AI Bias in the Legal Industry and the Rising Tension in Recruitment

Artificial Intelligence is becoming increasingly integrated into the legal profession, creating opportunities to improve efficiency while also raising important questions about equality, diversity and inclusion. A LexisNexis survey in February 2025 found that 46% of legal professionals across the UK and Ireland were already using AI for work, with a further 38% planning to. 

Recruitment matters here because AI-assisted systems can process large volumes of applicant information, identify patterns and support decisions much faster than would otherwise be practical. 

Yet efficiency does not necessarily mean fairness. 

As law firms increasingly adopt AI, the more important question is not simply whether the technology should be used, but whether firms can benefit from it without allowing existing inequalities to become further embedded within recruitment processes and decisions.

AI can reproduce the very inequalities firms are trying to remove

At their most basic level, algorithms are step-by-step instructions designed to complete a particular task. More sophisticated algorithmic systems can analyse vast quantities of data, identify correlations and use those patterns to support decision-making. 

Within recruitment, an algorithm may analyse historical application or employee data to identify characteristics associated with previously successful candidates and then use those patterns when assessing new applicants. For employers dealing with large applicant pools, the advantage is clear: super efficiency.

The difficulty is that the same ability to identify patterns can also entrench them. If an algorithm learns from historically bias employment data, it is necessary to question not only what the system is learning, but where those patterns came from.

A system learns from who has succeeded at a firm already; where senior lawyers entered the profession decades ago under recruitment practices nobody would defend today, the algorithm reads and replicates who the firm has historically hired.

Natalie Sheard's concept of ‘algorithm-facilitated discrimination’ is particularly useful in understanding this problem. Sheard prefers this term to ‘algorithmic discrimination’ or ‘algorithmic bias’ because AI does not operate separately from human decision-making. Discrimination may arise from the technology itself, but it may also result from the way an employer designs, implements or uses that technology. AI can therefore facilitate the unfavourable treatment of individuals because of characteristics such as disability, race, ethnicity, sex, gender or age, but responsibility cannot simply be transferred to the algorithm. The person or organisation using the system remains central to the decision-making process. Although Sheard’s empirical work focuses on Australian employers, the underlying principle is not jurisdiction-specific.

This point is particularly important because it prevents AI from being treated as though it is somehow independent from the people and organisations using it. If a law firm adopts an AI system, chooses the data it relies on and uses its recommendations in recruitment, the firm should still be accountable for the outcomes produced.

Where AI relies on historical employment data, this heightens concerns. If an algorithm learns from previous employment decisions, it is also necessary to consider what human biases may have shaped those decisions in the first place. 

Amazon's experimental recruitment tool provides a leading example. The system analysed historical employee data to identify characteristics associated with successful candidates but reportedly came to favour men and penalised gender-associated terms such as “women's chess club”. 

The system had not simply been instructed to select male candidates. Instead, it learned from historical data generated within a workforce that was already predominantly male.

Historical data do not automatically become neutral simply because they are processed by technology. Nor is simply removing protected characteristics from a system sufficient. Apparently neutral information can still correlate with those characteristics. A postcode, for example, may reflect patterns of residential segregation, while an applicant’s educational history may reflect the institutions from which a firm has traditionally recruited. 

Inequalities created through previous human decisions can potentially be carried into future algorithmic decisions. 

This is one of the clearest reasons why AI should not be viewed as the source of the problem in isolation. Bias may already exist within recruitment before an algorithm is introduced. What AI can add is speed, scale, and, perhaps most importantly, the appearance that the resulting decision has been produced objectively.  

The real weakness may lie in whether the Equality Act can protect applicants in practice

If an AI-assisted recruitment system disadvantages candidates because of a protected characteristic, can existing discrimination law respond effectively?

Michael Selmi takes a relatively optimistic position. Writing in the US context, he challenges the assumption that algorithmic decision-making necessarily requires an entirely new legal response, arguing that established disparate-impact principles can be applied to algorithmic decisions. He also argues that complexity does not protect the employer: a system producing statistically useful results is not enough on its own if the employer cannot demonstrate that the criteria are genuinely connected to job performance.

