Trending Articles
Trending Topics
Trending
This article is part of our Opinions section, where we invite industry professionals to share their views on the most pressing technology questions of our time.
AI has always had the risk of inherent bias, where machines will discriminate against different types of people. A recent article in the FT brought this concern to the forefront as it showed how medical chatbots downplayed the concerns of women and ethnic minorities. This proves that AI bias is not just a theoretical glitch, but is already causing actual, real-world consequences.
Despite the promise of an AI‑driven future free from human prejudice, evidence increasingly suggests that artificial intelligence not only perpetuates entrenched biases but magnifies them at scale. Rather than occasional errors, AI systems commit systematic missteps that can affect millions of individuals.
Studies have shown that AI often absorbs human prejudices embedded in training data and then amplifies them at scale. What’s worse, another study found that this in turn strengthens humans’ existing biases, creating a harmful feedback loop.
There is, of course, a strong moral and ethical dimension to this issue, but what’s discussed less often is the economic consequences bias carries with it. The biggest of these is how a biased chatbot or other AI may expose a company. For example, in the example from the FT, the healthcare providers are wide open to lawsuits, not to mention the loss of their patients’ trust.
And the health industry is not alone — similar challenges appear in recruitment, advertising, and even finance. Biased credit scoring can expose financial institutions to regulatory sanctions, while AI-driven discrimination in hiring or customer targeting can severely damage brand trust. In all these cases, the use of poorly vetted AI systems can lead to serious legal and reputational repercussions.
It is important to distinguish between algorithmic bias and intentional malfeasance: the presence of bias in an AI system does not necessarily mean that the people behind it hold such views themselves.
The issue often lies in the training data. Many LLMs, particularly those used for text generation, are trained on vast troves of internet-sourced content, which often go uncurated and frequently reflect societal prejudices, misinformation, or unverified authorship. As such, even when model architecture and engineering practices are neutral, the biases embedded in the training data can manifest in the model’s outputs in subtle and pervasive ways.
The quality of an AI model’s output is directly tied to the quality of its training data. For instance, training an LLM on text riddled with typographical errors will likely result in similarly flawed outputs. The same principle applies to more subtle and insidious issues, such as bias: if biased data is fed into a model, this will inevitably be reflected in its responses.
While it is possible to build safeguards to detect and mitigate these issues, doing so requires additional time, expertise, and financial investment. Responsible AI developers account for these costs upfront, recognising the long-term value of robust and ethical design.
Conversely, cutting corners in the interest of speed or cost-efficiency may yield short-term gains, but often results in greater risks, reputational damage, or downstream costs when biased outputs are exposed or cause harm.
This is because bias, of any kind, will erode trust in your company, your staff, and your brand. Trust isn’t some nebulous, nice-to-have concept; it’s a very real thing and losing it will have immediate repercussions for any business. One bad experience in an AI interaction, and people will swear off using it again; this will only magnify in a healthcare situation.
If trust in AI is a public good, it makes sense to have it regulated, though it will, and should, vary across industries. Governments around the world are coming around to the idea, and we’ve seen some interesting approaches.
In the United States, regulation remains fragmented, with a patchwork of state-level initiatives and federal sector-specific guidelines, rather than a unified national AI framework. This decentralisation can create enforcement gaps and allow room for less scrupulous actors to exploit regulatory inconsistencies.
In contrast, the European Union has adopted a more centralised and proactive stance. The EU AI Act, expected to take full effect in 2026, imposes strict obligations on transparency, explainability, and risk management. Its framework aims to ensure greater accountability and protection for individuals affected by AI-driven decisions, including provisions to mitigate bias and discriminatory outcomes.
Australia currently takes a middle-ground approach to AI regulation: more structured than the U.S.’s fragmented model, but less prescriptive than the EU’s binding AI Act. While it has introduced voluntary AI Ethics Principles and public-sector transparency mandates, there is no dedicated AI legislation yet.
Canberra is, however, moving toward risk-based mandatory guardrails, particularly for high-risk applications. This flexible framework allows for adaptability, but critics argue that the lack of enforceable rules leaves room for unchecked bias and insufficient accountability, especially in the private sector.
That said, wherever a company is based, the writing seems to be on the wall: The companies that thrive won’t be those who move fastest, but those who move most responsibly. Break the rules, and consequences will follow.
The best way to make sure that you stay ahead of these issues is to make sure that you have some kind of hybrid system in place where you have AI do the hard work, but make sure that you have humans in the loop to make sure everything that’s going on is following the rules set by your customers and legislators alike.
A good example is to make sure that the data being fed into your LLM, say, is “clean” of questionable biases – you wouldn’t let typo-ridden texts into your database, and you shouldn’t have biased material in there, either. Who trains your AI and how they’re doing it is as important as making sure data is kept safe.
Though testing can be cumbersome and costly, in the long run, you’re saving money and likely your reputation. Much like in the way a cybersecurity breach can ruin your reputation for security, bias in your AI can destroy your trustworthiness as a market player.
The upshot of all this is that being ethical when it comes to AI is going to be a business advantage. If you check for AI bias, make sure you have human oversight, and make sure what you’re doing is compliant with the law (and unwritten rules, for that matter), you’re going to be a step ahead of companies that don’t meet all these requirements.
Not doing so is a core business risk that can tank your company. An AI isn’t just another computer: If we can’t explain how an AI reached its decision, we shouldn’t be surprised when people stop trusting it. And once trust is gone, no amount of innovation can bring it back.
