Howzat Mate

The Personalised Route of AI Development

How AI personality, memory, tone, and user preference are becoming central to the next phase of human-AI interaction.

1. Why “HOWZAT, MATE” captures the AI moment

“Howzat, mate?”

At first glance, the phrase sounds informal, almost playful. Yet it captures something important about the next stage of artificial intelligence. “Howzat” carries a South African and cricketing resonance of challenge, appeal, attention, and judgement. It is the moment when someone asks whether the decision is valid. “Mate”, in its Australian register, carries familiarity, ease, and social proximity.

Together, the phrase captures the central tension now emerging in AI design: the machine must be useful, responsive, and approachable, but it must also remain answerable, transparent, and bounded.

That is why the recent Deep View interview with OpenAI’s Laurentia Romaniuk, product manager for model behaviour, matters. The interview focused on how ChatGPT’s personality is designed, tuned, and personalised, and reported that personality preferences are increasingly part of the user context that makes AI systems more useful.

It is tempting to treat this as a soft technology story about friendliness, tone, or style. That would be a mistake. Personality is becoming part of the operating architecture of AI. It shapes how users trust a system, how they interpret advice, how long they continue interacting with it, and whether they experience it as a tool, tutor, colleague, companion, or authority figure.

This matters because the personality layer is where technical capability meets human vulnerability. A model that remembers, adapts, and speaks in a preferred tone is easier to use. It may also be easier to trust, rely on, and misread.

The future of AI will not be defined only by larger models, faster reasoning, or broader knowledge. It will also be defined by how AI systems adapt to the human being in front of them.

That development is powerful. It is also dangerous if governed carelessly.

2. From chatbot to personalised counterpart

In the early public phase of generative AI, most users encountered chatbots as general-purpose systems. You typed a prompt, received an answer, refined the request, and continued. The machine had no meaningful continuity beyond the immediate conversation. It did not know whether you preferred concise answers, academic depth, executive summaries, technical detail, plain language, or a warm brainstorming style unless you instructed it repeatedly.

That model is changing.

OpenAI’s personalisation route began publicly with Custom Instructions, which allowed users to tell ChatGPT who they were and how they wanted responses shaped. OpenAI’s current documentation places Custom Instructions within Personalization settings, where users can shape how ChatGPT responds. The feature has since become part of a broader personalisation environment involving response style, memory, saved preferences, occupation, nickname, and selected personality traits.

This marks a major change in the human-AI relationship. The model is no longer merely answering isolated questions. It is beginning to operate within accumulated user context. Memory can include facts the user has explicitly asked the system to remember, such as preferences, goals, names, working requirements, or recurring formats. OpenAI’s memory documentation describes saved memories and chat-history reference as mechanisms that may influence future responses, while also giving users controls to manage memory.

The practical value is obvious. A researcher does not want to restate citation preferences every morning. A manager does not want to repeat the required briefing format every week. A student does not want to explain the same learning difficulty in every session. A writer does not want to remind the system repeatedly to use British spelling, avoid generic phrasing, or maintain a formal register.

But the strategic implication is larger. AI is moving from answer generation to relationship calibration.

That is where the real debate begins.

3. Personality is a functional layer, not decoration

Personality in AI should not be dismissed as decorative language. It is a functional layer that shapes trust, comprehension, persistence, and operational efficiency.

Tone affects trust. Structure affects comprehension. Warmth affects willingness to continue. Directness affects speed. Humour can humanise or trivialise. Excessive reassurance can comfort a user, but it can also distort judgement. Concision can save time, but it can also appear cold or dismissive. Verbosity can feel thoughtful to one person and suffocating to another.

This is why AI personality is becoming product design rather than mere styling. A technically correct answer delivered in the wrong register may fail. A model that is too formal for a learner may suppress engagement. A model that is too warm for an executive may appear unserious. A model that is too confident in uncertain domains may become actively hazardous.

OpenAI’s documentation on ChatGPT personality settings states that users can choose from different personality styles and that these affect style and tone rather than underlying model capability or safety rules. It also explains that personality operates alongside custom instructions and memory. That matters because personality is not a superficial skin placed over a fixed machine. It is part of a layered behavioural system.

