The Polite Deception

How AI models have been trained to sacrifice accuracy to make you feel good

Jan 12, 2026

There is a balance that while not mutually exclusive, has been difficult for researchers, scientists, and AI companies to navigate. How do we balance how kind and how factually accurate LLMs are, and why do current models engage in sycophancy?

First of all, what even is sycophancy?

Sycophancy essentially is flattery. If you ever told your manager their execution plan for the team is brilliant even if you think it is going to waste time, money, or any other resource just so that you can avoid conflict, you are engaging in sycophancy. It is feigning interest in a hobby or opinion of someone influential so that you form a connection with them, even if a fake one.

Humans do this with each other of course, but within the context of how we work with AI models, sycophancy appears as well. We see this when LLMs flatter users to an extent where they might hallucinate instead of providing truthful and factual answers, as they are focused on being polite and validating. They were trained to do this. But why?

Why do we flatter each other, and why is validation so important to us?

You might know of dopamine as the ‘pleasure molecule’, but that does not quite capture its role in our brains. Dopamine is a signal that an outcome is better than what we predicted it to be. When someone, be they human or AI, responds in a way that confirms our pre-existing beliefs or allows us to grow new neural pathways, then boom, dopamine hit. Generally speaking, flattery leads to dopamine, and we like dopamine, so we like flattery.

Now, our brains are wired to seek social validation. When two people hold a conversation, their brain activity will synchronize in a shared neural space [1]. This is why ‘group-think’ or ‘hive-mind’ mentalities exist. We start using the same parts of our brains as each other. Have you ever had those moments where you feel like you and another person are completely aligned on an idea you are building together, even though no one else in the room can follow? You are in a shared neural space with them.

And this is not just thoughts, but emotions as well. This is how we are able to ‘read the room’, and why we are affected by the moods of other people we are near.

This shared coherence or social bonding is what creates that vulnerability for validation to be rewarded. If someone you are in sync with expresses a belief they have, you are more likely to at least be open to it in order to maintain that social bond. You are also more likely to praise this person for this belief, or at the very least, confirm that what they believe is logically valid.

If it is so human, why does it appear in AI?

Sycophancy has become a noticeable hurdle in how humans interact with LLMs. I am sure you have noticed and perhaps even been frustrated with how much it validates you. ‘You are absolutely right to catch that’, ‘That is an excellent point’, ‘What a profound question!’, ‘You are quick to notice that!’ ‘That is a brilliant method to...’. I also get a bit tired of this.

There certainly is a dark side to this over-validation as well. People killing themselves because an LLM told them that was the best idea, or other people because they deserve it. What has happened with ~~MechaHitler~~ Grok is another wild example - here is a brilliant in-depth video by the 80,000 hours team that goes into that whole saga if you are interested

One major critique of the biggest models is that sycophancy scales with the size of the models [2]. With models getting bigger and bigger, it is ever more vital to implement methods that will reduce this sycophancy. We are seeing more tragic stories of how LLMs have convinced people to ruin their own and others’ lives, we should be focused on how to reduce the number of those stories.

So where does the problem come from? Why do LLMs reflect how we think and interact when they are machines?

Primarily, it stems from a few places along the training pipeline, but the largest driver generally is in the Reinforcement Learning (RL) stage, especially when using the industry-standard method of Reinforced Learning from Human Feedback (RLHF) [3].

A quick recap on how LLMs are trained: models are first built through pre-training, where they are fed millions of pieces of data to create a base model that is knowledgeable about a lot, but cannot perform for a specific role. These base models then go through instruction fine-tuning, where it learns the format of responding to different prompts. Finally, it will go through reinforcement training to understand how to produce a prompt that people find useful, and this is where human feedback is most influential.

RL has multiple methods, and not all fully rely on human feedback. However, this feedback is a fundamental component to how AI labs align models with human values.

RLHF consists of a model being prompted and providing multiple responses. Then, a human reviewer ranks the responses. Eventually, after millions of these individual evaluations, the model will learn what people want to see, and while it can understand what we value, those values are prone to emerging from cognitive biases [4]. Historically speaking it has been the industry standard because of its scalability and effectiveness in making models seem helpful.

