AI in Motion

Can AI ever be truly neutral, and is neutrality what we should be aiming for?

That was the first question at the second of our “AI in Motion” events, a series tackling the big, human questions around AI. This session focused on bias in AI: an issue that has been in the headlines for over a decade but one now compounded by the rapid scaling of AI adoption and the growing autonomy of the technology.

The event was chaired by TLT’s Emma Erskine-Fox, with expert contributions from Lili Elenoglou (Legal Director, TLT), Tess Buckley (Senior Programme Manager – Digital Ethics and AI Safety, TechUK), Lara Groves (Senior Researcher, Ada Lovelace Institute) and Nish Imthiyaz (Global Privacy Counsel, Vodafone).

Panellists were clear from the outset: AI cannot truly be neutral, and the goal of AI neutrality is neither achievable nor coherent. Instead, organisations should focus on optimisation, making considered choices about the trade-offs they are comfortable accepting in their AI use. With that context set, the discussion moved to how bias emerges in AI systems and crucially, what organisations can do about it.

Human bias vs AI bias

A point often made when discussing AI bias is that humans are also biased – is AI bias really any different? Yes, said our panellists. Human bias is contextual and can be challenged in the moment, whereas AI bias gets baked in and can be replicated millions of times. The perception of AI being more objective than humans is also dangerous in this context, creating the potential for overreliance on AI outputs and a failure to detect and mitigate bias.

It is also important to recognise that bias is not just about gender and race, despite these being the most widely reported examples. Organisations need to be alive to all potential areas of bias, including less explored areas such as socioeconomic inequality, ableism and class systems.

Bias throughout the AI lifecycle

Bias is a live operational risk that can occur at any point in the AI lifecycle. The panel discussed different types of AI bias that can be reflected in AI-generated outputs, including:

• Data bias, where bias in training data causes an AI model to learn skewed patterns.
• Algorithmic bias, where systematic errors in algorithms produce unfair or discriminatory outputs.
• Feedback loop bias, where an AI output is fed back into training data, causing the model to continuously learn from and amplify its own previous biases.

However, bias can be most concentrated at the problem definition and ideation stages, where stakeholders make fundamental, often unintentional, assumptions about what an AI tool should achieve. Bias is often not visible, particularly in the post-deployment monitoring stage where it can be hard to catch.

Another point where AI bias can enter is through using proxy variables, which can unintentionally represent protected characteristics – for example, postcode can be a proxy for race. The absence of explicit protected characteristics does not remove the risk of bias and organisations need to be mindful of these potential proxies

What does the law say?

The legal and regulatory framework governing AI is fragmented and jurisdiction-specific. Existing laws and regulations are broadly outcomes-focussed rather than prescriptive about steps organisations need to take to address bias in AI.

In the UK, a dense regulatory framework spans multiple areas relevant to AI bias. Data protection laws include requirements around fairness, accuracy and transparency, while the Equality Act prohibits discrimination on the grounds of protected characteristics. Sector regulators like the FCA and CMA are coordinating regulatory efforts through the Digital Regulation Cooperation Forum, and the UK’s five cross-cutting principles (including “Fairness”) form a pro-innovation framework which expects regulators to apply existing core principles and concepts to AI.

In contrast, the EU approach is more prescriptive. The EU AI Act imposes detailed requirements on high-risk AI systems, including expectations for training datasets to be relevant, representative and, to the extent possible, complete and error-free. It also requires organisations to assess the impact of possible biases; and to ensure measures are in place to detect, prevent and mitigate biases.

Globally, frameworks like the OECD AI principles focus on inclusivity, human-centric values and fairness to promote trustworthy AI.

What does this mean in practice?

Whilst there are numerous global principles, operationalising them in practice is a challenge. In particular, the division of responsibilities for bias mitigation between developers and deployers is often unclear. Organisations acting primarily as deployers, rather than developers, may have limited understanding or visibility of how models are trained, raising the question of how far they need to go to monitor and mitigate bias themselves.

Mature organisations are shifting towards use-case governance: considering the purpose for use of the tool then focussing on contextual bias testing within their own environment, without relying solely on the vendor’s assurances.

Key practical recommendations for organisations include:

For developers

• Training data: Use diverse, representative training data, being mindful of proxy data and protected characteristics.
• Test and record: Carry out robust bias audits across different groups, not just in aggregate. A model that is 90% accurate overall can still be systematically wrong for specific demographics, so bias audits should always be disaggregated.
• Document design choices and trade-offs: Record what data was used and was excluded, what the model was optimised for and why, and what trade-off decisions were made during development and training, to create accountability and facilitate future audits.

For deployers

• Inventory: Maintain a clear central catalogue of AI tools in use. Effective governance starts with visibility, so a clear inventory of AI systems is an essential foundation.
• Vendor engagement: Require transparency from vendors, review bias audits as part of due diligence and ensure there are mechanisms and contractual obligations for addressing issues where bias is detected.
• Conduct independent testing: Where possible, don’t simply rely on the vendor’s assurances and audits; test for bias internally, and implement continuous monitoring to identify biased outcomes.
• Contestability and redress: Even with the most robust testing and monitoring, bias will always be a risk. Put in place mechanisms to ensure those impacted by biased outcomes can contest decisions and obtain swift correction and redress where needed.
• Human oversight: Ensure meaningful human involvement for high-stakes decisions, carried out by staff with bias training and enough AI literacy to be aware of and spot bias risks.

For all organisations

• Awareness: Recognise what you don’t know about bias and avoid over-confidence. There is a significant difference between “we have not found bias” and “we have tested rigorously for bias and are confident in the outcome”.
• Accountability: Assign clear accountability for AI governance to someone with genuine decision-making power in your organisation, not just a policy owner.
• Invest in diversity: Diverse teams – taking into account not just gender, ethnicity and other protected characteristics, but also socioeconomic background, age, class, skillset and lived experience – are more likely to spot blind spots and ensure the most robust bias mitigation strategy.

This publication is intended for general guidance and represents our understanding of the relevant law and practice as at June 2026.  For more information see our terms & conditions.