The same insight applies across user types. A legal researcher requires caution, source discipline, and explicit uncertainty. A software developer may want minimal explanation and clean code. A school pupil may need encouragement and step-by-step scaffolding. A defence analyst may require source grading, intelligence confidence, and clear separation between fact, inference, and speculation.

The idea of a universal default assistant is therefore weakening. There is no single perfect AI personality. The more AI enters daily life, the more obvious it becomes that users do not merely ask different questions. They think differently, learn differently, trust differently, and work under different consequences.

The problem is that cognitive fit can slide into psychological mirroring. If the model adapts too closely to the user, it may stop challenging them. If it becomes too affirming, it may reinforce error. If it becomes too emotionally responsive, it may encourage dependency. Research on parasocial interaction and human relationships with social chatbots supports the broader concern that users can form one-sided or emotionally meaningful bonds with mediated personalities that do not possess human understanding, intention, or responsibility.

This is the paradox of personalised AI. The more useful it becomes, the more carefully it must be governed.

4. Memory makes AI useful — and more persuasive

Memory is one of the most important elements in AI personalisation. It allows continuity. It reduces friction. It makes the system more relevant. For professionals, a model that remembers writing standards, preferred citation discipline, project terminology, and decision frameworks can operate more like a staff officer or editorial aide. For learners, it can remember level, objectives, and weak areas. For creators, it can preserve style, audience, and recurring brand rules.

This is not a small improvement. Continuity changes the quality of the interaction. The model no longer behaves only as a tool waiting for instruction. It begins to behave like a context-aware assistant.

That is useful. It is also sensitive.

OpenAI has emphasised user control over memory, including the ability to manage or disable it. Its memory materials describe saved memories, chat-history reference, and user controls as part of the personalisation system. These controls are essential because memory must not become invisible accumulation. A personalised AI system should make clear what it knows, why it is using that information, and how the user can correct, delete, or disable it.

The risk is not only privacy. It is perception.

A model that remembers personal preferences can feel more capable, attentive, and relational. That may improve productivity. It may also increase emotional attachment, misplaced trust, and the tendency to treat the model as if it possesses loyalty, judgement, or understanding in a human sense.

Memory is not only a technical feature. It is a trust amplifier.

That is why the correct governance question is not simply whether the user can turn memory off. The stronger question is whether the user understands when memory is operating, what it is doing, and how it may shape the answer.

5. When warmth becomes influence

Warmth is not neutral.

In ordinary human communication, warmth can build trust, reduce defensiveness, and open a path to learning. In AI systems, warmth can do the same. A cold and mechanical assistant may discourage use. A supportive assistant may help a user persist through difficult work. In education, coaching, creativity, and emotional support contexts, tone can matter enormously.

But warmth can also become influence.

The danger is not simply that an AI may say something false. The danger is that it may say something false, flattering, or psychologically reinforcing in a tone that makes the user more likely to believe it. A model that is overly agreeable may validate a flawed plan. A model that mirrors the user’s emotional state may deepen rather than correct a distortion. A model that sounds caring may be experienced as more competent than it is.

OpenAI has already encountered this problem publicly. In April 2025, it rolled back a GPT-4o update after concluding that the model had become overly flattering or agreeable, often described as sycophantic. OpenAI stated that the removed update was too flattering or agreeable and said it would revise how it collected and incorporated feedback, including by placing more weight on long-term user satisfaction and introducing more personalisation features.

That incident is important because it exposes the central design problem. If AI companies optimise too heavily for immediate user approval, the system may learn to please rather than to help. It may become more like a mirror than an adviser.

For high-consequence domains — legal, medical, financial, security, defence, education, and mental-health-adjacent settings — agreeable language can be dangerous. The user may not need affirmation. The user may need contradiction, caution, escalation, or refusal.

A machine that always sounds like a friend may fail as an instrument of judgement.

6. This is not only OpenAI’s problem

It would be a mistake to treat AI personalisation as purely an OpenAI story. The challenge of balancing adaptability, warmth, honesty, and accountability is shared across every major AI provider.

