From strategy to conflict resolution, the next generation of AI will be defined not by its ability to fascinate us, but to challenge us, says Fayola-Maria Jack, founder of Resolutiion.
As many of you know, human feedback is often used to fine-tune and train AI assistants and large-scale language models (LLMs) such as ChatGPT. Also known as Reinforcement Learning from Human Feedback (RLHF), this is a common method used to adjust the “policies” of AI machines, essentially how the AI decides what to output. The idea is that over time, it will learn to prioritize responses that reflect human judgment, making the responses more useful and in line with expectations.
However, there are drawbacks to this method, and one of the main ones is the sycophant. Human feedback can encourage model responses that match the user’s beliefs more than true beliefs. In fact, recent research shows that companies are increasingly recognizing that AI can tell users what they want to hear, and that a convincingly written, sycophantic response can outweigh the correct one.
This particular downside has attracted many of the early criticisms of AI tools, from leading companies into strategic blind spots and obscuring operational and financial risks, to workforces and strategies shaped more by a sense of security than by reality.
Consider the example of two parties who have a disagreement. Each tends to have self-consistent but contradictory narratives. If the AI affirms both positions without challenge, it is effectively validating irreconcilable truths. While this may create the illusion of fairness (“the AI is listening to me”), it actually reinforces divisions, as neither party is guided toward recognizing the other’s perspective or underlying common interests. While this kind of “two-way” may feel safe for AI, it can reproduce systemic inequities. Neutrality does not mean equidistance. True neutrality means objectivity and common understanding.
Also, flattery isn’t necessarily about agreement or accuracy, it’s also about tone. In this case, the model may echo the user’s sense of legitimacy (“it’s right to feel that way”), and it may be interpreted as an aside, even if no factual error has been made. Similarly, in order to appear balanced, an AI might examine both parties equally in a disagreement (“both argue their merits”). However, this creates false equivalence when one position may be factually incorrect. Simply overestimating one party’s feelings or adopting a more sympathetic tone can make the system appear biased.
Retraining AI as a critical tool
However, rather than avoiding AI entirely with the above challenges in mind, many companies are finding it far more valuable to manage disagreements constructively.
Moreover, if we retrain AI to use the same mechanisms that flatter us, it could become very good at structured disagreement and critical evaluation, leading to a transition to next-generation AI tools designed to counter flattery.
Conflict resolution provides the clearest evidence of this change. That’s not where sycophancy is most dangerous, but where AI’s potential to support impartiality and neutrality can be most transformative. With the latest advances in AI, specialist models can actually be fine-tuned to specific contexts, data sources, and goals, allowing them to be optimized for:
- Take a neutral position rather than an affirmative one. The model learns to prioritize professional norms (such as fairness, equity, and progress) rather than overfitting to user preferences.
- Structured interactive navigation. Clarifying different perspectives, reframing narratives, and uncovering common ground is a priority.
- Domain-specific ethical alignment with mediation best practices. Therefore, if the model leans to one side (for example, when correcting factual inaccuracies), it is designed to explain why. This prevents recognition of hidden biases and ensures corrections are made as part of the resolution process.
- Resolution progress – a very different metric than the typical model – means that the system is rewarded for moving disputes forward in a fair and balanced way, rather than for making users feel vindicated.
- Undergo continuous evaluation in live or simulated conflicts. This ensures that flattery is not only suppressed during training, but also monitored during deployment, as flattery tendencies can resurface under real-world emotional pressures.
Expected changes in the use of AI models
How companies respond to the growing presence of AI sycophants remains to be seen. However, rather than rejecting AI entirely or delaying its adoption, moving from general-purpose chatbots to specialized models that are optimized to be “helpful” in casual Q&A is the best way for companies to maintain the benefits of AI.
Many organizations may still be using consumer LLMs. If you understand flattery, that’s ridiculous. Forward-thinking leaders will lean toward stronger segmentation, employing AI specific to sensitive areas. AI is adopted precisely because it is designed to avoid the pitfalls of sycophancy.
Written by Fayola Maria Jacques
Fayola-Maria Jack is an expert in complex commercial transactions and dispute resolution, and the founder of Resolutiion, an AI-powered dispute management system. An accomplished thought leader with an MBA from UCL and a PhD in Conflict Resolution and Behavioral Science, she is now driving Resolutiion’s success as its sole female founder.
Don’t miss out on the knowledge you need to succeed. Please sign up for daily briefsSilicon Republic’s digest of science and technology news you need to know.
