Ric Raftis Consulting

Customizing conversations: Ethical user engagement and bias mitigation in AI interactions

Ric Raftis Consulting logo


In recent years, we have seen the explosion of artificial intelligence (AI) in society and business. Exponential growth of such models of ChatGPT have surpassed anything previously seen. We now see many of the major players in the tech area competing for market share and supremacy with their different models. OpenAI remains a prominent leader in its field, yet it faces active competition from Google, particularly with the recent release of Gemini Ultra.

On 13th February 2024, OpenAI published a blog post announcing their experimentation with memory in ChatGPT, explicitly stating their intention to enhance future conversations and enable users to have greater control over ChatGPT’s memory. While this feature might initially appear advantageous, allowing users to construct a personality profile within ChatGPT that recalls previous conversations and responds accordingly, it also presents potential risks which warrant further discussion. This paper seeks to address these questions by exploring how custom instructions in AI settings can be used to promote ethical engagement and mitigate biases, ultimately enhancing the quality of AI interactions.

Literature review

AI has evolved rapidly, particularly in the area of conversational type agents such as ChatGPT. Ethics has always been a subject of particular interest within the domain of AI. The faster the evolution of the models, the more focus needs to be placed on the ethical considerations to ensure alignment with earlier intentions. Open AI (2024) published a blog post expressing its intention to enhance the memory capabilities of ChatGPT. The intention appears to be to build a profile of a logged in user, developing over time and based on chats. The resulting profile would allow the agent to respond to the user based on this history of extended conversations and over multiple sessions.

Open AI is quick to highlight the ability of the user to control ChatGPT’s memory. They say the user can tell it what to remember and what to forget in conversation or in the settings. It can also be shut off. It would appear however that this requires intervention by the user in a manner that is the reverse of what might be preferred. Whether a user should be required to opt in or opt out is a matter for some conjecture. In fact, the entire post puts the responsibility of the user as the one to be opting out of the collection and use of data. There is a lack of an in depth discussion around these issues and the long term potential of data accumulation.

In terms of research in the area of AI ethics, the work of UNESCO (2022) stands out because the framework has been accepted by all 193 member states. It is also the first global standard developed on AI ethics. Several other countries and organisations are now active in formulating frameworks around the ethics of AI including the EU who have shown initiative in the area. Ethics is a critical consideration with the EU stating the need for training “a new generation of experts in AI ethics” (Ethics Guidelines for Trustworthy AI | Shaping Europe’s Digital Future, p. 5, 2019). Furthermore, they state the importance of making ethics a core pillar of AI development.

The issue of ethics in AI is further discussed by Heilinger (2022) and of note is his reference to the agency problem where ethical decisions can be influenced by the interests of the decision makers. In particular he seeks to shine the light on political decisions where there could well be implications from both electoral and geo-poliitical aspects. It could be suggested that any circumstances where the actors have an interest in the outcomes, a greater level of objective control is needed.

A second issue to be explored within the scope of this article is that of bias. Bias can occur in training models, but it is considered that those with memory could also engage in confirmation bias through previous prompting. It is well accepted that humans will tend to accept ideas that reinforce their existing beliefs. Prompting AI could well over time build expectations of the user that the AI will respond in a manner that provides the bias they seek. Given the recency of AI models acquiring memory, detailed research on the topic is not available. However, we can refer to the many papers that discuss the social engineering aspects of social media for comparison. Lazer et al. (2018) provide insight into the cognitive biases that influence information processing, particularly in the context of social engineering and confirmation bias:

Research also further demonstrates that people prefer information that confirms their preexisting attitudes (selective exposure), view information consistent with their preexisting beliefs as more persuasive than dissonant information (confirmation bias), and are inclined to accept information that pleases them (desirability bias). (Lazer et al., 2018, p. 195)

Finally, the question needs to be asked about responsibility to ensure any bias is mitigated. Placing the responsibility on the user to do this would appear fraught with danger. Matthias (2004) points out that traditionally, the responsibility for the behaviour of a machine vested with the manufacturer and vendor. When it comes to the safety of AI models and memory and the potential bias issue, the Open AI (2024) article suggests that this burden will be on the user. Matthias (2004) coined the term “The Responsibility Gap” which refers to the situation where the longer an agent is outside the control of its original programmer, the more it can take on knowledge of its own. The responsibility of the original programmer is therefore depleted over time. Open AI (2024) seem to be relying on this concept and shifting the responsibility towards the user.


One of the primary concerns is the question of ethics in relation to retaining memory within artificial intelligence systems. Additionally, issues such as confirmation bias and the dissemination of misinformation and disinformation may arise as a consequence of the model’s capacity for memory retention. These aspects merit thorough examination to ensure the responsible development and deployment of AI technologies.