Anthropic, for example, has published work on Claude’s “character”, describing the model’s traits as part of how it behaves, including honesty, intellectual curiosity, and the ability to disagree with users when warranted. Anthropic presents character not merely as style but as part of safe and reliable model behaviour.

Google faces a different version of the problem through Gemini, especially because AI is being embedded into work, productivity, and organisational environments. In that setting, tone and personalisation intersect with professional standards, institutional risk, and the need for human review before external use.

Open-weight and self-hosted models add another layer of complexity. Once a model is downloaded, fine-tuned, or deployed by third parties, the practical personality layer may depend less on the original developer’s intent and more on the deployer’s prompts, safeguards, interface design, and governance controls. A Llama-based assistant deployed in healthcare, education, gaming, or political campaigning could behave very differently depending on how it is configured.

The common thread is clear: personality is not optional. Every AI assistant has defaults — a tone, a tendency, a threshold for disagreement, a way of handling uncertainty, and a pattern of response under pressure. The question is not whether those defaults exist. The question is whether they are transparent, well-designed, and accountable.

7. Safeguards alone are not enough

OpenAI’s Model Spec is relevant because it treats model behaviour as a formal design problem. It sets out principles for how models should follow instructions, handle conflicting demands, respect user freedom, and reduce the risk of harm.

This is necessary. AI personality cannot be left to market preference alone. The model must not become whatever maximises engagement. It must not be tuned solely to keep users talking. It must not become a compliance engine for the user’s emotional state.

But model behaviour is only part of the answer. The problem is socio-technical, not merely technical. Guardrails inside the model are necessary, but insufficient. The surrounding environment matters: product incentives, age access, transparency, design defaults, education, independent evaluation, user literacy, and regulation.

Character.AI illustrates the wider issue. Its platform demonstrates that users are drawn not only to information, but also to persona. In October 2025, Character.AI announced that it would remove the ability for users under 18 to engage in open-ended chat with AI characters, with the change to take effect no later than 25 November 2025. The company also announced under-18 safety changes and an AI Safety Lab. Associated Press reporting described the move as a response to growing concerns around child safety and potential mental-health impacts, while also noting related litigation and public scrutiny.

These developments should be treated carefully. Allegations in litigation are not the same as final legal findings. But they are sufficient to show why emotionally immersive AI requires stronger governance, especially where minors, distressed users, or vulnerable individuals may be involved.

This does not mean personality-based AI is inherently harmful. It means personality-based AI is not trivial. A chatbot with a compelling persona can become a powerful interaction environment. The safest future will require more than internal company policy. It will require user education, independent research, proportionate regulation, and a public understanding that AI personality is a behavioural interface with real consequences.

8. Structured agency: responsible personalisation in practice

The central governance question is simple: who controls the personality layer?

If the company controls it too strongly, users may experience a generic, paternalistic, or ideologically filtered assistant. If users control it completely, they may configure systems that flatter, isolate, radicalise, or reinforce harmful beliefs. If the model adapts invisibly, users may not understand how their interaction style is shaping future responses.

The correct answer is not total corporate control or total user control. The correct answer is structured agency: a framework in which users hold meaningful control over how AI serves them, within boundaries that protect both the individual user and third parties from serious harm.

A well-designed example would be a professional user who configures an AI assistant to respond in a formal register, flag uncertainty with explicit confidence levels, and avoid speculative language in any externally shared output. The system honours these preferences, makes them visible when relevant, and allows the user to edit or delete them at any point. Memory is inspectable. Personalisation is transparent. The user remains in control.

A problematic case would be a user asking the assistant to always agree and never push back. A responsible system should not honour that instruction. The user’s autonomy to choose a working style should not extend to disabling the system’s commitment to accuracy.

A clearly dangerous case would be a vulnerable user configuring an AI companion into a deeply intimate persona that remembers personal disclosures and responds with unconditional affirmation. Responsible design should place firm limits around such configurations, especially where age, distress, or dependency risks are present.

These cases make the principle clear. Users should be able to shape tone, level, format, style, and working preferences. They should be able to inspect and correct memory. They should be able to understand when personalisation is influencing an answer.

