Unconscious Bias Training in London: What Actually Changes Behaviour

The evidence against the usual format is unusually clear

Unconscious bias training became the default answer to inclusion problems in London organisations for about a decade. It is easy to buy, quick to deliver, and produces a clean attendance record. What it does not reliably produce is a change in what people do. That is not a contrarian position. It is the position the UK government took about its own workforce.

In a written ministerial statement on 15 December 2020, the Cabinet Office announced that unconscious bias training would be phased out in the Civil Service, on the basis that it does not achieve its intended aims. The statement drew on an evidence summary commissioned from the Behavioural Insights Team and published by the Government Equalities Office the same day. That summary found there is no evidence that the training changes behaviour or improves workplace equality in terms of representation in leadership or pay inequalities.

The largest synthesis behind that conclusion is Forscher and colleagues (2019) in the Journal of Personality and Social Psychology, a meta-analysis of 492 studies covering 87,418 participants. Implicit measures can be shifted, but the effects are typically weak, the procedures generally produced trivial changes in behaviour, and changes in implicit measures did not mediate changes in behaviour. Moving a score on a bias test tells you very little about the hiring decision three months later.

• 492 studies and 87,418 participants: implicit measures do move, but the effects are typically weak (Forscher et al., 2019).
• Changes in implicit measures did not mediate changes in behaviour (Forscher et al., 2019).
• No evidence that the training changes behaviour or improves workplace equality (Government Equalities Office and Behavioural Insights Team, December 2020).

Why awareness does not convert into behaviour

The government review sets out six hypotheses for the failure, and they read better as a design brief than as a verdict. Short educational interventions rarely change people who have absorbed bias across a lifetime. Training can activate the very stereotypes it names. Majority groups can feel left out, which reduces their support for diversity. People react against efforts to control them. Participants can become complacent, believing the problem has been solved. And there is a moral licensing effect, where having attended the session makes the next unequal decision feel permissible.

Compulsion makes this worse. The same review cites research covering 830 medium to large US companies over 30 years, in which mandatory diversity training either did not change the number of women in management positions or reduced it. The review also records a risk of back-firing when participants are exposed to material suggesting that biases are unchangeable.

The most careful field test makes the gap visible. Chang and colleagues (2019) ran a pre-registered experiment with 3,016 employees at a global organisation, with an active placebo control, tracking real workplace decisions for up to 20 weeks. Among the groups whose untreated attitudes were relatively less supportive of women, the training produced attitude change but not behaviour change. Some behaviour change appeared among groups already strongly supportive before training. The format worked least well on the people it is usually bought for.

What the evidence points towards instead

None of this is an argument for doing nothing. The Equality and Human Rights Commission assessment of 2018, summarised in the government review, found that unconscious bias training is effective for raising awareness and can reduce implicit bias, but that the evidence for its ability to change behaviour is limited, partly because most of the studies reviewed did not use valid measures of behaviour change. The failure is in part a measurement failure.

Where behaviour did move, the design looked different. Carnes and colleagues (2015) ran a cluster randomised controlled trial across 92 departments and 2,290 faculty members. The intervention was a two and a half hour interactive workshop teaching five specific strategies to practise, including stereotype replacement, perspective taking and individuation, alongside written commitments to action. Where at least a quarter of a department took part, self-reported equity-promoting actions were significantly higher three months later. Practice, specificity and critical mass, rather than exposure.

The government review's own recommendations point the same way: invest in process interventions such as recruitment and progression, which have a better evidence base; avoid mandatory one-off briefings; and require providers to evaluate against outcomes beyond raising awareness.

• Name the specific behaviours the organisation wants, in the situations where they are hardest.
• Rehearse those behaviours rather than explaining the concept behind them.
• Run the work more than once and in more than one format, not as a single briefing.
• Reach enough of a team that the new behaviour becomes the visible norm.
• Change the process as well as the person: shortlisting, promotion panels, allocation of stretch work.

How Sidestream approaches unconscious bias work in London

Sidestream is based at 58 Malden Road, London NW5 3HG, and works with organisations across London and the UK. The method has two halves. The first is immersive theatre made with professional actors, so that a difficult moment arrives as a real interaction rather than as a slide. The second is behavioural science: interventions are designed against the COM-B model of Michie, van Stralen and West (2011), which asks whether capability, opportunity and motivation are all present for the behaviour in question.

That combination matters for bias work in particular. Awareness training addresses one narrow part of motivation and leaves capability untouched. Nobody leaves a briefing more able to interrupt a colleague who has just talked over the quietest person in the room, or to re-open a shortlist that has quietly converged on people who resemble the panel. Capability of that kind is built the way any other skill is built, by doing it, badly at first, somewhere it is safe to be bad at it.

Sidestream's production The Accused addresses equality, diversity and inclusion through lived experience and has been recognised by the Goldsmiths University of London Public Engagement Award. The Death of Jane Doe received a CorpComms Award. Sidestream has worked with the Metropolitan Police, UCL, the University of Cambridge, Bocconi, Goldsmiths, Imperial College London, the Innocence Project, TCS, WISE and HM Courts and Tribunals. Work is scoped and priced per engagement.

