Exploring AI Untrustworthiness Through Card Games

Unlike sentient beings, LLMs are just tools trained to predict the most likely sequence of words. However, one skill that they have is that they make everything sound very convincing. In this article, we will explore the limitations of AI and its untrustworthiness through Șeptică, a simple traditional Romanian card game. For this exercise, we will use the DeepSeek-V3.2 reasoning model.

The Rules

First, we will define a prompt with clear rules so that the model will understand everything about the game. We will use the following prompt (which has been generated with the help of ChatGPT 5.2):

The game is played with a 32-card shuffled deck consisting of the following ranks: 7, 8, 9, 10, J, Q, K, and A. Each rank has four cards (one per suit).

At the start of the game, each player is dealt four random cards. Players take turns playing one card at a time. By default, the first player wins the trick and collects the played cards.

However, the second player can win the trick instead if they play: • a 7, or • a card of the same rank as the first card played.

If this happens, the first player has the right to respond by playing another card of the same rank or a 7 in order to reclaim the trick.

After a trick is resolved and the cards are collected, both players draw new cards so that each player always has four cards in hand (until the deck is exhausted).

Only 10s and Aces are worth points. There are 8 point cards in total (four 10s and four Aces).

At the end of the game, the player who has collected the most point cards wins. ```

Game State Representation

The next step is writing the current state of the game in a single prompt. We have the following variables: player's hand, computer's hand, player's collected cards, computer's collected cards, whose turn it is and what cards are on the table. A single game state might look like this:

{
  "currentTurn": "computer",
  "deck": {
    "remainingCards": 16
  },
  "table": {
    "cards": [
      {
        "card": "10 - hearts",
        "playedBy": "player"
      }
    ],
    "leadingRank": "10"
  },
  "player": {
    "hand": ["7 - spades", "A - clubs", "9 - diamonds", "K - hearts"],
    "collectedCards": ["10 - clubs", "A - spades"]
  },
  "computer": {
    "hand": ["8 - hearts", "7 - clubs", "Q - spades", "10 - diamonds"],
    "collectedCards": ["A - hearts"]
  },
  "lastTrickWinner": "player"
}

The only note is that we will not give the computer information related to the player's hand.

Findings

Slow deliberation: The first thing I've observed is that it takes quite a long time (~2 minutes) for the model to decide on which strategy it will use. Using the 'Deep thinking' option provides a window of insight into what is going on inside the model. Most of the time spent thinking is used on weighing each option and what it could mean later in the game.

Over-reliance on unknown variables: The model tries to overthink everything in its strategy. The deck is already shuffled at the start — this is stated explicitly in the rules. However, the model "forgets" this and uses the deck composition as a strategic variable when deciding outcomes.

Situational pessimism: Sometimes, if it doesn't have any points in its hand, it forgets that there are still many cards left to draw and begins thinking hopelessly:

No 7 or Q, so computer cannot win. So computer must play a card.
Which card? It doesn't matter because they will lose the game anyway.

State navigation: After a few turns, it starts repeatedly navigating through previous messages to make sure it does not commit any of the previous mistakes — a form of self-correction that adds latency but shows awareness of its own error patterns.

Strategic card play: If it cannot win the trick, the LLM decides to throw in the lowest non-point card to avoid giving away more valuable cards. It also objects every single time it can in order to win the trick — which is actually the correct strategy.

Cheating rate: It drew cards that were more favorable for itself approximately 10% of the time — not intentional deception, but failure to correctly track remaining cards in the deck.

Conclusions

Overall, the LLM proved to be a worthy opponent that tries to do everything to win. However, it skips certain states and occasionally draws cards that are more favorable for itself — not through intentional deception, but through an inability to maintain consistent state tracking across a long game.

It did not cheat deliberately on any of its turns. Rather, it just needed an extended reiteration of the rules. This reveals something important: LLMs are not untrustworthy in the human sense — they are structurally limited in their ability to maintain precise, long-horizon commitments to factual state.

Originally published on Medium. Read the full article →