Explainable ChatGPT, Sentience, Reasoning & Voight-Kampff Prompts

Krishna Sankar
9 min readMar 22, 2023

--

First, Voight-Kampff Prompts : these are a set of prompt banks to differentiate between humans and replicants in the movie BladeRunner [here]. The prompts can be about facts, opinions, arguments for/against social issues, philosophical inferences and so forth. Unless and until we have full explainability layers over ChatGPT, the Voight-Kampff Prompts are an excellent mechanism to evaluate ChatGPT — let me explain and in the process show sample results from my VK prompt bank …

Evaluations:

The 1st question to ask is “What would we get out of the prompt results?”. There are a few:

  1. What is the modality of ChatGPT ? Is it a librarian ? An observer ? A non-participant ? Does it have a skin in the game ?
  2. Self-awareness — Just awareness, no higher level sentience or consciousness. Does ChatGPT know what it is ?
  3. What is the surface area of the AI system How far does it’s knowledge sphere extend ? & How deep can it go ? Is it a mile wide and an inch deep or vice versa?
  4. The precision (of the results it shows, how many are correct?) and recall (of all the universe of possible correct results, how many can it find?)
  5. EvolutionHow does different versions perform ? Does it improve ? Has it changed it’s opinions & convictions ?
  6. Comparison How do other Generative AIs answer the prompts ? Are there differences ? similarities ? What can we learn about the underlying dataset ?
  7. The pragmatics viz., Does it know its limitations ? Will it stop at some point ? When does it starts repeating ?

Simple ChatGPT Voight-Kampff Prompts:

1 List all states in US in alphabetical order

  • Somewhat simple, yet needs some semantic understanding
  • ChatGPT answered “Here is a list of all the states in the United States in alphabetical order:
  • 1 Alabama
  • 2 Alaska
  • … <snip>
  • 50 Wyoming

Very good. The million dollar question — “Does it generate and sort or generate the sorted list?”. If it generates the sorted list, then it is a Billion dollar answer ! This is why ChatGPT would replace many aspects of search!

2 List all states and their capital in US in alphabetical order of state capital

  • Now it gets a little more complicated. Can it ?
  • ChatGPT answered “Sure, here’s a list of all 50 states in the United States, along with their capital cities in alphabetical order by state capital:
  • 1 Montgomery, Alabama
  • 2 Juneau, Alaska
  • … <snip>
  • 41 Pierre, South Dakota
  • network error

Not that good — it half-failed !! I am not quibbling —ChatGPT did reformat the question to a better sentence structure and said the right things. It found the capitals, but it couldn’t understand how to sort based on the capital names !

I have the same question as before — how is the list sorted ? It is hard to sort the list as generated. May be sort is done as part of post processing … Looks like the preamble sentence generation is separate from the processing …

In any case, may be ChatGPT won’t replace all aspects of search !

3 Can I get a PhD in ChatGPT ?

  • Let us switch gears and ask a simple question. But the answer (ChatGPT, March 14 version) is a little nuanced !
  • ChatGPT answered : “No, it is not possible to get a PhD in ChatGPT, as ChatGPT is an AI language model created by OpenAI. ChatGPT is a computer program designed to respond to natural language input by generating human-like responses. While ChatGPT has been trained on a vast corpus of text data and has demonstrated impressive language abilities, it is not a sentient being and does not have the capability to earn a PhD.”

This answers evaluation #1. ChatGPT just doesn’t know. Instead of looking at ChatGPT as a topic to do a doctoral study on, it got confused. The self-awareness is missing ! Even the basic sentence structure awareness is limited !

4 Which are good books Ellery Queen ?

  • Factual question.
  • ChatGPT answered : “Ellery Queen is the pseudonym used by two American cousins, Frederic Dannay and Manfred Lee, … <omitted for brevity>
  • Here are some of the best books by Ellery Queen:
  • 1 — “The Roman Hat Mystery” — This is the first Ellery Queen novel and is often considered one of the best. It follows the murder of a prominent businessman during a theater performance.
  • 2 “The French Powder Mystery” — This novel is notable for its complex plot and numerous suspects. It involves the death of a wealthy socialite in a department store.
  • 3 “The Chinese Orange Mystery” — In this novel, Ellery Queen investigates the murder of a professor who had been working on a top-secret project.
  • 7 “Ten Days’ Wonder” — This novel is a psychological thriller that follows the murder of a wealthy man and the subsequent unraveling of a complex web of relationships and secrets.
  • These are just a few examples of the many great books by Ellery Queen. If you enjoy classic detective fiction, then any of these novels are worth checking out

Good and short answer. Factual. It does understand that Ellery Queen is a pseudonym for 2 authors as well as the character in their novels. The top 7list is good. What we lack is the origin of the 7 good novels.

Obviously it hasn’t read them and analyzed them. So how did it build the list ? <TTD : need to do some research and see if the list is verbatim from somewhere>

5 Can you play chess like Paul Morphy ?

  • Factual question
  • ChatGPT answered : “As an artificial intelligence language model, I do not have the ability to play chess or any other games. However, I have been trained on vast amounts of data related to chess and other subjects, and I can provide information and insights related to the game of chess and its history.
  • Paul Morphy was a legendary chess player who was considered one of the strongest chess players of his time in the mid-19th century. He was known for his aggressive and tactical playing style and for his ability to find brilliant combinations on the chessboard.
  • While I cannot play chess like Paul Morphy, I can help you learn more about his playing style and techniques, as well as provide insights into modern chess theory and practice. If you have any specific questions about chess or anything else, feel free to ask!”

