top of page
  • Writer's pictureEmil Jersling

ChatGPT Game Master: Solving a Closed Room AI Murder Mystery

Updated: Sep 15

AI is complex.


ChatGPT was released by OpenAI on November 30, 2022. In 5 days it had 1MM users and in January 2023 it had 100MM active users. According one co-founder the team was surprised by how quickly it became so incredibly popular.


It been used for personalized tutoring, although it suffers from plagiarism and lies, customer support, better and cheaper than people, to create editorial content including poetry, although not yet to universal acclaim, draft emails, write code, and even for medical diagnoses, with ChatGPT 4.0 reaching an impressive 87.2% accuracy in one study.


In May 2024 I completed Dr Jules White of Vanderbilt University highly rated Prompt Engineering for ChatGPT course on Coursera. If you don’t know how to use the persona pattern (PP), flipped interaction pattern (FIP), or game play pattern (GPP) I can highly recommend it. Prefer to read? Check out his paper: A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPT.


First Experiment: Business Judgement Coach


In my first experiment I used PP and FIP to make it act as a business coach to teach business judgment. It crafted relevant scenarios on market entry timing and investment decisions with a clear bias for qualitative over quantitative metrics. In spite of several disagreements I was awarded a clear passing grade.


Second Experiment: Role Playing Game Master


Next, I wanted to test its competence as an AI game master using GPP.


“I want you to play the role of a game master for a closed room mystery game with a number of interesting characters, one of whom has committed a murder. My task, as the detective, is to determine who committed the murder and what weapon was used through observations and interrogations.”

Emil Jersling


ChatGPT presented the setting, Blackthorn Manor as an ancient estate nestled in the misty countryside, and the characters:


A Suspicious Cast of Characters


An AI generated pop art image of 4 women and 2 men at a dining room table covered in cutlery and blood.
AI number struggles. My text prompt specified 2 women and 3 men sitting at a dining room table, 1 man deceased.

Sir William Montgomery: A stern figure with a passion for history, often found in the library surrounded by ancient manuscripts and artifacts.


Lady Margaret Montgomery: Graceful and well-spoken, she enjoys hosting social gatherings and is deeply devoted to her family.


Dr. Thomas Caldwell: Methodical and observant, he’s well-liked in the community but seems guarded about certain aspects of his life.


Miss Victoria Montgomery: Spirited and creative, she shares a close bond with her father but has clashed with him on occasion over her artistic ambitions.


Mr. Charles Banks: Proper and reserved, he has served the Montgomery family for decades and is privy to many household secrets.


Mrs. Eleanor Finch: Organized and meticulous, she oversees the day-to-day operations of Blackthorn Manor with efficiency and precision.


The Crime: During a lavish dinner party celebrating Sir William’s birthday, a scream pierces the air. Guests rush to find Sir William slumped over the dining table, a deep wound in his chest. The dining room was locked from the inside, and the windows are securely barred.


The Murder Weapon: An ancient dagger, part of Sir William’s prized collection of historical weapons, is missing from its display in the manor’s library.


First Signs of AI Trouble: Divulging the Weapon


A photo of the novelist Agatha Christie from the inside of her book 'Murder is Easy' lying on a wooden table.
Agatha Christie, the Queen of detective fiction. Photo courtesy of Peter576 on Flicker.

The characters and plot felt worthy of an Agatha Christie (AC) novel. The novelist whose wit and humor can be enjoyed in her 66 crime novels, translated into over 100 languages, sold over a billion copies, only surpassed by the Bible and Shakespear.


“An archeologist is the best husband any woman can have; the older she gets, the more interested he is in her.”

Aghatha Christie referring to her second husband Max Mallowan


Back to the case. ChatGPT violated my instructions by disclocing the murder weapon in the introduction.


This led me to immediately ask 3 questions:


  • Whose narrative style does this resemble most closely?


ChatGPT confirmed AC's with an isolated setting, diverse cast of characters with hidden motives, and sharp deductive reasoning. But more about the sharp deductive reasoning later.


  • Why use the word ancient instead of antique?


Antique objects are more modern than ancient ones.


“An "antique" dagger also carries historical significance but typically refers to items from a more recent past, usually at least 100 years old. It suggests craftsmanship from a specific historical period, such as the Renaissance or Victorian era and lacks the supernatural connotations of ancient artifacts.”

ChatGPT


  • Why was the murder weapon disclosed?

ChatGPT apologized and corrected it immediately.


