Using a chatbot is pretty far from the highlight of anyone's day.

They started out being almost completely useless: inflexible and with a limited range of responses that could send you round and round in circles. Now that the underlying technology is improving, however, while the user experience has undoubtedly improved, you've begun to ask yourself: Wait, how much of what I'm writing here is being remembered and noted down by the bot? Is this secure?

Privacy Worries

If you've ever wondered how much a chatbot remembers, you've probably also worried about your privacy. Every time we type a question into an online assistant or customer support window, it feels as though we’re giving it a little piece of our personal information. What we ask the bots helps the large language models that underpin them improve their responses over time, but also creates a dilemma: how is it possible to let the system learn—and deliver a better user experience—without exposing details of what they have learned from others and ourselves?

Plausible Token Amplification

Researchers at NTT believe they have found an answer. Their new Plausible Token Amplification, or PTA, makes it possible for an AI model to learn from previous interactions, while still keeping those interactions private. The technique was presented at the International Conference on Machine Learning (ICML) 2025, one of the world’s most selective machine-learning conferences, held in Vancouver this July.

To understand what PTA does, let's unpack its name. In an AI model, a token is a small fragment of text—a word or part of a word—that the system uses to process language. Amplification means giving greater importance to something. And plausible in this context means that, even after the data are modified for privacy, they still look natural and believable to the model. Plausible Token Amplification therefore means highlighting the words that matter most, while keeping the data both useful and private.

The Problem with In-Context Learning

The issue (or problem) that PTA is designed to combat is the unwanted effects of "in-context learning," which is where a model improves its responses by looking at example question-and-answer pairs within the chatbot prompt. For instance, if you show it a few customer queries and their correct replies, it can begin to infer how to answer a new question of a similar kind. All very good, of course, but the side-effect of this ability is that those examples may contain private details from past users.

Differential Privacy Helps... But Isn't Perfect

To prevent this happening, until now bot creators have tried to use differential privacy, which adds controlled noise—small random changes—to obscure sensitive information. Doing this protects the inputted data, but the noise also makes the model less accurate. What differential privacy ends up doing is hiding the very clues that guide the model's reasoning.

Deciding Which Words Matter

PTA fixes that imbalance. Before the differential privacy noise is added, the system identifies the tokens that carry the main meaning—words like “delivery,” “refund,” or “address”—and gently strengthens their presence in the data. When the random noise is then applied, those important words remain visible enough for the model to learn the underlying rule, without exposing anyone’s personal details.

PTA is supported by Bayesian analysis, a mathematical way of reasoning that updates what we believe as new evidence appears. In this case, it helps the model estimate which words are most likely to define the right rule or intent. PTA thereby gives the system a clearer sense of which clues matter most before the noise is applied, restoring much of the accuracy lost in standard privacy-protected learning.

Better Bots, Safer Data

Long story short: this matters for anyone who interacts with AI systems that learn from conversation: customer service platforms, healthcare bots, educational tutors, or office assistants that improve through use. With PTA, these systems can learn from real interactions, while at the same time getting better at keeping private information private. A banking chatbot, for example, could become better at recognizing transaction-related questions, without ever blurting out details from previous customers. A medical triage assistant could refine its understanding of symptoms, while also keeping patient data secure.

NTT's Plausible Token Amplification is a way for people to trust that AI can be both intelligent and careful with their data. It's about knowing that progress in AI needs to serve human wellbeing, rather than random performance metrics. Designing privacy-preserving systems that can communicate naturally and effectively, NTT is creating technology that exists not just for its own sake, but for the people who use it. It's building intelligence with empathy: learning responsibly, responding thoughtfully, and earning trust with care.

Innovating a Sustainable Future for People and Planet

For further information, please see this link:
https://group.ntt/en/newsrelease/2025/07/07/250707b.html

If you have any questions on the content of this article, please contact:
Public Relations
NTT Service Innovation Laboratory Group
https://tools.group.ntt/en/rd/contact/index.php

Daniel O'Connor

Daniel O'Connor joined the NTT Group in 1999 when he began work as the Public Relations Manager of NTT Europe. While in London, he liaised with the local press, created the company's intranet site, wrote technical copy for industry magazines and managed exhibition stands from initial design to finished displays.

Later seconded to the headquarters of NTT Communications in Tokyo, he contributed to the company's first-ever winning of global telecoms awards and the digitalisation of internal company information exchange.

Since 2015 Daniel has created content for the Group's Global Leadership Institute, the One NTT Network and is currently working with NTT R&D teams to grow public understanding of the cutting-edge research undertaken by the NTT Group.