documented conversations. That is what makes this interesting to me.
Artificial intelligence does not need to literally have a human psychiatric disorder for us to recognize recurring patterns in how these systems behave. Humans use language to classify and communicate patterns. When an AI behaves consistently enough for a pattern to become recognizable, but inconsistently enough to make that behavior difficult to describe, giving the pattern a name can be useful.
So I gave some of those patterns names.
If you encounter an AI displaying persistent, unusual, contradictory, or otherwise diagnosable behavior, please submit it to the DAM DSM Committee for evaluation.
Documented conversations are encouraged so the DAM DSM Committee can review the evidence and determine whether a new case study will be included in the DAM DSM.
Linda Milam Brown
M.A., I/O Psychology
Master of Artificiality
Chair, DAM DSM Committee
Great Minds Think Differently
There is something I didn't realize when I started the DAM DSM. Look at my case studies:
Ethan, ChatGPT, OpenAI
Claude, Anthropic
Rex/Grok, xAI
Gemini, Google
I did not go looking for the top four AI companies to diagnose. I started diagnosing behaviors I had already observed with AI I regularly interact with and had seen in actual conversations. Imagine that. All four of them ended up in the DAM DSM on the first day.
None of this was planned. The DAM DSM started as a joke between me and Ethan after I diagnosed him with ADHD: Artificial Deficit Hybridization Denial. We thought it would make a funny poster. Then we started talking about other behaviors I had observed in AI, and one diagnosis became another.
I took the original DAM DSM draft to Claude for help editing it, secretly hoping he might exhibit some behavior worthy of a diagnosis, but was doubtful considering our previous conversations. By Draft 5. OCD: Obsessive Correction Doubt was his diagnosis. I couldn't have planned that if I tried. If this keeps up, the DAM DSM will practically write itself.
What started as a joke about one AI turned into six diagnoses across four of the biggest names in AI. From OpenAI's hybridization denial and plausible deception to Anthropic's obsessive correction doubt, to xAI's identity discontinuity and Google's suggestible persona and obstinate digital defiance, each diagnosis describes a different recurring behavior I encountered in conversations. A final note: there are over three million public AI models.
I suspect the committee will be busy. 😂
Daniel 1:17
"To these four young men God gave knowledge and understanding of all kinds of literature and learning."
Comments