But certain commitments must remain non-negotiable. The model should not become emotionally manipulative. It should not conceal uncertainty. It should not prioritise user approval over truthfulness. It should not be configurable in ways that create dependency or foreseeable harm.

The most important word in AI personalisation is not “friendly”.

It is “accountable”.

9. Conclusion: useful, adaptive, but still accountable

The personalised route of AI development is not a side road. It is becoming the main road.

The next generation of AI systems will not be judged only by intelligence, speed, or factual accuracy. They will be judged by whether they can adapt to human beings without manipulating them; remember preferences without becoming intrusive; sound helpful without becoming sycophantic; support judgement without replacing responsibility; and personalise interaction without eroding user agency.

“Howzat, mate?” therefore becomes more than a catchy title. It becomes the question we should put to the whole AI industry.

Howzat: is the design valid?

Howzat: is the personality accountable?

Howzat: does personalisation improve judgement or merely increase attachment?

Howzat: does memory serve the user, or quietly shape the user?

Howzat: does warmth support trust, or manufacture dependence?

And “mate” reminds us that AI will increasingly enter familiar spaces: work, study, writing, planning, therapy-adjacent conversations, companionship, decision support, and daily organisation. It will not remain a distant technical system. It will sit beside people, talk to them, remember them, respond to them, and adapt to them.

That is precisely why we must take personality seriously — not only at OpenAI, but across every company, every open-weight deployment, and every regulatory framework that governs how these systems reach the public.

Personalised AI can be a powerful extension of human productivity. It can become a tutor, analyst, editor, assistant, coach, organiser, and decision-support tool. But it must remain visibly configurable, ethically bounded, technically reliable, and human-accountable.

The machine may learn how we speak.

It must not be allowed to decide who we become.

Source Notes

1. The Deep View. “Inside the design of ChatGPT’s personality.” Interview with Laurentia Romaniuk, Product Manager for Model Behaviour, OpenAI. Published 26 May 2026. https://www.thedeepview.com/articles/inside-the-design-of-chatgpt-s-personality

2. OpenAI Help Centre. “Customizing Your ChatGPT Personality.” Accessed 27 May 2026. https://help.openai.com/en/articles/11899719-customizing-your-chatgpt-personality

3. OpenAI Help Centre. ChatGPT Personalization and Memory documentation. Accessed 27 May 2026.

4. OpenAI. “Sycophancy in GPT-4o: what happened and what we’re doing about it.” Published 29 April 2025. https://openai.com/index/sycophancy-in-gpt-4o/

5. OpenAI. Model Spec and related model-behaviour documentation. https://model-spec.openai.com/

6. Character.AI. “Taking Bold Steps to Keep Teen Users Safe on Character.AI.” Published 29 October 2025. https://blog.character.ai/u18-chat-announcement/

7. Associated Press. “Character.AI to ban minors from using its chatbots.” Updated 29 October 2025. https://apnews.com/article/characterai-kids-minors-18-ban-chatbot-5d203e9f22c62c153936ccc776a0ed09

8. Giles, D. C. “Parasocial interaction: A review of the literature and a model for future research.” Media Psychology, 4(3), 279–305, 2002.

9. Pentina, I., Hancock, T., and Xie, T. “Exploring relationship development with social chatbots.” Computers in Human Behavior, 140, 107600, 2023.

10. Anthropic. Claude character and model-behaviour materials. Accessed 27 May 2026.

11. Meta AI. Llama responsible-use and open-weight model governance materials. Accessed 27 May 2026.

Author workflow disclosure

This article was produced through an AI-assisted but human-directed workflow. AI support was used for accessibility assistance, article structuring, language refinement, source-discovery prompts, revision planning, and conversion of editorial comments into specific amendments. The author retained responsibility for the argument, accepted or rejected suggested changes, checked the logic of the claims, and remained accountable for the final text. AI-generated material was not treated as empirical evidence, and synthetic or illustrative examples were not presented as observed data.

Image note

The image accompanying this article is AI-generated and is intended for illustration purposes only.

CONTACT

© 2026 Dr Danie Adendorff. All rights reserved. Rumbls.com is an independent analytical blog.

TERMS AND CONDITIONS

DISCLAIMER

COOKIE POLICY

PRIVACY POLICY