A New Microsoft Patent Teaches AI Chatbots When to Be Warm and When to Be Brisk
Ask a chatbot for help with a grief and ask it to proofread an email, and you often get the same cheerful voice. A new Microsoft patent application tries to fix that by matching the amount of warmth to the job.
What Microsoft's empathy-tuned chatbot add-on does
Imagine you tell an AI assistant you're struggling after a rough week, and it answers in the same chirpy voice it uses to fix your spreadsheet formulas. Or you ask for help with a cover letter and get a long, gushing pep talk. Chatbots tend to have one emotional setting, and it's often wrong for the moment.
Microsoft's filing describes a way to give a chatbot different levels of warmth for different kinds of requests. The company studied hundreds of real conversations in which people said how much empathy they wanted before chatting and how much they felt they got afterward. People wanted much more for personal problems and distressing situations than for learning a skill or working through an assignment.
The patent then uses that pattern to train small add-on pieces, one per type of request, that nudge the chatbot's tone up or down depending on what you're asking about.
… cluster the tasks based on empathy levels and/or types associated with each task to identify shared empathy task clusters; determine empathy patterns for each task cluster; generate synthetic multi-turn conversations for each task cluster based on the empathy patterns …
Translation: The system groups similar tasks together and creates practice dialogues to learn the right tone.
How Microsoft trains a chatbot's tone by task
The filing describes a training phase and a use phase. In training, the system collects example multi-turn conversations (back-and-forth chats) between people and AI models. Each conversation is tied to a task, such as writing help or personal issues, and to how much empathy the user expected and how much they felt they received.
The tasks are then grouped into clusters that share similar empathy needs. In the company's example, eight tasks were sorted into four clusters, including "Distressing/Social/Personal Situations" and "Learning Skills." Each cluster gets its own empathy pattern.
Because real conversations are limited (the example dataset had 755), the system generates synthetic conversations, meaning AI-written practice chats steered toward each cluster's pattern. Those are used to train an expert adapter, a small add-on layer that changes how the main model talks without rewriting the model itself. The filing says these can be LoRA adapters, a common lightweight method.
A scoring system (a reward model, which grades conversations on how well they match the target empathy) helps fine-tune the adapter. At use time, the model spots your task and answers with the empathy level that fits. The claims also cover other tone traits, such as humor, formality, friendliness and curiosity.
… train an expert adapter based on the synthetic multi-turn conversations to adapt an output of the generative language model system to achieve a task-specific target level and/or type of empathy.
Translation: An add on module is trained using those practice dialogues so the chatbot knows how much warmth to show.
Why chatbot tone matters for Microsoft's AI products
For you, the practical payoff is an assistant that reads the room. Support bots, coaching apps and health-adjacent tools are where a tone-deaf reply costs the most, and the filing names mental health support, customer service and coaching as target uses. The company reports that adapters cut the gap between desired and perceived empathy by about 72.66% in its tests, and held up better across long chats than a plain instruction telling the model to be empathetic.
The filing is a patent application, and the examiner's first review rejected the claims as filed, which is routine and usually followed by amendments. Nothing here says it will ship in a product, but it shows Microsoft treating tone as something to be engineered per task.
This is the 19th Microsoft patent we've tracked in voice & speech AI since May, including one that filters out other voices and one that picks reply format.
The problem here is real. Anyone who has poured out a bad day to a chatbot and received a bulleted list of tips knows how jarring a wrong tone feels, and the filing's own data shows people want very different warmth for a personal crisis than for learning a skill. A mismatch costs trust, and in sensitive moments it can sting.
The approach fits the size of that problem. Tuning one dial per type of request, using what people said they wanted, is a sensible step, and the finding that a long instruction fades as a chat gets longer matches what many users notice.
The open question is whether the data behind it is big enough. It rests on 755 conversations from 109 mostly professional participants. That is a modest sample for deciding how a global audience wants to be comforted.
Get our take in your Top Stories
Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Microsoft patent news breaks.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
33 drawing sheets from US 2026/0339933 A1 · click any drawing to enlarge
Show the other 25 sheets
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →
Be the first to weigh in