In other words, the underlying problem is not entirely new. 

Employers have long used selection methods that disproportionately disadvantage certain groups, and replacing a conventional recruitment method with an algorithm does not necessarily transform the nature of that discrimination.

Artificial assistance is still assistance within a decision-making process that ultimately remains connected to the employer and the use of AI should not allow responsibility to be shifted away from the organisation simply because part of the decision-making process has been automated.

Selmi’s argument about complexity persuasive is compelling.  

If an employer cannot explain why particular candidates were selected or rejected, the technical sophistication of the system should not automatically protect the employer from scrutiny. 

In fact, the opposite concern arises: the more difficult a system is to understand, the more important it becomes to ask how its decisions are being reached and whether those decisions can be justified. A firm should not be able to rely on the complexity of an algorithm as a defence where it cannot explain why an individual was treated unfavourably.

Technological complexity may make discrimination harder to identify, but it should not make employers less accountable for the decisions made through systems they choose to use.    

Alysia Blackham's argument is more convincing when this debate is considered within the British context.

Blackham accepts that the Equality Act 2010 may respond to some forms of algorithm-facilitated discrimination. Section 19 can apply where an apparently neutral provision, criterion or practice places people sharing a protected characteristic at a particular disadvantage. An apparently neutral AI recruitment system could therefore potentially fall within the existing framework where its use results in unequal outcomes.

The difficulty is that indirect discrimination can still be justified where the employer demonstrates that the practice is a proportionate means of achieving a legitimate aim. Blackham identifies the danger that this may create a feedback loop. An algorithm may appear effective because its predictions closely reflect existing outcomes, even where those outcomes have already been shaped by inequality. The system could therefore reproduce inequality while simultaneously appearing to generate evidence supporting its own effectiveness.

More importantly, Blackham questions whether existing rights can realistically be enforced. This is where her argument exposes a deeper weakness in the Equality Act 2010 framework. The British system relies heavily on individuals recognising discrimination and bringing claims themselves, yet algorithmic recruitment can make that considerably more difficult.

An unsuccessful applicant may not know whether AI was involved in the decision, which factors influenced the outcome, or whether other applicants sharing their protected characteristic experienced the same disadvantage. The speed, scale and lack of transparency associated with AI can therefore make discrimination difficult to identify before a legal challenge can even begin. This suggests that the deeper problem is not simply whether the Equality Act provides rights in principle, but whether individuals are realistically able to recognise when those rights may have been breached and enforce them in practice.  

Which brings us back to Selmi point, that discrimination existed long before AI and that algorithmic decision-making does not create an entirely new category of discrimination.

However, this reveals a problem rather than resolving one. If weaknesses already existed in the way discrimination was identified and challenged, AI does not remove them; it may simply make them more difficult to see.

The real issue is therefore not whether the Equality Act 2010 technically applies to algorithmic discrimination, but whether the existing framework is capable of providing effective protection and remedies in practice.

Selmi is right to emphasise that the underlying discrimination is not new, but Blackham shows why that does not make the existing legal framework sufficient. If flaws were already present before AI became common, automated recruitment may carry those same flaws into a process that is faster, more complex, and more difficult for applicants to challenge.

The question, then, is not simply whether the Equality Act applies. It does. The more important question is whether a right can provide meaningful protection if the person holding it is unable to recognise when it has been breached or realistically enforce it.

Human oversight only matters if firms are prepared to question the technology

These concerns do not necessarily justify removing AI from recruitment altogether. The evidence instead points towards using AI as a supporting tool while retaining meaningful human involvement.

Thomas International's guide on AI in recruitment identifies several potential benefits, including the ability to process large numbers of applications, automate repetitive tasks, and apply more consistent assessment methods across recruitment processes. AI can also assist with screening, scheduling, and candidate assessment. However, the same guidance acknowledges that these systems can learn and reproduce existing human biases, and therefore recommends that automated AI-powered workflows are routinely tested and verified. 

