Blog

What belongs here

The articles will focus on the design decisions that survive contact with real conversations: when an AI should disagree, what it should forget, how to spot a false memory, what changes after a model update, and when a friendly voice crosses into manipulation. Expect practical design explanations and, where work has actually been done and documented, build notes, annotated examples, methods, and tests. A proposed evaluation method will be described as a method; measured findings will need their own evidence and limitations.

The aim is to help product teams, designers, and researchers reason about the whole conversational system. A useful answer may need a prompt change, but it may also need a different memory rule, interface affordance, review process, or product incentive. This is not a stream of rewritten AI announcements.

Explore AI personality, companions, and human–AI interaction

These topics overlap, but each asks a different question. Start with the behavior you want to understand rather than a label for the technology.

AI personality: what stays recognizable?

AI personality is the pattern of behavior people encounter across conversations: voice, initiative, disagreement, and boundaries. The practical challenge is to make those choices consistent without making every reply sound the same. A persona description is a starting point; examples, memory rules, and evaluation turn it into a design that can be examined.

AI companions: what should carry across visits?

AI companions are built for ongoing interaction. That makes memory, expectations, and the freedom to leave central design questions. This publication approaches companions through continuity and user control, not claims that an artificial character has human feelings. Useful questions include how a companion handles forgotten context, corrects false recall, and avoids pressure to keep talking.

Conversational AI: what happens when dialogue goes wrong?

Conversational AI includes systems people use through text or speech. Here, the focus is dialogue behavior: asking a useful follow-up, recovering from misunderstanding, disagreeing clearly, and explaining when the system cannot help. A successful exchange should help the person accomplish their purpose, not merely produce another fluent response.

Human–AI interaction: what can the person understand and control?

Human–AI interaction connects the conversation to the interface and the decisions around it. Can a person see what the AI remembers, challenge an answer, undo a choice, or stop an action? Transparency and control need to work in the product, not just appear in its description. This is where character design meets responsible use.

Memory and evaluation connect the topics

Across all four areas, the same practical tests recur: a correction, a return visit, a missing memory, a request to forget, and a change of model. The AI personality engineering guide below introduces a working framework and an illustrative study-companion example. The English article series will develop narrower questions; planned topics are not presented as completed research.

How to read the work

I am Cao Hung Nguyen, the writer behind The AI Whisperer. My working position is simple: believability is coherence, not deception. An AI can have a clear character and still disclose that it is an AI, admit uncertainty, refuse a request, and respect the user’s control over memory. When a piece presents an illustrative conversation, it should say so. When it reports an observation or a test, it should describe the setup and what the result cannot establish.

The site is moving its editorial focus to English while keeping its German archive accessible. The German transition note explains why; it is not an English-language article in this new series. You can also read the About page for background on Cao’s work.

Working on a conversational product?

If your team is defining a character, revising memory behavior, or trying to find where a companion loses its boundaries, send Cao a question or feedback. Bring a concrete interaction or failure case; we can start there.