How an AI answers a simple science question — step-by-step (Example: “Why did Pluto stop being a planet?”)

How an AI answers a simple science question step-by-step (Example: “Why did Pluto stop being a planet?”)

By way of a short demonstration, explained in approachable scientific style but focused on how a modern AI processes the question rather than on the astronomy alone.

When you ask a human “Why did Pluto stop being a planet?” you get an answer shaped by memory, textbooks, and a judgement about what level of detail you want. When you ask an AI, a similar goal exists  produce a correct, clear, and concise explanation  but the steps the machine takes look different. Below I walk through how a contemporary large language model (LLM) typically handles that single question: the modular stages it passes through, the computations involved at a conceptual level, and the way the final answer is assembled. 

1) The user query and intent detection

User input (example): “Why did Pluto stop being a planet?”

First the system interprets the text: this short sentence expresses an information-seeking intent (a factual explanation). Modern AI pipelines run lightweight analyses to classify intent, detect language, and identify named entities (“Pluto”) and temporal cues. This preliminary step decides which processing route to follow: a simple direct reply, a longer explanatory answer, or a retrieval-augmented search if up-to-date facts are needed.

Why it matters: correct intent detection prevents irrelevant content and helps the model choose the right depth and style.

2) Text normalization and tokenization

The input sentence is normalized (lowercased or not, punctuation handled) and broken into tokens the atomic pieces the model actually computes with (subwords, words, or bytes depending on architecture).

Example simplified tokenization:

• “Why” → token₁

• “did” → token₂

• “Pluto” → token₃

• “stop” → token₄

• “being” → token₅

• “a” → token₆

• “planet” → token₇

• “?” → token₈

These tokens become the numerical input the network processes (vectors of numbers).

Why it matters: tokens are how language becomes math. Tokenization shapes which document fragments the model can match and how rare words are represented.

 

3) Context encoding and embeddings

Each token is mapped to a high-dimensional vector (an embedding) that captures aspects of its meaning learned from huge amounts of text. The model then encodes the full question into a contextual representation using layers of computation (often transformer layers). These layers compute how each token relates to every other token (the “attention” mechanism).

At the end of this stage the model holds an internal representation that encodes the question’s semantics: that the user asks for an explanation about Pluto’s change of status.

Why it matters: embeddings let the model find and combine relevant facts and patterns it learned during training.

 

4) Retrieval (optional, but common in high-accuracy systems)

Many production systems add a retrieval component: they query an external knowledge store (articles, textbooks, official resolutions) using the question’s embedding to fetch relevant documents. In our example the retrieval step would likely return the International Astronomical Union’s (IAU) 2006 resolution and explanatory articles about Pluto, Kuiper belt objects, and the historical discovery of Pluto.

Retrieval keeps the answer anchored to factual sources and lets the model cite up-to-date or authoritative material even if the base model’s training cut-off didn’t include those specific documents.

Why it matters: retrieval reduces hallucinations and improves accuracy for factual questions.

 

5) Reasoning and planning a response

Given the encoded question and any retrieved documents, the model prepares an output plan: decide the main points to cover (a short chronology, the IAU definition, why Pluto no longer fits, and a brief note on remaining controversies), the order to present them, and the tone (concise but explanatory).

Modern models don’t write a literal plan in human words; rather, the planning emerges from the next stage’s computations that predict the most useful next token sequences conditioned on the input and any retrieved evidence.

Why it matters: a clear plan helps the answer be coherent and focused.

---

6) Generation: producing the answer

The model now generates text token by token, each new token chosen probabilistically from the model’s prediction of what comes next given the context, the retrieved snippets, and the desired style. Constraints (length limits, safety filters, or format instructions) influence choices.

A safe, public system also routes the output through content filters that block harmful or disallowed content. For factual queries this usually does not intervene, but the filter ensures no inappropriate language or misinformation patterns are produced.

Sample AI answer (what the model might produce):

Pluto was reclassified in 2006 because the International Astronomical Union (IAU) formalized what counts as a “planet.” The IAU definition requires three things: (1) the body orbits the Sun, (2) it has enough mass for its own gravity to make it nearly round, and (3) it has cleared its orbital neighborhood of other debris. Pluto meets the first two criteria but not the third: it shares its region of the solar system with other objects in the Kuiper belt and has not cleared smaller bodies from its orbit. As a result, Pluto was designated a “dwarf planet,” a category for round objects that orbit the Sun but have not cleared their neighborhoods. This change reflects a refinement of terminology driven by discoveries of many similar small worlds beyond Neptune and a desire for a clear, usable classification scheme.

That output is assembled token by token, guided by attention to the question and, if used, retrieved IAU text.

7) Post-processing, citations and user framing

If retrieval was used, the system can append citations or suggest further reading. It may also produce a short glossary for unfamiliar terms (“cleared its orbit,” “Kuiper belt,” “dwarf planet”). If the user asked follow-ups, the system can adapt style and depth.

Why it matters: citing sources and explaining jargon make the answer verifiable and educational.

 

8) Safety, uncertainty, and disclaimers

A good AI indicates uncertainty when appropriate (“Some scientists still debate wording and classifications”) and avoids inventing precise numeric facts unless confident. For historical, consensus decisions like the IAU reclassification, the answer is firm; for unsettled science the model will hedge and recommend primary sources.

 

9) An example interaction flow (user → AI)

1. User: “Why did Pluto stop being a planet?”

2. AI detects intent: factual explanation.

3. AI tokenizes and encodes the query.

4. AI retrieves authoritative sources (IAU 2006) — if available.

5. AI plans: history → definition → Pluto’s status → consequences.

6. AI generates the explanatory paragraph(s).

7. AI supplies a short glossary and suggests sources for deeper reading.

   
   

Takeaway: what this tells us about AI explanations

An AI’s answer isn’t a single “reason” hidden in a black box; it’s the result of layered processes  interpreting language, mapping meaning to vectors, retrieving pertinent documents, and generating fluent text under safety and clarity constraints. For simple historical-factual questions like Pluto’s status, the process looks efficient and deterministic to a user: a short question yields a compact, sourced explanation. For open-ended or controversial questions, the same machinery will produce longer, more cautious answers and may offer multiple viewpoints.

Glossary

• Tokenization: splitting text into the atomic units the model processes.

• Embedding: a numeric vector that represents a token or sentence’s meaning.

• Attention: the mechanism allowing the model to weight relationships between tokens.

• Retrieval: fetching external documents to ground answers in facts.

• IAU (International Astronomical Union): the body that set the 2006 planet definition.

• Kuiper belt: a region beyond Neptune populated by small icy bodies.

---

Further reading (suggested)

• Soter, S. (2006). What is a Planet? The Astronomical Journal.

• Margot, J.-L. (2015). A Quantitative Criterion for Defining Planets.

• IAU Resolutions B5/B6 (2006).

• Brown, M. (2008). The Kuiper Belt and the Demotion of Pluto.

---

This short walkthrough shows both the astronomy and the plumbing: a straightforward factual question lets us see an AI’s typical pipeline in miniature  from token to explanation  and illustrates why retrieval and careful framing matter for accuracy and trust.