This is important because the efficiency and consistency offered by AI do not remove the risk of discrimination; they instead make ongoing scrutiny of how these systems operate especially important.

The distinction that matters is between using AI to support recruitment and allowing it to control the process. 

Human involvement must therefore be meaningful. Simply keeping a recruiter somewhere in the decision-making process is not enough if that person automatically accepts whatever the system recommends. 

Real oversight requires understanding the criteria being applied, questioning the output, and being prepared to challenge or overrule the system where its recommendations appear inconsistent, unfair, or difficult to justify.

Law firms should not wait for applicants to expose discrimination themselves

AI has clear potential to make recruitment faster, more efficient, and more manageable. However, the evidence also demonstrates why efficiency cannot be treated as proof of fairness. Algorithms can learn from historical employment decisions, reproduce existing patterns of inequality, and make it difficult for unsuccessful applicants to understand why they were rejected.

At the same time, rejecting AI altogether does not automatically produce a fairer recruitment process. Human decision-making can itself result in discrimination, while appropriately designed and supervised technology may support greater consistency and efficiency. 

The better approach is to treat AI as a tool that assists, rather than replaces, accountable human decision-making.

The central issue is ultimately one of responsibility. 

The burden should not fall mainly on unsuccessful applicants to identify and challenge discrimination after it has already occurred, particularly where the recruitment process is designed in a way that makes discrimination difficult to detect. If law firms choose to use AI, they should also accept responsibility for ensuring that the systems they adopt operate consistently with the equality and diversity standards they claim to uphold.

In practice, this means testing outcomes, examining whether particular groups are being disproportionately disadvantaged, scrutinising the historical data on which systems rely, and ensuring that recruiters are capable of questioning and, where necessary, overriding automated recommendations. The Equality Act 2010 may provide legal protection, but Blackham's analysis shows why firms should not rely on individual applicants to identify possible discrimination and invoke those protections themselves.

As AI becomes more common within legal recruitment, firms need to move beyond asking whether their systems are legally permissible or technologically efficient, and ask whether they are consistent with the EDI commitments they have already made. Those commitments mean very little if a machine is quietly undoing them at the first sift.

REFERENCES:

• Equality Act 2010

• Blackham A, ‘Artificial Intelligence and Equality at Work: Evaluating the Adequacy of the UK Legal Framework for Responding to the Discrimination Risks of New Technologies’ (2026) Industrial Law Journal (advance online publication, dwag013) https://doi.org/10.1093/indlaw/dwag013 accessed 26 October 2026

• Selmi M, ‘Algorithms, Discrimination and the Law’ (2021) 82 Ohio State Law Journal 611

• Sheard N, ‘Algorithm-Facilitated Discrimination: A Socio-Legal Study of the Use by Employers of Artificial Intelligence Hiring Systems’ (2025) 52 Journal of Law and Society 269

• Beaumont-Oates W, ‘AI Recruitment and How it Works’ (Thomas International, updated 26 June 2024) https://www.thomas.co/resources/type/hr-blog/ai-recruitment-and-how-it-works accessed 26 October 2026

• Dastin J, ‘Amazon Scraps Secret AI Recruiting Tool that Showed Bias against Women’ (Reuters, 10 October 2018)

• Shabir M, ‘AI Adoption Amongst Lawyers Passes Tipping Point’ (LexisNexis, 3 April 2025) https://www.lexisnexis.co.uk/blog/future-of-law/ai-adoption-amongst-lawyers-passes-tipping-point accessed 26 October 2026

At Diversity+, we are committed to helping organisations make diversity, equity, and inclusion a core aspect of their business, and help individuals overcome adversity at all stages, including of their legal careers. We offer bespoke programmes for organisations across the UK, and bespoke mentoring for individuals. Reach out to us at hello@diversityplus.info for more information and a free consultation call.