NLP Explained: Why Your Phone Thinks You're Swearing

NLP has changed significantly in the past two years. The tools you are using in March 2026 are not the same as the tools from 2023, and three specific upgrades are worth knowing about.

Adaptive Thinking is a feature where the model chooses how hard to think before responding, using more reasoning for tricky inputs and less for simple ones.

Claude 4.6 now uses what Anthropic calls Adaptive Thinking.

In plain terms, the model can choose how hard to think before responding. A message like “Hello” gets an instant reply because the model does not need to reason.

But a sarcastic sentence like “Oh great, another Monday” requires a moment of processing to identify that “great” is not being used sincerely.

The system allocates more computational effort to genuinely ambiguous or complex inputs and less to simple ones.

This is the waitstaff equivalent of a seasoned server who spots immediately that your "I'm fine" is a polite fiction, while also not overthinking it when you just want the bill.

Context compaction is how modern language models compress earlier parts of a conversation into summaries so they can remember the meaning without keeping every word.

Earlier language models had a memory problem.

In a long conversation, they would effectively “forget” what was said at the beginning by the time they reached the end. The context window, the amount of text a model could hold in view at once, had a hard limit.

2026 models handle this with context compaction. Rather than dropping old content when the window fills up, the model compresses earlier parts of the conversation into summaries and retains those.

The details are condensed, but the meaning is preserved.

Think of it as the waitstaff keeping a tidy notepad. They do not memorise every word of every previous conversation.

But they keep a running summary: "Table 4 has been here before, prefers the fish, asked about allergies." The important context stays accessible even when the verbatim detail is gone.

Research from the Oxford NLP Group on long-context language modelling, published at nlp.cs.ox.ac.uk, traces how this problem was identified and the approaches researchers developed to address it.

One practical note here: if you are feeding sensitive client information into an AI tool during a long conversation, context compaction means the model is building and retaining a compressed version of everything you have shared.

The waitstaff's notepad does not disappear between sessions in all tools. Before you start pasting contracts, client details, or confidential business data into any AI window, our AI security guide: what's safe to share covers exactly what is fine, what is risky, and what to avoid.

Once you start noticing NLP in action, you'll see it everywhere:

Voice assistants like Siri, Alexa, and Google Assistant use NLP to turn your spoken questions into text, understand what you meant, and keep track of short conversations over multiple turns.

Siri, Alexa, Google Assistant. Always ready to misunderstand you in creative new ways. The genuine improvement over the past three years is context.

Earlier versions treated each query in isolation.

Current versions maintain conversational context across a short exchange, which is why you can now say “What’s the weather tomorrow?” and follow it with “What about the day after?” without the assistant losing the thread

Autocorrect and predictive text use NLP models trained on huge amounts of writing (plus your own messages) to guess what you meant to type and which words you are likely to use next.

The system that insists you meant “ducking” has processed enough clean, family‑friendly text that a specific four‑letter word appears far less frequently in its training data than its polite substitute. The pattern wins. Your intention loses.

The more interesting part: predictive text on modern phones now uses a local, on‑device NLP model that has learned from your specific writing patterns over time.

It knows you always text your mum on Sunday evenings and that you typically use complete sentences with her. It adjusts accordingly.

Spam filters use NLP to analyse the wording, tone, and structure of emails so they can spot suspicious messages even if they have never seen that exact text before.

Your inbox’s bouncer.

Deep learning spam filters now use NLP to understand the intent and tone of an email, not just flag known trigger phrases.

A message that has never been seen before can still be correctly identified as spam because its sentence structure, urgency framing, and request pattern match what spam tends to look like.

Sentiment analysis uses NLP to scan large volumes of text and work out whether the overall tone is positive, negative, neutral, or mixed.

Businesses use this to process customer reviews, social media mentions, and support tickets at scale.

Instead of someone reading every review manually, NLP identifies whether the tone is positive, negative, neutral, or mixed, and flags the ones that need human attention.

The practical limit: sarcasm still causes significant problems. "Oh, the delivery took three weeks. Brilliant service." registers as positive in unsophisticated systems because "brilliant" is a positive word.

Adaptive Thinking models handle this better, but it is not fully solved.

NLP streamlines content and communication by generating drafts for emails, social posts, and product descriptions while analysing customer feedback at scale.

Sentiment analysis on customer feedback at scale. Subject line testing before you commit to a campaign.

The pattern that works: use NLP tools to produce the first draft and handle volume. Use your own judgement for tone, accuracy, and anything that carries your name.

NLP tools process vast amounts of customer reviews quickly to reveal hidden themes and patterns you’d miss manually.

Not to replace customer conversations. To make sure you are not missing signals that are hiding in the volume.