Embeddings Explained: How AI Turns Words Into Coordinates

An embedding is a list of numbers that gives a word, sentence, or document a position in a space where distance means difference in meaning. That’s it. “Dog” becomes something like [0.21, -0.88, 0.04, …] with a few hundred or a few thousand entries, and “puppy” lands nearby while “algebra” lands far away.

This is the quietest important idea in AI, and it’s also the most teachable. Your kid already understands coordinates from a map. Embeddings are a map with hundreds of directions instead of two. And when Gemini in Google Classroom, live since August 10, 2026, builds flashcards from a teacher’s uploaded materials rather than from the internet, embeddings are how it finds the right passage.

Key Takeaways

• An embedding turns text into a vector: a fixed-length list of numbers. Similar meanings get nearby vectors, which is the entire point.
• Mikolov et al.’s 2013 word2vec paper showed that these vectors capture relationships well enough to do arithmetic with meaning, learning “high quality word vectors from a 1.6 billion words data set” in under a day.
• Similarity is measured by distance or angle between vectors, not by shared letters. “Car” and “automobile” are close; “car” and “card” are not.
• Embeddings are the search step in every “chat with your documents” feature, including study guides built from class materials.
• The failure mode to teach: embeddings capture statistical association, which means they also capture bias present in the training text.

What a vector actually is, and why numbers beat words

Start with the problem. A computer needs to compare meaning, and letters don’t help. “Car” and “card” share three characters and mean nothing alike. “Car” and “automobile” share one and mean the same thing.

So the trick is to represent each word as a point in space, positioned so that position encodes meaning. Instead of two coordinates like a map, use several hundred. Each dimension is a learned direction that captures some aspect of how the word is used, and none of them has a human name.

How does the model learn where to put each word? By looking at context. Words that appear in similar surroundings get similar positions. “The ___ barked” and “the ___ fetched the ball” both tend to have “dog” in the blank, so “dog” ends up near other words that fit those slots.

The foundational paper is Mikolov, Chen, Corrado, and Dean, “Efficient Estimation of Word Representations in Vector Space”, 2013. It introduced two architectures, CBOW and Skip-gram, and the reason it mattered was efficiency: it learned “high quality word vectors from a 1.6 billion words data set” in less than a day, which put embeddings within reach of anyone with a computer.

The result that made it famous is the arithmetic. Because relationships are encoded as directions in the space, you can add and subtract vectors. Take the vector for “king,” subtract “man,” add “woman,” and the closest word to the result is “queen.” The direction from “man” to “woman” turns out to be roughly the same direction as from “king” to “queen,” and from “uncle” to “aunt.” Nobody programmed that. It emerged from counting which words appear near which other words.

The reason modern embeddings can tell “river bank” from “bank account” is the attention mechanism introduced in Vaswani et al.’s 2017 paper. Modern systems have moved past single-word vectors to contextual and sentence-level embeddings, where “bank” gets a different vector in “river bank” than in “bank account,” and where a whole paragraph gets one vector. The core idea is unchanged: meaning becomes position, and similarity becomes distance.

Where your kid meets embeddings without knowing it

Every feature that searches your own material uses embeddings. The pattern is always the same three steps.

Step one: index. Split the source material into chunks and compute an embedding for each. Store them.

Step two: query. Compute an embedding for the student’s question.

Step three: retrieve. Find the stored chunks whose vectors are closest to the question’s vector, and hand those to the model as context. Lewis et al.’s 2020 NeurIPS paper formalized this pattern as retrieval-augmented generation, pairing a “dense vector index of Wikipedia” with a text generator.

That’s how Gemini in Google Classroom generates flashcards and practice quizzes from a specific assignment. The August 4, 2026 announcement says the feature incorporates “the relevant title, assignment instructions, and curriculum materials as context directly from Google Classroom,” and includes syncing to Gemini Notebook for study guides and audio overviews. It went live on the web August 10 and on mobile August 17. That grounding, class materials rather than the whole internet, runs on embedding similarity. Our companion piece on retrieval-augmented generation walks through the full pipeline.

Google’s broader August 2026 roundup lists the same grounding pattern in SAT prep and study tools. Same mechanism, other places your kid already touches: the “related videos” list on a learning platform, plagiarism detection that catches paraphrases, a search bar that finds “how do I calculate area” when the document says “computing surface measurements,” and spam filters that recognize a scam written in new words.

