Can You Really Trust AI? Spotting Hallucinations and Bias

If you’ve been anywhere near a Wi-Fi signal in the past two years, you’ve likely heard about the uptake of AI. Boardrooms are buzzing about upskilling, efficiencies, and how generative AI will transform our work with tailored, more accurate information than a search engine could ever provide.

At least, that's the dream.

In a high-pressure economy where time is short and expectations are high, AI has become a quiet necessity, embedded in workflows whether overtly or covertly and embraced as an easy fix.

The C-suite isn’t the only one feeling the shift. Entry-level roles are being reshaped as AI takes over tasks like data cleaning, draft writing, and admin, the so-called “professional rites of passage.” And it’s not stopping there. 

In almost every industry and facet of life, people are turning to AI for advice far beyond spreadsheets and scheduling. From mental health support to career guidance, chatbots are increasingly acting as stand-in experts, a risky role when the tech still makes basic mistakes. 

What are these mistakes?

These most often fall into two categories:

AI “hallucination” is when a system like ChatGPT or Gemini confidently presents false or irrelevant information as fact. Sometimes it’s an outright fabrication. Other times, it’s factually correct but contextually wrong. Either way, the output can sound polished, convincing… and completely unhelpful.

AI bias arises when the system’s training data overrepresents some perspectives while leaving others out. That bias, rooted in human decisions and historical inequities, can quietly embed itself into the algorithms, skewing results and deepening harm. From hiring and policing to credit scoring and healthcare, biased data collection has long mirrored societal inequity. AI simply scales it.

Why This Happens

When the most advanced AI can’t “understand” the full totality and nuance of context, it acts much like a human and tries to fill in the gaps. Predicting the next likely word or context based on patterns in its training data. When that data is incomplete or skewed, the system fills in the gaps with assumptions. This can lead to:

• Factual errors - invented statistics, wrong dates, or fabricated sources.

• Contextual errors - advice that doesn’t match the user’s cultural or situational reality

• Overfitting - confidently producing answers that overly rely on narrow or unrepresentative sources presented as fact

  
  

OpenAI has said its latest model is six times less likely to hallucinate than its predecessor, claiming it can “proactively flag potential concerns” based on existing conversations. Yet, hallucination when AI simply makes up a response rather than admitting it doesn’t know is still a live risk, especially as more people turn to chatbots for sensitive areas like healthcare.

This is where existing biases in human knowledge collide with AI’s limitations. For example, in the UK, much of the foundational research in medicine and biology has relied on study groups made up primarily of white, often male, people of European ancestry. These studies underpin medical textbooks and training, yet they provide a one-size-fits-all picture of health that can miss vital differences. A well-known example is the use of images of white skin to diagnose dermatological conditions across all patients, even though some conditions appear differently on darker skin. Without diverse and representative inputs, AI trained on this data risks giving incomplete or unsafe medical guidance.

The same applies to social and legal contexts. A Stanford-Oxford study found that large language models still exhibit “dialect prejudice,” applying negative traits such as “dirty,” “lazy,” or “stupid” to speakers using African-American English (AAE), and linking them to less prestigious jobs or criminality. This isn’t overt hate speech; it’s the subtler, insidious bias that can seep into hiring algorithms, legal assessments, and healthcare decisions without detection.

As AI moves deeper into healthcare, recruitment, financial services, and justice systems, these flaws will matter more, and they will disproportionately harm communities already facing systemic inequities.

What can we do?

Bias and hallucination are not rare glitches; they are structural challenges in AI design. While progress is being made, responsible implementation demands ongoing vigilance. If your business uses AI or is considering it, consider the following:

• Data sources: Are they regularly updated? Where does the AI draw its data from? Has the dataset been evaluated for bias against marginalised groups?

• Transparency: Does the system clearly explain how it reaches conclusions?

  
  

Navigating a rapidly evolving industry can feel overwhelming, especially when the inner workings of these tools are unclear. Approaching AI with critical awareness allows us to engage more confidently and actively encourage the mitigation of bias and hallucination.