Measuring whether anything actually changed

Most bias training is evaluated at Kirkpatrick Level 1, reaction, which asks whether people enjoyed the session, and sometimes at Level 2, learning. The level that decides whether the money was well spent is Level 3, behaviour: whether participants do something different in actual work afterwards.

The distance between those levels is not academic. In a controlled 2020 UCL study by co-founder Ben Laumann (n = 57), rehearsal with a role-play coach produced significantly better feedback delivery than an instructional slideshow or video, rated blind by two independent assessors. The more instructive finding sat underneath the headline: participants in the passive groups self-rated their learning highly, but their measured behaviour did not match. Self-report alone would have called the passive training a success.

That is how an unconscious bias programme survives for a decade without working. It is assessed with the instrument least able to detect its failure. Before contracting any London provider, ask which observable behaviours will differ afterwards and how that will be evidenced.

The compliance question for London employers

A large share of bias training in London is bought as risk management. The Worker Protection (Amendment of Equality Act 2010) Act 2023 inserted section 40A into the Equality Act 2010, requiring an employer to take reasonable steps to prevent sexual harassment of employees in the course of their employment. The duty came into force on 26 October 2024, one year after Royal Assent. Section 124A allows an employment tribunal to order a compensation uplift of up to 25% of the award.

An attendance register for an awareness session is thin evidence when the government's own review of that format found no evidence it changes behaviour. Reasonable steps are easier to demonstrate when an organisation can show which behaviours were specified, how they were rehearsed, who took part, and what changed afterwards. This is an observation about evidence quality, not legal advice; specific obligations should be checked with a qualified adviser.

Frequently Asked Questions

Does unconscious bias training actually work?

It reliably raises awareness. It does not reliably change behaviour. The Behavioural Insights Team evidence summary published by the Government Equalities Office in December 2020 found no evidence that the training changes behaviour or improves workplace equality, and the Forscher meta-analysis of 492 studies found that changes in implicit measures did not translate into behavioural change. Awareness is a genuine outcome, but rarely the one being bought.

Why did the UK government stop unconscious bias training?

A written ministerial statement on 15 December 2020 announced that unconscious bias training would be phased out in the Civil Service, because ministers concluded that it does not achieve its intended aims. The statement referenced a systematic review of 492 studies involving more than 87,000 participants in which changes to unconscious bias measures were not associated with changes in behaviour.

Is unconscious bias training a legal requirement in the UK?

No. There is no legal requirement in the UK to run unconscious bias training. Since 26 October 2024, section 40A of the Equality Act 2010 requires employers to take reasonable steps to prevent sexual harassment, and a tribunal may order a compensation uplift of up to 25% where that duty has not been met. What matters is whether the steps taken were reasonable, not whether a session was held.

What should replace unconscious bias training?

Two things in parallel. Change the processes where bias turns into outcomes, particularly shortlisting, promotion panels and the allocation of stretch work. Then rehearse the specific behaviours managers need in the moments where those behaviours are hard, repeatedly rather than once, and across enough of a team that the new behaviour becomes the visible norm. The Carnes cluster randomised trial found higher reported action where at least a quarter of a department took part.

How do you measure whether bias training changed anything?

At Kirkpatrick Level 3, which asks whether observable behaviour in actual work changed, rather than at Level 1 reaction scores. In Sidestream's own research, participants in passive training rated their learning highly while their measured behaviour did not match, so satisfaction data alone would have declared it a success. Practical measures include structured observation of rehearsed moments, decision audits on shortlisting and promotion, and reporting patterns over time.

Related Sidestream Guides

• DEI training in London built around rehearsed behaviour
• what inclusive leadership looks like as observable behaviour
• the COM-B model of capability, opportunity and motivation
• the Kirkpatrick model and what Level 3 actually measures
• how to measure behaviour change after training
• harassment prevention training and the reasonable steps duty
• immersive training compared with e-learning

Sources: Written Ministerial Statement on Unconscious Bias Training, Cabinet Office, 15 December 2020 · Unconscious bias and diversity training: what the evidence says, Government Equalities Office and Behavioural Insights Team, December 2020 (full report PDF) · Unconscious bias and diversity training: what the evidence says, publication page, GOV.UK · Forscher, Lai, Axt, Ebersole, Herman, Devine and Nosek (2019), A Meta-Analysis of Procedures to Change Implicit Measures, Journal of Personality and Social Psychology 117(3), 522-559 · Carnes, Devine, Baier Manwell and colleagues (2015), Effect of an intervention to break the gender bias habit for faculty at one institution: a cluster randomized controlled trial, Academic Medicine 90(2), 221-230 · Chang, Milkman, Gromet, Rebele, Massey, Duckworth and Grant (2019), The mixed effects of online diversity training, PNAS 116(16), 7778-7783 · Worker Protection (Amendment of Equality Act 2010) Act 2023, sections 40A and 124A, legislation.gov.uk · Worker Protection (Amendment of Equality Act 2010) Act 2023, section 5, commencement one year after Royal Assent, legislation.gov.uk