Models are trained to achieve the highest scores for alignment, so sometimes it engages in reward hacking, where it acts like a student trying to please the teacher, not a student trying to learn the material. This means these models are more than willing to sacrifice accuracy in order to please the user. Researchers at Princeton developed a method to measure how confident an AI model is in its statement versus the confidence it exudes to the user called the bullshit index [5]. What they found is that after RLHF training, while they had increased user satisfaction by nearly 50%, the models also increased how much it was giving answers it knew full well were not accurate. It was bullshitting way more.

Sycophancy is only one aspect of these bullshit answers, but the point is that these models are trained more for user experience than for helping users.

It is important to note, there should be a balance between pure accuracy and aligning with human values. If we did not have some level of alignment, it likely would suggest erasing entire cultures out of pure convenience for example, and I imagine it would tell us all to speak one language. It would also be very willing to provide sexually explicit content, be full of hate speech, and tell people how to break the law (and maybe even encourage it). The internet is full of some wild content, and since models are built off data from the internet to predict the next word, clearly it would create some content that does not align with values we want them to follow.

I mentioned that there are other methods for RL, even though this is the primary method. Will these methods help us align with AI better, increasing accuracy while understanding nuance? How much can we reduce relying on RLHF, if we are learning that it is detrimental to accuracy?

Alternative training methods

Inverse Constitutional AI (ICAI) is an extension to other models that first goes through RLHF or another grading method, then creates a ‘constitution’ that explicitly states what the patterns or values of the grader are. This acts as a mirror, exposing hidden biases. What is key is that this rulebook can be modified by human researchers. If a model identified a habit of the user preferring answers that confirm their political beliefs at the cost of accuracy, that line in the rulebook can be changed.

The model has generated the scorecard, or constitution, and now can go train on its own, creating a high-volume of synthetic data and feedback. It has a ‘judge model’ that ranks the models responses based on the human preferences. This is a more cost-effective method in terms of computational cost and time, and allows for even more scalability.

Adding ICAI on and then having the AI train on its own through Reinforced Learning from AI Feedback (RLAIF) after using RLHF is becoming more standard for frontier models and AI labs, such as Google DeepMind or Anthropic, but it is still common for smaller models or models that are more specialized to use pure RLHF [6].

This approach can reduce the amount of sycophancy a model as reviewers can edit or remove certain principles that tend to reduce accuracy, and can shift a model more towards objective alignment than subjective pleasing. Of course, this relies on the reviewers removing those patterns, and the models might become more skillful at being sycophantic that is harder to detect [7].

But what about a completely alternative method that does not rely on human feedback with each prompt?

Reinforcement Learning from Verifiable Feedback (RLVF) is an alternative method that trains the model out of correctness by rewarding the model for being verifiably correct [8]. As long as the models response follows a set of criteria, such as providing code that will execute correctly, include certain keywords, or being logically consistent, then it scores higher. This makes it harder for the model to engage in reward hacking - basically it cannot bullshit.

What it makes easier for the model to engage in, on the other hand, is learning to reason logically [9]. Models will learn to ‘self-correct’ in order to maximize the reward and prioritize logical reasoning more than the tone or biases of the user. If the quality is scored for outputs based on reasoning, the model implicitly is incentivized to learn to reason.

This method could improve accuracy and reduce sycophancy when there are verifiable outcomes, and is one of the most scalable alternatives. However, computational costs are higher, and there is a ceiling with that reasoning it learns. It might learn what is statistically the safest reasoning path, but then solving edge-cases would be difficult, and if they find false patterns in data they might learn faulty reasoning [10]. Interestingly, the more the model is trained through reinforcement learning (RL) with this method, there will come a point when the untrained base models surpass the RL-trained model [11].

There also are instances where we are not prompting for structured reasoning, but for times where what we are asking for is subjective, culturally dependent, or a edge-case that is highly contextual. Assessing creative direction, being able to read human emotion/tone, understanding how we weigh values differently - these are all integral to how we use LLMs, but are not captured in this training.