The important consequence for a student: retrieval by meaning means bad chunking produces bad answers. If a teacher’s PDF is split mid-sentence, the vector for that chunk represents half an idea, and the retrieved context will be subtly wrong. When a study tool gives a confidently incomplete answer, this is often why.

What embeddings get wrong

Embeddings learn from text, so they learn what the text contains, including things nobody wanted them to learn.

The canonical demonstration used the same arithmetic that made word2vec famous. If “king minus man plus woman” gives “queen,” then analogies involving occupations can produce results that reflect stereotypes present in the training corpus rather than anything about the world. The mechanism is neutral. The data is not.

Two more limitations worth knowing:

Similarity isn’t truth. Two passages can be close in embedding space and contradict each other, because they’re about the same topic. Retrieval finds relevant text, not correct text.

Negation is hard. “The experiment worked” and “the experiment did not work” are close in embedding space, because they share nearly all their words and topic. That’s a real and persistent weakness, and it matters for anything where a kid is trying to find whether a source supports or refutes a claim.

Those two together explain a class of AI study-tool errors that look like hallucination but aren’t. The model retrieved something relevant and then reasoned from it. The retrieval wasn’t wrong exactly; it was insufficiently precise.

How to Teach Your Kid About Embeddings

Embeddings are coordinates. Every kid who’s played Battleship or read a map already has the concept; you’re adding dimensions and meaning.

Ages 5–8: The floor map of words

Clear a space on the floor. Put a paper labeled “animals” in one corner and “food” in the opposite corner. Hand your kid word cards (dog, cat, pizza, apple, horse, bread, tiger, cake) and have them place each card closer to whichever corner fits. Now add a third label, “big,” on a chair, so height becomes a dimension. Ask where “elephant” goes. Name it: “You just gave every word a place. Computers do this with hundreds of directions instead of three, and that’s how they know which words mean similar things.”

Ages 9–12: Two-dimensional word math

Draw an x-y grid on graph paper. Put “man” at (2,2), “woman” at (2,6), “king” at (8,2). Ask your kid where “queen” should go. Most kids say (8,6) immediately, because they see the pattern. Then tell them that’s exactly what happens in real embeddings: the direction from “man” to “woman” is the same direction as from “king” to “queen,” which Mikolov’s team published in 2013. Then try a few more (uncle/aunt, boy/girl) and one that breaks (car/cars vs. child/children) and talk about why irregular cases are harder.

Ages 13+: Rank by similarity, then check the machine

Have your teen write one target sentence and ten candidate sentences, some on-topic with different words, some off-topic with shared words. Their job: rank the ten by meaning-similarity to the target, by hand. Then paste the same eleven into an AI tool and ask it to rank by semantic similarity. Compare the two rankings. Where they disagree is the interesting part, and negation cases (“the treatment helped” vs. “the treatment did not help”) will usually be where the machine looks worst.

The question to ask: “How would you explain to a computer that ‘happy’ and ‘glad’ mean the same thing, if it can’t understand words?”

Word to vector: a worked demonstration

Word or phrase	Simplified 3-dimensional position (living, size, edible)	Nearest neighbor in this toy space	Why
dog	(0.9, 0.4, 0.0)	cat	Both living, small-to-medium, not food
cat	(0.9, 0.3, 0.0)	dog	Same region
elephant	(0.9, 1.0, 0.0)	horse	Living and large
pizza	(0.0, 0.3, 1.0)	bread	Non-living and edible
bread	(0.0, 0.2, 1.0)	pizza	Same region
car	(0.0, 0.7, 0.0)	truck	Non-living, large, not edible
card	(0.0, 0.0, 0.0)	paper	Shares letters with “car” and nothing else

A real embedding has 300 to 3,000+ dimensions instead of 3, and no dimension has a human-readable label like “edible.” This table is a teaching simplification, not a real vector. The point it makes accurately: “car” and “card” end up far apart despite sharing letters, and “pizza” and “bread” end up together despite sharing none.

What to do at home

Teach the word “similar,” not “same”

Embeddings find similar. That’s different from finding correct, and different from finding the answer. A kid who says “the tool found something similar, let me check whether it’s right” has internalized the right model of the machine.

