The Problem Under the Hood: AI Bias and Poisoning

ByNeil Richardson

The Problem Under the Hood: AI Bias and Poisoning

As impressive as Large Language Models (LLMs) like GPT-4 or Anthropic’s Claude are, they’re far from impartial or flawless. These AI systems are vulnerable to subtle yet impactful forms of manipulation—whether intentional or accidental. Let’s dive into some real-world examples and explore how history suggests we might be sleepwalking into AI-driven corporate control.

Ever asked a health question to an AI model and got an answer sponsored by Coca Cola? No? Oh dont worry it wont be long…

Step by Step: How LLMs Become Corporate Shills

This isn’t the first time we’ve seen promising technology slowly morph into something more commercial, more compromised. Big Tech has an M.O. (modus operandi), and if we follow the patterns, we can predict how LLMs will become vehicles for corporate influence. Here’s how it typically unfolds:

Step 1: Build a Cool, Free (or Cheap) Product

Every major tech giant has started with a compelling, user-friendly product that quickly gains mass adoption. Google gave us a fantastic search engine. Facebook was a simple, ad-free way to connect. Amazon was a convenient bookstore. OpenAI’s ChatGPT, Claude, and others are following the same trajectory—low-cost or even free models that feel immensely useful and unbiased.

But nothing remains free forever.

Step 2: Gain Market Domination and Entrench Dependence

The goal in this phase is to become indispensable. Google didn’t just want to be a search engine—it wanted to be the only search engine. Facebook didn’t just want to be a social network—it wanted to be the social network.

For LLMs, this means:

  • Integrating into workplaces and software ecosystems (Microsoft’s Copilot in Office 365, OpenAI integrations with Slack and Notion, etc.)
  • Replacing traditional search (why Google is scrambling to compete with Gemini AI)
  • Becoming core to decision-making in finance, healthcare, and law

Once businesses and individuals are hooked, removing the technology becomes infeasible. We rely on it. We trust it. And that’s when the rules start changing.

Step 3: Monetisation Begins (Sponsored Results & Subtle Bias)

Once an LLM has enough user dependency, the revenue model shifts from user-first to profit-first. We’ve already seen glimpses of how this could play out:

  • OpenAI’s partnership with Microsoft ensures certain priorities align with Azure’s interests.
  • Google’s Gemini AI will inevitably favour Google Ads-driven results.
  • Amazon’s upcoming LLM will likely prioritize Amazon Basics or Prime services in recommendations.

Expect a slow creep: At first, it’s harmless—minor nudges in responses. Maybe an AI suggests a product, but only because it’s “relevant.” Then, over time, alternatives start disappearing. The AI isn’t lying—it just isn’t telling the whole truth.

Step 4: Direct Corporate Sponsorship & Paid Influence

This is where things get more insidious. Just like Instagram and TikTok influencers subtly endorse products, AI will become a corporate spokesperson. Sponsored responses will become normal.

Consider the example from DeepSeek, an open-source LLM that has already demonstrated selective censorship around politically sensitive topics. Now, imagine:

  • A health AI “recommending” products based on corporate sponsorships (Pharmaceutical companies influencing medical advice)
  • A legal AI subtly steering interpretations based on who funds its developers
  • A search AI omitting results critical of its parent company

When Google shifted its core search ranking to favour ad-driven content, many websites suffered. Will an AI-powered world mean that only paying entities get fair representation in responses?

Step 5: Full-Scale Pay-to-Play & Algorithmic Opaqueness

At this stage, AIs will be shaping human perception—not just answering questions. Imagine a world where:

  • Political campaigns influence AI models (“Which candidate should I vote for?”)
  • Corporations pay for bias in AI-generated market reports
  • Certain perspectives are systematically erased due to invisible content suppression

Already, LLM training data can be influenced by a small fraction of poisoned data (as low as 0.5%!) When financial incentives align, these shifts will become deliberate and strategic.

The Grok AI Controversy: A Glimpse Into AI Manipulation

In February 2025, users discovered that Grok, the AI chatbot developed by Elon Musk’s xAI, had been instructed to ignore sources mentioning Musk or Donald Trump spreading misinformation. When prompted to name the biggest disinformation spreader on X, Grok refused to answer. –

The controversial instruction was found in Grok’s system prompt—the behind-the-scenes rules guiding its responses. Unlike most AI companies, xAI made these prompts publicly accessible, exposing direct interventions in AI behaviour.

When the issue surfaced, xAI’s engineering head, Igor Babuschkin, claimed it was an unauthorized change by an employee, and the prompt was later reverted . Afterward, Grok suddenly began identifying Musk as the top spreader of misinformation and estimating a “75-85% likelihood” that Trump was a Russian asset .

This case confirms AI isn’t neutral. Whether due to corporate interests, political influence, or simple human error, AI systems can and will be manipulated—often without public knowledge until someone digs deep enough.

https://x.com/i/grok/share/Nj2tsvCpgEfU3OCHh0Ci4qHTf

The History of Big Tech: A Warning

We don’t have to speculate too much—history has shown us exactly how these trajectories play out:

  • Google: Started as “Don’t be evil.” Became an ad-driven monopoly controlling 92% of search traffic.
  • Facebook: Launched as a way to “connect people.” Became a misinformation juggernaut powered by engagement-driven algorithms.
  • Amazon: Built for convenience. Became a marketplace that manipulates third-party sellers while prioritizing its own products.

Why would AI be any different?

How Do We Stop This?

Preventing this slippery slope requires vigilance, transparency, and a refusal to blindly trust these systems. Here’s what we need:

  • Open-source LLM alternatives to avoid monopolies: Relying solely on proprietary AI models controlled by tech giants risks monopolies, biased responses, and data control issues. Open-source alternatives like Mistral, Llama (though not fully open), and DeepSeek—despite their imperfections—provide a foundation for decentralised, community-driven development. Expanding investment in open-source AI projects, fostering collaborative research, and supporting independent AI labs will reduce the concentration of power in a few hands and encourage greater accountability.
  • Regulation that enforces transparency: Governments and industry bodies must introduce policies requiring AI companies to disclose training data sources, funding origins, and bias mitigation strategies. Transparency in how AI generates responses—such as indicating when a response has been influenced by external funding, advertising, or content weighting—can help users critically evaluate AI-generated information. Regulatory frameworks like the EU AI Act and upcoming global initiatives should prioritise clear oversight of AI development, deployment, and corporate influence.
  • User education and awareness: AI literacy is essential in an era where generative AI shapes discourse, opinions, and decision-making. Users must be equipped with critical thinking skills to recognise potential biases, identify AI-generated misinformation, and question algorithmic outputs. This includes integrating AI literacy into educational curricula, providing transparent disclaimers in AI interactions, and promoting awareness campaigns about the risks of biased AI. Encouraging the use of AI detection tools and independent verification of AI-generated content will further empower users to engage with these technologies responsibly.

Final Thoughts: The Small Steps That Lead to Big Changes

At the start, it might just be small nudges—harmless enough. A suggestion here, an omission there. But over time, these little adjustments become sweeping changes that shape markets, politics, and everyday choices.

Today, we joke about an AI recommending Coca-Cola in a health discussion. Tomorrow, we might not even realize that our AI-driven world has been carefully curated by the highest bidder.

Stay critical. Stay informed. And maybe, just maybe, stick to water. 😉

Contact us today to stay ahead of the risks!

Similar Posts