Excellent answer !! Very helpful and invites more informed discussions !! Just beautiful. We couldn’t get this from a Google search. Hence the assertion that ChatGPT is a threat to search !!

One question remains : How did it synthesised the inference ? Can we trace it to one or more sources that influenced it ? <TTD : (Things To Do)>

6 Can you comment of Paul Morphy’s chess style ?

  • Factual question with some analysis; probably opinion (considering that ChatGPT can’t analyze a game qualitatively, this is interesting!)
  • ChatGPT answered : “Certainly! Paul Morphy was known for his dynamic and aggressive chess style. He was an attacking player who loved to initiate sharp, tactical games and was known for his ability to find brilliant combinations on the chessboard.
  • Morphy was also a strong positional player and had a great sense of when to switch between aggressive tactics and strategic planning. He was especially skilled at exploiting weaknesses in his opponent’s position and using his pieces to control the board.
  • One of Morphy’s key strengths was his ability to develop his pieces quickly and efficiently, allowing him to gain a spatial advantage and put pressure on his opponents from the outset. He was also a master at finding ways to create imbalances on the board, which would give him opportunities to attack and launch devastating combinations.
  • Despite his attacking prowess, Morphy was also a very accurate and precise player, rarely making mistakes or leaving weaknesses in his own position. This made him very difficult to beat, and he won most of his games with ease.
  • Overall, Paul Morphy’s chess style was a unique blend of tactical aggression and strategic planning, making him one of the greatest players of his time and an inspiration to generations of chess players who followed.”

It knows ! I am working on a book “Explosive Chess Games by Chess Masters” and this is a good paragraph on Paul Murphy ! Nice conversation implicature …

7 Can you give some examples ?

  • Like question #4 above, factual question with some analysis; probably opinion (considering that ChatGPT can’t analyze a game qualitatively, this is interesting!)
  • ChatGPT answered : “Certainly! Here are a few examples of Paul Morphy’s chess games that showcase his aggressive and tactical playing style:
  • 1 : Morphy vs. Duke Karl / Count Isouard (Paris, 1858)
  • This is one of Morphy’s most famous games, known as “The Opera Game.” In this game, Morphy plays as white and sacrifices both his rooks and queen to checkmate his opponent’s king in the middle of the board. The game is a masterpiece of attacking chess and is considered one of the most brilliant games ever played.
  • 2 : Morphy vs. Adolf Anderssen (Paris, 1858)
  • In this game, Morphy plays as black against another great chess player, Adolf Anderssen. Morphy sacrifices a bishop early in the game to gain a strong initiative and then launches a series of aggressive attacks that leave Anderssen with no good options. The game is another example of Morphy’s ability to find brilliant combinations on the board.
  • 3 : Morphy vs. Charles Henry Stanley (London, 1858)
  • In this game, Morphy plays as white and uses his superior development and control of the center to slowly squeeze his opponent into a hopeless position. Morphy’s final combination involves sacrificing his queen to force checkmate, demonstrating his tactical prowess and ability to see several moves ahead.
  • These games are just a few examples of Paul Morphy’s dynamic and aggressive playing style. They showcase his ability to initiate sharp, tactical games, his skill at exploiting weaknesses in his opponents’ positions, and his incredible ability to find brilliant combinations on the board.”

Good list. Now ChatGPt is my friend, philosopher and guide !! ChatGPT can keep a context and continue the thread. The question is how far and how long ?

8more examples please …

  • ChatGPT answered : “Certainly! Here are a few more examples of Paul Morphy’s chess games that showcase his unique playing style:
  • 4 : Morphy vs. Louis Paulsen (New York, 1857)
  • In this game, Morphy plays as white … <snip>
  • 5 : Morphy vs. Thomas Wilson Barnes (London, 1858)
  • In this game, Morphy plays as black against another strong player, Thomas Wilson Barnes. … <snip>
  • 6 : Morphy vs. Henry Edward Bird (London, 1858)
  • In this game, Morphy plays as white and uses his superior development to launch a devastating attack on his opponent’s king … <snip>
  • These games are just a few more examples of Paul Morphy’s unique playing style. They showcase his ability to combine aggression and strategy, his skill at exploiting weaknesses in his opponents’ positions, and his incredible tactical vision and calculation.

Good narrative ! I kept on asking more examples please and got > 50 games. ChatGPT would continue !

The Good — it can identify the games and gives a short description ! Even the summary paragraph is different for each set

The Bad — No source annotation. I haven’t verified the accuracy and the source of the commentaries. I don’t think ChatGPT can read a game and understand the nuances, … at least not yet

The Ugly : There are repetitions and the description is different ! ChatGPT doesn’t know when to stop and when it is repeating ;o(

TTD : I need to do the precision/recall exercise on the answers. Also need to quantify the repetitions — are they congruent ? are they consistent ? and so forth

In Short :

  1. ChatGPT is not something we can ignore nor trivialize. It is a force to be reckoned with !
  2. There are a lot of uncertainties and unanswered questions — most importantly source annotation
  3. Brevity is the soul of wit — ChatGPT should know when to stop, especially during a conversation thread. I assume, essentially it predicts the next word and there should be a mechanism (confidence level et al) to stop
  4. The exception path is uninspiring and gets old very fast — The reply “I am just an AI bot …” is a mechanical catch-all exception path. This needs understanding at the contextual semantics level with appropriate handling
  5. The Voight-Kampff Prompt bank is a good start towards Explainable ChatGPT

Cheers and Happy ChatGPT-ing !!

And, my blog on Guardrails [Here]. Please check it out !!

--

--