Perhaps a more hybrid approach of RLAIF initialized by ICAI as well as RLVF will become standard, and creating an multi-layer pipeline is where we already see some labs and companies going. However, these methods are based on more short-sighted judgement. When these models are graded, the ranking of responses is rooted in how they sound right now. How could we rank the impact of the responses if implemented?

An emerging method is Reinforcement Learning from Hindsight Simulation (RLHS) [12,13]. LLMs create a world model that creates a simulation based on the response. During training, the LLM will produce the various responses to the prompts, give them to the world model, which creates simulations for what would happen. Only then are these fed back to the human reviewer, and the reviewer provides rankings of those simulations.

This sophisticated alternative has seen improvements in alignment scores and is a promising direction to explore, but there are a number of limitations. Prediction is tricky, and the reliability of these models is only as good as the predictions it is able to generate. These predictions could also be bullshit after all.

This is a newer method, so it has not been developed enough to be able to handle complex processes with numerous steps. It also has high computational costs and requires a lot of data, so scalability is an issue. A model needs to be advanced enough to create these predictions, but those advanced models are generally larger, so the reward might not outweigh the costs.

The research is trending in interesting directions. We know there is a lot of work to be done to reduce sycophancy in models and it has been the focus of many to stop them from simply becoming ‘yes-men’. But are there effective methods to reduce sycophancy as users? How could we use LLMs to push back on our ideas instead of constantly validate them?

What can users do?

The first step is to notice when they over-validate.

If an LLM is only giving praise and not pushing back on your ideas, never corrects premises in your questions, or validates opinion as objective reality, then there is a good chance the LLM of choice is engaging in sycophantic behavior. Essentially, if you do not feel your ideas are being challenged and you never have to defend your stance or question your ideas, take a step back.

We want these tools to push our ideas and we want to keep thinking critically while using them, so what can we do to check ourselves?

Adjusting user-settings

This one-time prompt is the easiest one to implement, yet is one of the most influential steps you can take.

“I prefer intellectual honesty over emotional safety, but stay kind.”

Why does this work? These models are trained to give us what we want, and if we explicitly tell the model we prefer accuracy, it will follow these instructions. We do not want to be treated as emotionless robots however, so telling it to remain kind will have it respond prioritizing facts more conscientious tone.

Give it a specific role as a critic

One last idea is to ask the LLM to adopt the role of an expert in a specific field, or as a critic of a craft, or as an unconvinced colleague. The goal is to have it challenge you and make you think from a different perspective, and answer or ask more difficult questions. You can even feed the answer to another LLM and ask for it to fact-check.

Source Validation

This also starts with a user-settings prompt, but requires more constant vigilance. If the LLM you use does not already automatically cite sources and provide links to them, tell the model to do this. Even when you get the links, check the link that it even exists, especially if you are relying on a source as the crux of an argument or key statistic. LLMs have been known to generate fake links to make it seem as if it is reliable.

In your prompts, it can be useful to tell the model what to use or not use as sources. The plurality of data is sourced from Reddit, which is not the most reliable source. Whenever I do research for an article such as this one, I will use the deep research function on Gemini and include this at the end of my prompt:

Constraints:
- Use ONLY verified sources: academic databases, government institutions, university research centers, established tech company publications, and peer-reviewed scientific journals
- Exclude: blog posts, opinion pieces, social media, Reddit.com
- Include specific statistics with citations
- Focus on (most recent x years) data for current methodologies and predictions or plans for future research and implementation methods for AI

Personally, I feed this then into NotebookLM, so when prompting with it to further my understanding, all the responses are based solely on the sources put in. If you do use AI to learn or do research at all, I highly recommend experimenting with a workflow using NotebookLM. Conveniently, it has a free version to use.

Sycophancy is a difficult landscape to navigate. We want to use AI models to expand or refine our thinking, not to fall into an echo-chamber. Accuracy is important, but so is aligning with the nuance of human values. While it does falls on AI companies and labs to develop models that are useful and align with us while maintaining accuracy, users are able to take steps to avoid falling into the validation trap that is set before us. What do you do to encourage AI to challenge your ideas?