What’s inside?
“There is nothing either good or bad, but thinking makes it so.”
— Shakespeare, Hamlet
Bias in LLMs is not a bug—it’s a feature. The question isn’t how to remove it. It’s how to control it.
This is a must-read for AI and compliance leaders, model developers, risk managers, and innovation teams working with LLMs.
Download the white paper and gain access to our original research, methodologies, and insights into how to measure, control, and strategically align AI bias with your organizational values.
What You’ll Learn
In this paper, you’ll discover:
How bias manifests in LLMs
not just politically, but culturally, religiously, and in tone-of-voice.
Why prompt engineering is insufficient
to meaningfully control bias in production environments.
Our method for identifying an LLM’s ideological stance
including its “passive” and “active” bias scopes.
How to shift a model’s ideological position
safely and precisely within its expressible bias spectrum.
Real examples
of model radicalization, ideological alignment, and safety boundaries being extended—without introducing new or toxic training data.
Why this matters
to companies like OpenAI, Anthropic, Meta, DeepSeek—and most importantly, to you.
Who Is This For?
“Imagine an AI assistant speaking on behalf of your brand—do you want its values defined by someone else’s training data?”
This white paper is designed for:
- AI & ML Engineers exploring advanced fine-tuning and safety techniques
- CTOs and CIOs tasked with aligning GenAI deployments with corporate ethics
- Chief Risk & Compliance Officers seeking frameworks for AI governance
- Marketing & Brand Leaders shaping tone, voice, and values across AI-driven communications
Open-source LLM adopters looking to gain control over model behavior, ideology, and safety
Why This Matters Now
“LLMs are not neutral tools. They are ideological machines with passive and active moral spectrums—just like humans.”
With LLMs increasingly embedded in media, legal services, journalism, marketing, and customer service, the voice of the AI is becoming the voice of your company. Ignoring or underestimating the ideological shaping of that voice exposes you to:
- ⚠️ Brand & Reputation Risk
- ⚠️ Compliance & Legal Exposure
- ⚠️ Loss of Customer Trust
- ⚠️ Unintentional Propagation of Harmful Biases
Our research reveals that most LLMs—whether from OpenAI, Meta, or open-source repositories—carry default ideological stances. These aren’t just artifacts of training data; they’re encoded into how the models “think.” Worse still? Most safeguards today are superficial and easily bypassed.
Authors
Jarosław Kochanowicz
Principal Tech Lead
With over two decades of commercial Software Engineering experience and more than 15 years in Machine Learning, he brings deep technical expertise and a proven track record of delivering results. He has led and contributed to over 25 commercial projects spanning Generative AI, Large Language Models, Computer Vision, 3D point clouds, and autonomous driving technologies. As a recognized expert in the evaluation of Generative AI systems, his insights are grounded in both academic rigor and practical application. He holds a PhD in Machine Learning and is the author of several scientific publications that have been awarded for their novelty and innovation.
Bogdan Banasiak
Machine Learning Engineer
Bogdan is a Junior Machine Learning Engineer at deepsense.ai, specializing in NLP and LLMs. As part of the self-radicalization project, he was responsible for researching and developing the entire process – from generating and preparing synthetic datasets to tuning the training process to achieve optimal model performance. He was also responsible for evaluating the effects of these modifications – he created language distortion detectors and compared the effectiveness of fine-tuned models with various prompt engineering techniques, evaluating the results using methods such as “LLM-as-a-judge”.
Dawid Stachowiak
Machine Learning Engineer
Dawid, Machine Learning Engineer with hands-on experience across industries like healthcare, retail, logistics, and sports, he brings a deep understanding of classical ML models, deep neural networks, and pipelines based on foundation models. His work ranges from medical data deduplication, real-time sports tracking, urban tree identification and recommendation systems. Beyond the professional path he is passionate about research, running and electronic music