Guess in 20: how to play and evaluate answers

How to start the game and what counts as a result

The tool is a web form with a text field for entering instructions. You select the output language, press Generate and get one result in a separate panel. In the game 'Guess in 20' the AI chooses a common object, and you ask questions that can be answered only 'yes', 'no' or 'sometimes'. The goal is to guess the object in twenty questions or fewer. The result is either the guessed object or running out of attempts.

The quality of the result is evaluated according to three criteria: accuracy (AI answers correspond to the chosen object), completeness (the game ends successfully or gives enough information for a guess) and suitability (answers are logical, do not contradict each other and comply with the rules). It is important to understand that AI may make mistakes, and the queue and availability may vary.

Example: you start the game with the phrase 'Choose a common object. I will ask up to twenty questions with yes or no answers'. The AI answers: 'Okay, I chose an object'. Then you ask: 'Is this object bigger than a palm?' — 'Yes'. 'Is it used in the kitchen?' — 'Sometimes'. And so on.

Choose a common object. I will ask up to twenty questions with yes or no answers.

Checking answer accuracy

Accuracy is the key criterion. If the AI chose, for example, 'apple', and on the question 'Is it edible?' answers 'No', that is an error. To check accuracy, you need to ask after guessing or at the end of the game: 'What object did you choose?' and compare with the answers. If the answers contradict the object, accuracy is low.

In our example: the chosen object is 'umbrella'. Questions: 'Is it bigger than a palm?' — 'Yes'. 'Is it used outdoors?' — 'Yes'. 'Is it needed for rain protection?' — 'Yes'. 'Is it an umbrella?' — 'Yes'. All answers are accurate. If the AI on the question 'Is it needed for rain protection?' answered 'No', that would be an error.

You should bear in mind that AI may answer 'sometimes' to questions where the answer is ambiguous. For example, 'Is it used in the kitchen?' for an umbrella — 'Sometimes' (if the umbrella is stored in the kitchen). This is permissible, but may confuse. Check how much the answers correspond to reality.

Completeness: enough information for a guess

Completeness means that within twenty questions the object can be guessed with sufficient probability. If the AI gives too general answers or goes off topic, the game becomes uninteresting. Evaluate how many questions were needed to guess. If you guessed in 10–15 questions, completeness is good. If after 20 you did not guess — the object may have been too difficult or the answers not informative enough.

In the umbrella example: the questions covered size, place of use, function. This was enough. If the AI on the question 'Is it used outdoors?' answered 'Sometimes', and on 'Is it needed for rain protection?' — 'No', it would be harder to guess. Completeness suffers if the AI often answers 'sometimes' unnecessarily.

Check that answers do not contradict each other. For example, 'Is it bigger than a palm?' — 'Yes', 'Can it be put in a pocket?' — 'Yes' may be contradictory for an umbrella. Such inconsistencies reduce completeness.

Suitability: logic and rule compliance

Suitability is how logical the answers are and how well they comply with the game rules. The AI must answer only 'yes', 'no' or 'sometimes'. If it gives expanded answers or asks counter-questions, that is a violation. In our tool answers must be brief. Check that the AI does not go beyond the limits.

It is also important that the AI does not change the chosen object during the game. If you asked 'Is it edible?' — 'No', and then 'Is it an apple?' — 'Yes', that is an explicit contradiction. Suitability is low. In the umbrella example all answers are logical and consistent.

Another aspect — the language. You can choose the output language. Make sure the AI answers in the chosen language. If you chose Russian, but answers are in English, this reduces suitability.

Specific example of evaluation

Consider a fictional example to illustrate. You start the game with the instruction: 'Choose a common object. I will ask up to twenty questions with yes or no answers'. The AI answers: 'Object chosen'. You ask questions: 1. 'Is it bigger than a palm?' — 'Yes'. 2. 'Is it used indoors?' — 'No'. 3. 'Is it used outdoors?' — 'Yes'. 4. 'Is it needed for sun protection?' — 'Sometimes'. 5. 'Is it needed for rain protection?' — 'Yes'. 6. 'Is it an umbrella?' — 'Yes'. The game ends in 6 questions.

Evaluation: accuracy — all answers correspond to the umbrella (high). Completeness — information is sufficient for a quick guess (high). Suitability — answers are brief, logical, in Russian (high). However the 'sometimes' answer on the sun protection question may be debatable: an umbrella can protect from the sun, but usually umbrellas are associated with rain. This is permissible, but reduces unambiguity.

If the AI on the question 'Is it needed for rain protection?' answered 'No', and then you guessed umbrella, accuracy would be low. If on 'Is it bigger than a palm?' it answered 'No', that would also be an error. Therefore always check answers for compliance.

Question: 'Is it bigger than a palm?' — Answer: 'Yes'. Question: 'Is it used outdoors?' — 'Yes'. Question: 'Is it needed for rain protection?' — 'Yes'. Question: 'Is it an umbrella?' — 'Yes'.

Tool limitations

The tool has limitations. Input in the first field — up to 8000 characters, in the details field — up to 1000. For speech — up to 3000 characters, for images — up to 4 MB, for audio — up to 20 MB or 5 minutes. Model answers may contain errors, and the queue and availability may vary. No batch processing, automatic saving, integrations, OCR, settings panels, DOCX or PDF export. Only copying or downloading simple text TXT. Vectors only SVG and PNG. Speech eSpeak WAV. Subtitles SRT. No photorealistic image generation, audio or video generation, voice cloning, premium plan, registration.

In the game 'Guess in 20' the AI may sometimes give inaccurate answers or get confused if the object is difficult. It does not guarantee that you will guess in twenty questions. Also it is not intended for professional tasks — it is an entertainment tool.

Important: do not send audio to the AI if it is transcribed locally. Images are vector compositions, not photorealistic generation. Do not expect the tool to replace live communication.

How to improve game quality

To improve quality, formulate questions clearly and consistently. Start with general categories: 'Is it living?', 'Is it edible?', 'Is it used at home?'. Then move to more specific. Avoid ambiguous questions, to which the AI may answer 'sometimes' without benefit.

If the AI gives contradictory answers, try rephrasing the question. For example, instead of 'Is it used in the kitchen?' ask 'Is it kitchenware?'. This will help get a more accurate answer. You can also ask for a hint if the game rules allow.

Remember that the AI is not perfect. If answers seem strange, the object may have been chosen poorly. In such a case start a new game. Evaluating quality is a subjective process, but following the criteria of accuracy, completeness and suitability you can objectively assess the result.