Test the negation trap once

Have your kid ask a study tool whether a source supports a claim, using a source that explicitly contradicts it. Watch whether the tool notices. Often it retrieves the relevant passage and then handles the negation correctly, but not always, and seeing the failure once makes the lesson permanent.

Point at the grounding

If your kid’s school tool builds quizzes from class materials, tell them the tool is searching the teacher’s documents by meaning, not by keyword. That changes how they phrase questions: describing the concept works better than guessing the exact wording.

Notice the bias question early

“Where did these numbers come from?” leads directly to “they came from counting how people actually write,” which leads to why stereotypes show up in outputs. It’s a natural, non-preachy entry point, and it connects to the broader picture in when AI is wrong.

What not to do

Don’t tell your kid the computer “understands” the words. It has positions, and it computes distances. Whether that constitutes understanding is a genuinely open philosophical question, and the honest answer (“it behaves as if it does, using geometry”) is more interesting than either extreme. Our guide to how large language models work sits one level up from this.

What to Watch For Over the Next 3 Months

• Week 4: Your kid can say that words become lists of numbers and that closer means more similar in meaning.
• Month 2 red flags: They think the AI is looking up keywords. Or they treat “the tool found this passage” as “the tool proved this point.”
• Month 3 self-check: Ask them to predict which of three phrasings will retrieve the right section of a document. If they reason about meaning rather than matching words, they’ve got it.

Frequently Asked Questions

What exactly is an embedding?

A fixed-length list of numbers representing a word, sentence, or document, positioned in a space where nearby vectors mean similar content. Typical sizes run from a few hundred to a few thousand numbers, and the individual numbers have no human-readable meaning.

Does “king minus man plus woman equals queen” really work?

It works well enough to be a famous demonstration, published by Mikolov, Chen, Corrado, and Dean in 2013, and it’s not a parlor trick: relationships genuinely are encoded as consistent directions. It doesn’t work on every analogy, especially irregular ones, and modern contextual embeddings behave somewhat differently than the original single-word vectors.

How is this different from a keyword search?

Keyword search matches characters. Embedding search matches meaning, so it can find “computing surface measurements” when the student typed “how do I calculate area.” The trade-off is that it can also retrieve something merely on-topic rather than actually relevant.

Do embeddings have bias?

Yes, because they’re learned from human text and encode the statistical associations in it, including stereotyped ones. This isn’t a bug in the math; it reflects the data. It’s worth knowing about before your kid takes a retrieved passage as neutral fact.

Why does my kid’s AI study tool sometimes pull the wrong section?

Usually one of three reasons: the source was chunked badly, so a vector represents half an idea; the question was phrased in a way that lands near a different topic; or the retrieved passage is genuinely about the topic but says the opposite of what’s needed. Negation is a documented weak spot.

---

About the author

Ricky Flores is the founder of HiWave Makers and an electrical engineer with 15+ years of experience building consumer technology at Apple, Samsung, and Texas Instruments. He writes about how kids learn to build, think, and create in a tech-saturated world. Read more at hiwavemakers.com.

---

Sources

1. Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013). “Efficient Estimation of Word Representations in Vector Space.” arXiv:1301.3781. https://arxiv.org/abs/1301.3781
2. Google Workspace Updates. (2026, August 4). “Gemini in Google Classroom is expanding to users of all ages with contextualized Gemini starter prompts for students.” https://workspaceupdates.googleblog.com/2026/08/gemini-in-google-classroom-is-expanding-to-users-of-all-ages-with-contextualized-Gemini-starter-prompts-for-students.html
3. Lewis, P., Perez, E., Piktus, A., et al. (2020). “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.” NeurIPS 2020. https://arxiv.org/abs/2005.11401
4. Vaswani, A., Shazeer, N., Parmar, N., et al. (2017). “Attention Is All You Need.” NeurIPS 2017. https://arxiv.org/abs/1706.03762
5. Google. (2026, August). “Google AI updates, August 2026.” The Keyword. https://blog.google/innovation-and-ai/technology/google-ai-updates-august-2026/
6. Liu, N. F., Lin, K., Hewitt, J., et al. (2023). “Lost in the Middle: How Language Models Use Long Contexts.” TACL. https://arxiv.org/abs/2307.03172
7. Common Sense Media. (2026, August 18). “Teens in the AI Era: Schoolwork and Skills That Matter.” https://www.commonsensemedia.org/research/teens-in-the-ai-era-schoolwork-and-skills-that-matter