AI in consultation analysis: the legal risks

Generative AI is being used to analyse consultation responses and produce Equality Impact Assessments across the UK public sector. The error rates are documented, the evidence base for safe use is thin, and the legal consequences under the Public Sector Equality Duty and the Gunning principles may be more serious than most decision-makers currently appreciate.

How often does AI get consultation analysis wrong?

Almost everyone who uses AI tools for research or analysis can recall a moment when the output was clearly wrong. The harder question is how often AI gets things wrong in ways plausible enough not to prompt a second look.

The answer, based on published research, is more often than most practitioners assume. Researchers examining AI tools in legal practice found that even specialist legal AI tools hallucinated false information between 17% and 33% of the time. A public sector field experiment found that while AI improved some document tasks, it reduced quality scores in data-analysis tasks by 12%. Analysing free-text consultation responses and drawing equality conclusions from them is precisely that kind of data-analysis task.

There is also a documented psychological mechanism that makes this worse. Research on automation bias in public sector decision-making shows that decision-makers tend to over-rely on algorithmic outputs even when warning signs are present elsewhere. If an AI-generated Equality Impact Assessment (EqIA), the formal assessment public bodies must carry out to consider the effect of their decisions on people with protected characteristics, looks authoritative and well-structured, the instinct to scrutinise it carefully is reduced.

What specific risks does AI introduce to consultation analysis?

The evidence base for using AI to analyse public engagement and consultation responses is currently very limited. That matters, because limited evidence has not meant limited adoption.

Two problems are emerging in practice. First, members of the public are submitting AI-generated responses to consultations, including coordinated campaigns where hundreds of near-identical submissions are used to manufacture the appearance of broad support for a particular position. An AI tool analysing those responses may report a pattern of opinion that reflects artificial inflation rather than genuine public sentiment.

Second, free-text analysis, the part of consultation most important for understanding what people actually think, is where AI performance is least reliable. Nuance, local context, and lived experience described in consultees’ own words are precisely what current AI tools are worst at preserving accurately.

The combination creates a specific risk: an AI-generated analysis of a consultation influenced by AI-generated responses may produce a picture of public opinion that is doubly distorted, and that neither the analyst nor the decision-maker can detect without checking the underlying data carefully.

What are the legal consequences if AI-generated analysis gets it wrong?

The Public Sector Equality Duty (PSED), established under section 149 of the Equality Act 2010, requires public bodies to have due regard to equality when carrying out their functions. Courts have consistently held that due regard requires genuine, informed, and documented consideration of equality impacts before a decision is made. An EqIA shaped by AI that contains undetected errors may not satisfy that standard. The assessment would exist on paper. The due regard would not.

Alongside the PSED sits the legal test for fair public consultation. The Gunning principles, derived from R v London Borough of Brent ex parte Gunning (1985) and confirmed by the Supreme Court in R (Moseley) v London Borough of Haringey (2014), require that the product of consultation is conscientiously taken into account when a decision is made. Conscientious consideration means decision-makers must genuinely engage with what consultees said. If the summary they receive has been produced by an AI tool that has mischaracterised or omitted what consultees said, the conscientious consideration required by law has not taken place, regardless of whether a report was produced.

What should consultation leads and decision-makers do now?

The risk is manageable. It is not self-managing. Three steps represent the minimum sound practice currently requires.

• Treat AI outputs as a first draft. Any AI-generated analysis of consultation responses, or AI-assisted EqIA, must be reviewed against the underlying data before informing a decision. This is not a counsel of perfection. It is what verification means in a legally sensitive context.
• Audit the data before analysis. Where coordinated AI-generated responses are present, identify and handle them separately. Document their existence. Treating manufactured responses as equivalent to genuine consultation is itself a process failure.
• Document your quality assurance. Record the fact that AI tools were used and the steps taken to verify outputs. A decision-making record that shows AI was used without any verification trail is a vulnerability in any subsequent legal challenge. Bear in mind too that AI systems can reflect and replicate bias present in their training data. Using AI to produce an EqIA does not remove the need to consider the equality implications of the tool itself.

The greatest risk posed by generative AI in public decision-making may not be that it occasionally gets things wrong. It is that it often sounds sufficiently authoritative and plausible that organisations stop asking whether it is right. That is not a technology problem. It is a governance problem.

How tCI Can Help

Organisation Wide Learning Hub Access
Equip your entire team with professional consultation skills through one platform. Self paced courses, live virtual classrooms, practical toolkits and expert resources that build a shared baseline of competence across your organisation. Trusted by councils, NHS bodies and regulators nationwide.

Bespoke Training Workshops
Training that works with your real projects, not hypothetical scenarios. Sector tailored sessions help teams apply good practice to live challenges: sharpening consultation documents, building defensible codebooks, strengthening equality analyses. Half day or full day workshops for health, local government, planning and public service teams.

Coaching for Complex or High Risk Consultations
Expert guidance when the stakes are highest. One to one and small group coaching for senior officers navigating legally exposed or politically contentious decisions. Strengthen your judgement on proportionality, evidence standards and challenge management. Essential for organisations that may face judicial review risk or major service changes.

Whether you’re preparing for a high stakes service change, building long term consultation capability, or need confidence that your evidence approach will stand up to scrutiny, we can help.

Contact tCI: hello@consultationinstitute.org.