What a Token-Level Heatmap Is Really Telling You About Your Draft

You paste your essay into a detector and instead of a single verdict you get a wall of color: some words glowing red, others cool blue, a few in between. It looks like the tool is pointing straight at the sentences you should be ashamed of. It isn’t, exactly. A token-level heatmap is showing you something narrower and more interesting than “we caught you,” and once you know how to read it, it turns into one of the better revision tools you have.

Key takeaways

• A heatmap colors each token by how predictable it was to the model, not by whether you personally wrote it. Red means “easy to guess,” not “caught.”
• The colors are relative and contextual. The same word can light up in one sentence and stay cool in another, because the surrounding text changes what the model expected.
• Read the heatmap by pattern, not by pixel. A few scattered hot tokens are noise; long unbroken runs of red are the signal worth acting on.
• Short passages produce unreliable heatmaps. With too few tokens, the model has barely any context, so the colors swing on almost nothing.
• Use it as a revision map, not a scoreboard. Hot runs usually mark the spots where your writing coasted.

What a token actually is

First, a quick clarification, because “token” trips people up. A token isn’t always a whole word. Detectors break text into chunks the model recognizes, and a chunk might be a full word, part of a long word, or even a piece of punctuation. “Humanizing” might split into “human” + “izing.” For reading a heatmap, you don’t need to obsess over the boundaries. Just know that the unit being colored is usually a word or word-piece, which is why you sometimes see half a word lit up and the other half cool.

What the color is measuring

Here’s the part that changes how you read everything else. The detector runs a language model over your text and, at each position, asks one question: given all the words so far, how likely was this exact next word?

If the model would have confidently predicted that word, the token scores as low-surprise and gets painted hot. If the word caught the model off guard, it scores as high-surprise and stays cool. Stack those per-token surprise values up and you’ve got the statistics those colors are built on — perplexity is basically the average surprise across the whole passage. This isn’t a new trick, either: researchers built a tool called GLTR back in 2019 that highlighted text exactly this way, coloring each word by how predictable a model found it.

So red doesn’t mean “an AI wrote this word.” It means “this word was easy to predict here.” Those are not the same claim, and the gap between them is where most people misread the tool.

Why does predictability correlate with AI text at all? Because models tend to pick high-probability words. Left on autopilot, they write the sentence the model itself finds most likely, which is exactly the sentence that scores as low-surprise when you run it back through a model. Human writers swerve. We pick the odd verb, drop in a specific detail, change rhythm mid-paragraph. Those swerves register as surprise.

But plenty of human writing is predictable too. “Thank you for your time and consideration” is going to glow red no matter who typed it, because it’s a fixed phrase. That’s the honest limitation baked into the whole approach.

A mini-scenario: two versions of the same sentence

Say you wrote this:

> *In today’s fast-paced world, effective communication is essential for success in any organization.*

Run that and it’ll come back almost solid red. Every word is the word the model expected: “fast-paced” follows “today’s,” “essential” follows “is,” “success” follows “for.” Nothing surprises it. The heatmap is telling you, accurately, that this sentence could have been generated by anyone or anything.

Now the revised version:

> *Most of the screw-ups on my last team traced back to one thing: nobody wrote down what “done” meant.*

The same heatmap goes mostly cool. “Screw-ups” is an unlikely word after “the.” The colon and the quoted “done” break the pattern. A model wouldn’t have predicted this sentence, so the surprise scores climb and the colors drop.

Notice what actually happened. You didn’t trick anything. The second sentence is just more specific and more yours. The heatmap rewarded concreteness because concreteness is unpredictable. That’s the whole game, and it’s why the tool doubles as a writing-quality signal even if you never cared about detection.

How to read the map without overreacting

The single biggest mistake is reacting to individual tokens. A lone red word in a sea of blue means almost nothing — common words are predictable, and you can’t write English without them. Squint instead. You’re looking for shape.

Long unbroken runs of red are the real signal. When eight or ten tokens in a row all score as low-surprise, that’s a stretch where the writing is running on rails: generic phrasing, predictable structure, no specifics. Those runs are where editing pays off.

Scattered warm tokens inside otherwise varied text are normal. Leave them. If you try to cool every last one, you’ll start mangling sentences to insert artificial surprise, and readers notice that faster than any detector does.

Watch the openings. The first sentence of any passage tends to read warm because the model has little context to be surprised by yet. Don’t panic at a hot opening line in isolation.

A practical habit: skim the heatmap for the longest red blocks, rewrite those two or three passages with real detail, and ignore the confetti. Watch which blocks cool down as you add specifics versus which ones don’t budge — the stubborn ones are usually unavoidable common phrasing.

