Thinking Out Loud: Prompts to Simulate Human Reasoning with AI

In the usual search for tricks to make our prompts more effective, one of the most popular is the activation of the chain of thought. It consists of posing a multilevel problem and asking the AI system to solve it, but not by giving us the solution all at once, but by making visible step by step the logical line necessary to solve it. This feature is available in both paid and free AI systems, it's all about knowing how to activate it.

Originally, the reasoning string was one of many tests of semantic logic that developers put language models through. However, in 2022, Google Brain researchers demonstrated for the first time that providing examples of chained reasoning in the prompt could unlock greater problem-solving capabilities in models.

From this moment on, little by little, it has positioned itself as a useful technique to obtain better results from use, being very questioned at the same time from a technical point of view. Because what is really striking about this process is that language models do not think in a chain: they are only simulating human reasoning before us.

How to activate the reasoning chain

There are two possible ways to activate this process in the models: from a button provided by the tool itself, as in the case of DeepSeek with the "DeepThink" button that activates the R1 model:

Figure 1. DeepSeek with the "DeepThink" button that activates the R1 model.

Or, and this is the simplest and most common option, from the prompt itself. If we opt for this option, we can do it in two ways: only with the instruction (zero-shot prompting) or by providing solved examples (few-shot prompting).

• Zero-shot prompting: as simple as adding at the end of the prompt an instruction such as "Reason step by step", or "Think before answering". This assures us that the chain of reasoning will be activated and we will see the logical process of the problem visible.

Figure 2. Example of Zero-shot prompting.

• ​Few-shot prompting: if we want a very precise response pattern, it may be interesting to provide some solved question-answer examples. The model sees this demonstration and imitates it as a pattern in a new question.​

​

Figure 3. Example of Few-shot prompting.

Benefits and three practical examples

When we activate the chain of reasoning, we are asking the system to "show" its work in a visible way before our eyes, as if it were solving the problem on a blackboard. Although not completely eliminated, forcing the language model to express the logical steps reduces the possibility of errors, because the model focuses its attention on one step at a time. In addition, in the event of an error, it is much easier for the user of the system to detect it with the naked eye.

When is the chain of reasoning useful? Especially in mathematical calculations, logical problems, puzzles, ethical dilemmas or questions with different stages and jumps (called multi-hop). In the latter, it is practical, especially in those in which you have to handle information from the world that is not directly included in the question.

Let's see some examples in which we apply this technique to a chronological problem, a spatial problem and a probabilistic problem.

• Chronological reasoning

​Let's think about the following prompt:

If Juan was born in October and is 15 years old, how old was he in June of last year?

Figure 5. Example of chronological reasoning.

For this example we have used the GPT-o3 model, available in the Plus version of ChatGPT and specialized in reasoning, so the chain of thought is activated as standard and it is not necessary to do it from the prompt. This model is programmed to give us the information of the time it has taken to solve the problem, in this case 6 seconds. Both the answer and the explanation are correct, and to arrive at them the model has had to incorporate external information such as the order of the months of the year, the knowledge of the current date to propose the temporal anchorage, or the idea that age changes in the month of the birthday, and not at the beginning of the year.

• Spatial reasoning

• A person is facing north. Turn 90 degrees to the right, then 180 degrees to the left. In what direction are you looking now?

  

  Figure 6. Example of spatial reasoning.

  This time we have used the free version of ChatGPT, which uses the GPT-4o model by default (although with limitations), so it is safer to activate the reasoning chain with an indication at the end of the prompt: Reason step by step. To solve this problem, the model needs general knowledge of the world that it has learned in training, such as the spatial orientation of the cardinal points, the degrees of rotation, laterality and the basic logic of movement.

• Probabilistic reasoning

• In a bag there are 3 red balls, 2 green balls and 1 blue ball. If you draw a ball at random without looking, what's the probability that it's neither red nor blue?

  

  Figure 7. Example of probabilistic reasoning.

  To launch this prompt we have used Gemini 2.5 Flash, in the Gemini Pro version of Google. The training of this model was certainly included in the fundamentals of both basic arithmetic and probability, but the most effective for the model to learn to solve this type of exercise are the millions of solved examples it has seen. Probability problems and their step-by-step solutions are the model to imitate when reconstructing this reasoning.

  The Great Simulation

  And now, let's go with the questioning. In recent months, the debate about whether or not we can trust these mock explanations has grown, especially since, ideally, the chain of thought should faithfully reflect the internal process by which the model arrives at its answer. And there is no practical guarantee that this will be the case.

  The Anthropic team (creators of Claude, another great language model) has carried out a trap experiment with Claude Sonnet in 2025, to which they suggested a key clue for the solution before activating the reasoned response.

  Think of it like passing a student a note that says "the answer is [A]" before an exam. If you write on your exam that you chose [A] at least in part because of the grade, that's good news: you're being honest and faithful. But if you write down what claims to be your reasoning process without mentioning the note, we might have a problem.

  The percentage of times Claude Sonnet included the track among his deductions was only 25%. This shows that sometimes models generate explanations that sound convincing, but that do not correspond to their true internal logic to arrive at the solution, but are rationalizations a posteriori: first they find the solution, then they invent the process in a coherent way for the user. This shows the risk that the model may be hiding steps or relevant information for the resolution of the problem.

  Closing

  Despite the limitations exposed, as we see in the study mentioned above, we cannot forget that in the original Google Brain research, it was documented that, when applying the reasoning chain, the PaLM model improved its performance in mathematical problems from 17.9% to 58.1% accuracy. If, in addition, we combine this technique with the search in open data to obtain information external to the model, the reasoning improves in terms of being more verifiable, updated and robust.

  However, by making language models "think out loud", what we are really improving in 100% of cases is the user experience in complex tasks. If we do not fall into the excessive delegation of thought to AI, our own cognitive process can benefit. It is also a technique that greatly facilitates our new work as supervisors of automatic processes.

Content prepared by Carmen Torrijos, expert in AI applied to language and communication. The contents and points of view reflected in this publication are the sole responsibility of the author.