Every term below is defined by its effect on output. If a setting does not change what you would ship, it does not need your attention — and several of the most-discussed terms in AI writing fall exactly into that category.
The ones that change your text
- Token
- The unit a model counts in — roughly three quarters of an English word. It matters for two practical reasons: you are billed by it, and every model has a ceiling on how many it can consider at once.
- Context window
- How much text the model can hold in mind at once, measured in tokens. Exceed it and the beginning of your document quietly stops influencing the end, which is why long drafts lose their own thread.
- Temperature
- How much randomness is allowed when picking the next word. Low values give you the same phrasing twice; high values give you variety and a larger share of sentences that need checking. For business writing, the useful range is narrower than most guides suggest.
- Top-p
- An alternative randomness control that limits the candidate words to the most likely set. In practice, changing temperature alone is enough; adjusting both usually means you are tuning without a target.
- System prompt
- The standing instruction that applies to everything in a session — tone, length, what to refuse, whose voice to use. One well-written system prompt is worth more than twenty clever one-off prompts.
- Few-shot
- Giving the model examples before asking for the output. Two or three real samples of your own writing will move the tone further than any adjective you can put in the instruction.
- Zero-shot
- Asking with no examples. Fine for formatting and summaries, weaker for anything where voice matters.
- Chain-of-thought
- Asking the model to show its reasoning before answering. Useful for structure and planning; rarely worth exposing in the finished text.
- Grounding
- Tying output to a supplied source rather than to the model's training. The single most effective way to reduce invented facts: give it the document, then ask.
- Hallucination
- A confident, fluent, false statement. Not a bug you can switch off — a property of how the next word is chosen. The defence is grounding and verification, not a better prompt.
The ones that describe how it was built
- Training data
- What the model learned from. It determines what the model sounds like and what it does not know — including anything that happened after its cutoff.
- Cutoff date
- The point after which the model has no reliable knowledge. Anything time-sensitive — pricing, rules, people's roles — has to be supplied by you.
- Fine-tuning
- Additional training on a narrower set of examples. Worth it for a repeated, well-defined task; rarely worth it just to make output "sound like us".
- Embedding
- A numeric representation of meaning, used to find related text. This is what makes search over your own documents work.
- RAG (retrieval-augmented generation)
- Fetching relevant passages from your documents first, then writing from them. The practical version of grounding, and the difference between a tool that guesses about your business and one that reads it.
- Vector database
- Where those embeddings are stored so they can be searched quickly. You only meet this if you are building something.
- Perplexity
- A measure of how surprised a model is by text. Used as a rough signal by detectors, and unreliable for that purpose.
- Weight / parameter
- The learned values inside the model. Bigger is not automatically better for a specific writing job, which is why small models handle short, repetitive tasks well.
- Distillation
- Training a smaller model to imitate a larger one. Why cheap tools are often adequate for short output and weaker on long documents.
- Watermarking
- Marking output so it can be identified later. Not reliably present across tools, and not something to build a policy on.
The ones that affect cost and access
- API
- A way for software to call the model directly. Relevant when you want the same task run a hundred times rather than typed a hundred times.
- Rate limit
- How many requests you can make in a window. The thing that breaks a batch job at 2am, not a single writing session.
- Seat
- A licensed user. The pricing unit that makes team plans expensive faster than expected.
- Credit
- A consumed unit rather than a user — common in tools where output length varies. Credit pricing is harder to forecast than seat pricing.
- Free tier / trial
- Limited access with no payment. Worth using to test your real workload, not a demo workload.
- Annual vs monthly
- Yearly billing usually discounts the monthly rate in exchange for committing twelve months. The discount is real; the question is whether the tool survives a year in your workflow.
- Data retention
- How long your input is kept, and whether it is used for training. The clause worth reading before pasting anything confidential.
- Opt-out
- A setting or contract term that excludes your content from training. Sometimes default-on, sometimes buried, sometimes only available on business plans.
The ones that come up in arguments
- AI detector
- A tool that estimates whether text was generated. False positives on careful human writing are common enough that no serious workflow should treat a score as evidence.
- Paraphrase
- Rewriting the same meaning in different words. Useful for clarity; useless as a way to make generated text something it is not.
- Plagiarism
- Using someone's words without attribution. Distinct from AI use, and the two are frequently confused in policy documents.
- Disclosure
- Telling a reader that AI was involved. A client or audience question rather than a technical one, and best answered explicitly.
- Brand voice
- The consistent way an organisation writes. The hardest thing to get from a model and the most valuable thing to supply it, usually as examples rather than adjectives.
- Guardrail
- A rule limiting what the model will produce. Useful in products, largely invisible when you are writing.
- Latency
- Time to first word and to finish. It matters more than benchmarks suggest, because a slow tool gets used less.
- Prompt injection
- Instructions hidden in content the model reads. Relevant when a model is pointed at documents you did not write — including web pages.
- Copyright / ownership
- Who can claim the output. Varies by jurisdiction and by tool terms; worth reading the actual clause rather than the summary.
- Style guide
- Your written rules — capitalisation, numerals, tone. Paste it in as a system prompt and half of your editing disappears.
What this costs, from our 30-day benchmark
Our 30-day test recorded roughly 22 minutes of editing per 1,000 words on the low-cost dedicated writer against about 8 minutes on the general assistant — near $7 per piece of your time at $30 per hour. Of the settings above, grounding and a written style guide moved that number; temperature and top-p did not.
Two settings worth changing
Ground it in your documents, and paste your style guide in as a standing instruction.
Try Rytr →Frequently Asked Questions
Which AI writing terms actually matter day to day?
Context window, grounding and your standing instruction. Those three change the output you ship. Most of the rest describes how the model was built or how you are billed.
What is grounding, in plain language?
Giving the model the document you want it to write from, instead of letting it recall something similar. It is the most effective way to cut invented facts.
Are AI detectors reliable?
Not for decisions. They produce false positives on careful human writing, so a score should never be treated as evidence about how a document was produced.
Does a bigger context window help?
Only for long documents. Once a draft exceeds the window, the opening stops influencing the ending, which is when structure starts to drift.