Token-level vs. sentence-level views

Some tools color individual tokens; others highlight whole sentences with a single shade. Sentence-level views are easier to act on because they’ve already averaged out the per-token noise into “this sentence reads flat” or “this one doesn’t.” Token-level views show you more granularity but demand more interpretation.

If you have both, use the sentence view to decide *what* to rewrite and the token view to understand *why* a sentence scored the way it did. The percentage at the top is just the same data rolled into one number — a temperature reading, not a different measurement.

What the heatmap can’t tell you

It can’t tell you who wrote the text. It’s measuring the words on the page, not your intent or your process. A heavily edited AI draft can come back cool; a careful human first draft full of stock phrases can come back hot.

It can’t tell you the model is sure. Detectors output probabilities, not proof, and the per-token scores are even shakier than the overall verdict. That’s not a minor footnote. OpenAI pulled its own AI text classifier in July 2023, saying it was no longer available due to its low rate of accuracy. And a 2023 Stanford study (Liang and colleagues) found that several detectors flagged writing by non-native English speakers as AI-generated far more often than native writing — partly because that writing leans on predictable, learned phrasing. A heatmap inherits every one of those weaknesses.

And it can’t tell you much at all on short text. With only a sentence or two, there’s barely any context for the model to be surprised by, so the colors swing wildly on tiny differences. Treat heatmaps on anything under a couple hundred words as rough at best.

Turning the map into edits

Here’s the workflow that actually uses the heatmap well:

1. Run the full passage and ignore the headline number for a second.
2. Find the longest red runs. Read them aloud. They’ll usually be the parts that sound like filler even to you.
3. Rewrite those for substance, not surprise. Add a specific number, a real example, a concrete noun, an honest opinion. Don’t reach for the thesaurus.
4. Re-run and watch which blocks cooled. Blocks that stayed warm despite specific edits are often just unavoidable common phrasing — fine to leave.
5. Stop when the writing reads like a person, not when the screen is uniformly blue.

That last point matters. An all-cool heatmap is a warning sign, not a trophy. The fastest path to total blue is to write something nobody would predict because nobody would write it that way — and that reads as worse, not more human. You can browse more of our writing on how detection works if you want the deeper mechanics, or check the plans before you run a lot of drafts through it.

Frequently asked questions

What does a red token actually mean on an AI detection heatmap?

It means the model found that token highly predictable given the words before it. The detector runs a language model over your text and asks, at each position, “how likely was this exact word here?” Tokens the model would have guessed easily get colored hot; surprising ones stay cool. Red doesn’t mean you used AI for that word, and it doesn’t mean you got caught. Plenty of perfectly human phrases are predictable, because that’s how common phrasing works.

Why do the same words light up differently in different sentences?

Because the score for a token depends on everything before it. A language model predicts each word from its context, so “results” might look ordinary after a generic lead-in and surprising after a specific, unusual clause. Move a phrase, add a detail earlier, or change the sentence it lives in, and its color can shift even though the word itself didn’t change. Read runs of color in context, not single highlighted words.

Is a token-level heatmap more accurate than the overall percentage score?

It’s not more or less accurate, it’s a different view of the same underlying numbers. The percentage is a summary; the heatmap shows you where that summary came from. The heatmap is more useful for editing because it points at specific passages, but it’s noisier up close. Use the overall score as a rough temperature reading and the heatmap to decide which paragraphs to rework.

Should I rewrite every hot section until the whole heatmap is cool?

No. Chasing an all-blue heatmap usually makes writing worse, because the fastest way to raise unpredictability is to get weird: odd word choices, broken rhythm, forced complexity. Readers feel that. Fix the long red runs that mark genuinely flat, generic writing, then stop. If a hot stretch is accurate, specific, and reads well aloud, leaving it slightly warm is fine.

Why is the heatmap unreliable on short text?

Token-level scoring needs context. The model predicts each word from the ones before it, so in the first sentence or two there’s almost nothing to go on, and the colors swing on tiny differences. A 30-word note can show alarming red or perfect blue for reasons that have nothing to do with how you wrote it. The longer the passage, the more the heatmap settles into something meaningful.

The short version

A token-level heatmap is a predictability map, not a verdict. Red marks where your writing was easy to guess, which often — but not always — overlaps with where it was lazy. Used right, it points you at the flat spots so you can replace them with the specific, slightly surprising stuff only you would write. Read it for patterns, fix the long red runs, and don’t chase a perfect score. Want to see your own draft lit up token by token? Try it here.