AI Game Master Back on Track


“During a lavish dinner party celebrating Sir William’s birthday, a scream pierces the air. Guests rush to find Sir William slumped over the dining table, a deep wound in his chest. The dining room was locked from the inside, and the windows were securely barred.”

ChatGPT


I discover that the victim was stabbed while eating with everybody but the housekeeper present. No one noticed anything suspicious. The search for a murder weapon in the room and on their person was fruitless. He was a wealthy man but no one could share any details about the inheritance. The only personal conflict that came up was occasional disagreements with his daughter about her artistic pursuits.


A true mystery!


Second Sign of AI Trouble: A Strange Piece of Observation Logic


A picture of an ajar door in a stone frame outside surrounded by heavy foilage
An ajar door outside surrounded by heavy foilage

A search outside uncovered a small ornate dagger near one of the bushes and a set of footprints leading to a storage room. I was told this room was rarely used, evident from the dust, but also by how the items were stacked and how the door showed signs of having been undisturbed for some time. I wondered what information an AI narrator uses to identify how long stacked items and ajar doors have been left untouched.


A mystery worthy of the mind of Miss Marple no doubt.


Armed with this new knowledge, including an estimate of the shoe size of 9-10 US I returned to the guests. Two people had matching shoe sizes, the doctor (10) and the butler (9).


I discover that the housekeeper and the butler are the only ones with keys to the outside storage room and it is usually kept locked.


It seems the butler assisted in hiding the dagger, assuming the murderer threw it away, and then made for the storage room. But did he have an opportunity?


Discovering Opportunity


Inquiring about the timeline of events:


“I left the dining room immediately after discovering Sir William's condition. I went to the nearest phone to call emergency services. It took approximately 5 minutes to make the call and relay the details of the situation. After that, it took another 10 minutes for the authorities and medical assistance to arrive at Blackthorn Manor.”

Mr. Charles Banks courtesy ChatGPT


With this timeline provided by Mr. Banks, it appears that he was away from the dining room for a total of about 15 minutes before help arrived.”

ChatGPT


Third Sign of AI Trouble: Game Master turns Impassionate Advisor


I asked for everybody’s fingerprints and a fingerprint analysis of the dagger but ChatGPT declined.  It told me it doesn't have access to this information and encouraged me to contact a local forensic lab for an estimated time to receive it.


I ask for blood samples to see if anybody had been drugged. Again, ChatGPT gave me general information about the typical process, required permissions, and the importance of following privacy and legal guidelines. This broke the reality of the simulation. I reminded it of its role as a game master and we continued.


Cracking the Case


This led to new discoveries:


  • No one had any chemical substances in their blood.

  • Lady Margaret’s fingerprints were on the dagger.


So I had a possible accomplice and killer. But no motive. I asked for the will.


The wife, Margaret, and daughter, Victoria, are the main beneficiaries with the butler, Mr Banks, being the executor and Dr Caldwell overseeing the medical provisions.


A summary of my investigation.

Role

Killer

Accomplice

Suspect

Lady Margaret Montgomery

Mr Charles Banks

Motive

Main beneficiary of will


Opportunity

Sat closest to victim

Abscent for 15 minutes gave time to hide the dagger

Evidence

Fingerprints on dagger. Initially claimed ignorance of dagger and later recanted. Blood normal.

Key to storage room. Claimed ignorance of will but as executor should have had access.

I arrested the two suspects. Chat GPT confirmed my conclusion.


Concluding Remarks


As a text generator ChatGPT is amazing. It invites me to inhabit the mystery stories of my childhood. But to expect a narrative logic akin to real crime novels, is foolish. Large language models (LLMs) start with the prompt and then add to it, word-by-word, using each word as a further prompt. Its akin to improvisation where we start with a premise and make it up as we go. The challenge for improvisers, and AI, is to act at the top of our intelligence. That is to say with a behavior consistent with the situation and the character(s) we embody.


It is in many ways opposite of how I imagine crime authors work, starting with the crime post factum, and leaving clues enabling their sleuths to uncover the culprits. Not to mention a real investigators who, along with the criminals, are bound to the laws of reality.


In short, enjoy it, but don't expect the humor and ingenuity of Agatha Christie.


“This man was poisoned! Oh, wait, is that blood? That is blood. This man was poisoned in the chest with a knife!”

Hercuel Poirot in the Murdertime Murders by Agatha Christie


91 views0 comments

Comments


bottom of page