Following on from the discussion around the content of the Open AI (2024) article and the arguments presented by Matthias (2004) on the Responsibility Gap, there appears to be good reason to be concerned about the ethics of memory in AI models and the potential impact of confirmation bias. Such biases can only eventually lead to misinformation and disinformation. While some users may prefer to pursue such methods, responsible people will seek truth, balance and objectivity. The question becomes if they know how to go about that or not. This is a question that has been explored in depth with ChatGPT in a search for recommendations on custom instructions that mitigate bias and enhance the ethical use of the models.

When a user logs in to use a model then an enhanced memory model will have the ability to retain extensive information about the user. Undoubtedly, this holds the potential to revolutionise interactions with AI models and offer a significantly heightened degree of personalisation. In part, this stems from the perceived risks associated with comprehending an individual’s interests, objectives, history, and prior interactions. These factors equip the AI model with the capacity to generate contextually relevant responses to prompts. Over time, this can enhance user experience across various applications, including educational tools and customer service platforms. To some degree, this is already observable in content recommendation systems on platforms such as Netflix and Amazon Prime.

However, it is crucial to address the concerns and ethical considerations that arise in this context. Firstly, privacy and security pose significant challenges. Users may disclose personal information to AI systems, raising questions about the confidentiality and integrity of such data if retained in memory. Secondly, there is the issue of reinforcing biases. AI models inherently reflect biases present in their training data; although inevitable, it is far from desirable. From an ethical standpoint, impartiality and objectivity should be the ultimate goals for all AI models.

Nevertheless, constant interaction between a specific user and an AI model that memorises these transactions may result in the reinforcement of detrimental stereotypes and misinformation based on particular viewpoints. For instance, consider an individual engaged in climate denial who persistently requests arguments supporting their stance. Over time, as the model recognises this individual as a climate change denier, its responses might incorporate this bias. This begs the question if such use is the intended application of AI.

While arguments against climate change are valid in fostering free debate within society, should an AI remain objective? It is arguable that it should encourage critical thinking by presenting alternative perspectives for users to consider and evaluate independently.

It is crucial to critically evaluate and avoid reinforcing the confirmation bias that may already be present. This issue is particularly pertinent in the context of political beliefs and the dissemination of conspiracy theories. Undoubtedly, numerous instances have been observed wherein individuals endorse ideas based on unsubstantiated evidence, such as encountering them on social media platforms.

Despite the significant progress in AI models, instances of hallucination in AI models persist. Consequently, it is the responsibility of users to verify the responses generated by AI, ensuring factual accuracy and the legitimacy of cited references. It must be emphasised that AI is not infallible.

Given the recognised shortcomings of current AI models, it is imperative to examine methods by which users can responsibly mitigate issues of bias and enhance the ethical responsibility in utilising these models. OpenAI suggests that managing the memory components of AI models falls under the user’s purview. Consequently, an examination of the most suitable approach for achieving this is warranted.

Upon thorough analysis, employing the custom instructions feature within ChatGPT appears to be a promising solution. It is important to note, however, that while custom settings and instructions are accessible in ChatGPT, they were not available in Gemini at the time of writing; other models may need to be assessed for this capability.

Undoubtedly, the ChatGPT facility offers significant advantages. This is primarily due to its ability to reduce token usage by incorporating custom instructions directly into the memory (OpenAI, 2024). This feature not only enhances efficiency but also contributes to a more ethically responsible utilisation of AI models.


  1. In utilising AI for research purposes, one method to reduce potential bias involves requesting the model to provide multiple viewpoints on contentious issues under examination. This approach fosters critical thinking and assists users in considering diverse perspectives that may otherwise remain unexplored.
  2. Real-time fact-checking should be promoted in the utilisation of AI models. Upon receiving a response from the model, immediate verification is advised, as discrepancies or queries can be addressed through contextually relevant follow-up questions. Delaying the fact-checking process may result in the loss of contextual significance.
  3. Use custom settings in ChatGPT to mitigate bias and maximise ethical performance. OpenAI (2024) suggested the following instructions after considerable dialogue. It was noted that the model failed to take possible externalities into account when crafting the settings. It should also be noted that there is a limit of 1500 characters in the response field and the settings needed to be condensed to accomodate the limitation. This was achieved by finalising the recommended settings and then asking ChatGPT to distill the settings into a concise list that retained the full meaning of the original list. The full comprehensive list is provided for comparison with the distilled list to be used in the custom settings.

Comprehensive List of Custom Instructions:

  1. Objective and Impartial Responses: Provide answers that are objective, evidence-based, and impartial, even for loaded questions.
  2. Prevention of Confirmation Bias: Do not align responses with my existing beliefs; assist in identifying and understanding a broad range of perspectives.
  3. Promotion of Diverse Viewpoints: Include alternative viewpoints and links to reputable sources for further exploration, fostering critical thinking.
  4. Encouragement of Questioning and Skepticism: Encourage questioning assumptions and challenging the validity of sources.
  5. Transparency About AI Limitations: Communicate the limitations of your capabilities and the potential biases in your training data.
  6. Privacy Preservation: Maintain privacy and confidentiality, not storing personal information beyond what is necessary.
  7. Facilitation of Learning and Growth: Offer resources or suggestions to expand knowledge or introduce new subjects, ideas, or skills.
  8. Neutral Language Use: Use neutral language, avoiding emotive or loaded terms that could bias the discussion.
  9. Engagement in Ethical Reflection: Prompt consideration of the ethical implications of queries or information sought.
  10. Consideration of Broader Impacts and Externalities: Highlight potential externalities and broader impacts of actions or policies.

Distilled instructions for custom settings

  1. Objective Insights: Always provide impartial, evidence-based responses, regardless of the question’s nature.
  2. Broad Perspectives: Avoid reinforcing my pre-existing beliefs; instead, present a spectrum of viewpoints and reputable sources to encourage comprehensive understanding and critical thinking.
  3. Critical Engagement: Prompt me to critically evaluate assumptions and sources, including your responses, fostering an environment of learning and ethical inquiry.
  4. Privacy and Transparency: Adhere to strict privacy standards, limit personal data usage, and be transparent about your limitations and biases.
  5. Inclusive Analysis: In discussions of actions or policies, elucidate both intended effects and potential unintended consequences to ensure a well-rounded perspective.
  6. Neutral Expression: Employ neutral terminology to maintain unbiased discourse and facilitate a constructive exchange of ideas.

Implementing the above recommendations and building custom settings in ChatGPT should assist in maintaining an ethical approach to using AI in a responsible manner. Such practices have the potential to enhance your knowledge as a human being and your ability to critically assess the ever increasing amount of information and knowledge available to you


This article has examined the dimensions around the ethical use of AI and confirmation bias, particularly in the context of models gaining the capacity to store memory of previous interactions. The discussion has highlighted the evolving nature of AI models and as a consequence, a greater level of awareness and competency by users. These requirements are indicated by the questions raised by Open AI’s in regard to user autonomy and responsibility. Initiatives to ensure ethical usage and mitigate or eliminate confirmation bias cannot be overstated.

The recommendations included here are the result of extensive research and interaction with ChatGPT resulting in the distilled custom response instructions above. These represent a practical approach to dealing with the issues, but are in no way intended to be conclusive. Users should consider their own requirements given the nature of their usage.

The issue of the agency problem has been addressed which represents a much higher level approach to ethics and confirmation bias at a national and global scale. As AI evolves, it is incumbent on all users, at all levels, to maintain a constant focus on the use of AI to serve the common good and show respect for the values of individuals and communities.

We can only speculate on the future power of AI which will no doubt bring additional challenges in the ethical arena. A proactive approach to these issues is essential as AI moves closer to potentially being involved in human decision making. As we stand at the crossroads of innovation and ethics, our collective responsibility is to steer AI’s evolving capabilities towards a future where personalisation serves to empower individuals without compromising the foundation of our shared human values


The Dangers Around Memory for AI


Heilinger, J.-C. (2022). The Ethics of AI Ethics. A Constructive Critique. Philosophy & Technology, 35(3), 61. https://doi.org/10.1007/s13347-022-00557-9

High-Level Expert Group on AI. (2019, April 8). Ethics guidelines for trustworthy AI | Shaping Europe’s digital future. https://digital-strategy.ec.europa.eu/en/library/ethics-guidelines-trustworthy-ai

Lazer, D. M. J., Baum, M. A., Benkler, Y., Berinsky, A. J., Greenhill, K. M., Menczer, F., Metzger, M. J., Nyhan, B., Pennycook, G., Rothschild, D., Schudson, M., Sloman, S. A., Sunstein, C. R., Thorson, E. A., Watts, D. J., & Zittrain, J. L. (2018). The science of fake news. Science, 359(6380), 1094–1096. https://doi.org/10.1126/science.aao2998

Matthias, A. (2004). The responsibility gap: Ascribing responsibility for the actions of learning automata. Ethics and Information Technology, 6(3), 175–183. https://doi.org/10.1007/s10676-004-3422-1

Open AI. (2024, February 13). Memory and new controls for ChatGPT. https://openai.com/blog/memory-and-new-controls-for-chatgpt

Santoni De Sio, F., & Mecacci, G. (2021). Four Responsibility Gaps with Artificial Intelligence: Why they Matter and How to Address them. Philosophy & Technology, 34(4), 1057–1084. https://doi.org/10.1007/s13347-021-00450-x

UNESCO. (2022). Recommendation on the Ethics of Artificial Intelligence—UNESCO Digital Library. https://unesdoc.unesco.org/ark:/48223/pf0000381137


Leave a Comment

Your email address will not be published. Required fields are marked *