How “Don’t Harass” Became “Don’t Disagree”: The Three-Stage Capture of Evaluative Speech in British Law
Anti-discrimination law exists to stop people being bullied, excluded, and degraded. Nobody serious disputes that. What I want to show you is something more specific and more worrying: a structural mechanism inside the UK’s Equality Act 2010 that allows institutions—particularly universities—to slide, step by step, from prohibiting acts of harassment to punishing people for holding and expressing evaluative judgments. The slide is not inevitable. But the statutory architecture makes it predictable, and documented cases confirm it is not theoretical.
This is the argument I develop at length in a forthcoming paper (preprint on SSRN and in peer review). Here I want to lay out the core claim in plain language, because this matters to anyone who teaches, studies, researches, writes, or simply thinks out loud in a British institutional setting.
The Mechanism: Section 26 and the “Purpose or Effect” Gateway
The critical provision is section 26 of the Equality Act 2010, which defines harassment. Under s 26, Person A harasses Person B if A engages in “unwanted conduct” related to a protected characteristic, and that conduct has the purpose or effect of violating B’s dignity or creating a hostile, degrading, humiliating, or offensive environment.
Read that again. Purpose or effect. The disjunction is the entire mechanism.
Under the “purpose” limb, the inquiry is into what the speaker intended. Did they set out to degrade someone? That is a subjective test directed at the actor, and it is relatively narrow. Under the “effect” limb, the inquiry shifts entirely: it asks what the recipient experienced. Did they feel their dignity was violated? If so—subject to some qualifications—the conduct may constitute harassment regardless of what the speaker intended.
Now, the statute does include a qualifier. Section 26(4) says that when assessing whether conduct had the relevant effect, a tribunal must consider: (a) the perception of the recipient, (b) the other circumstances, and (c) whether it is “reasonable” for the conduct to have that effect. In tribunal proceedings, this reasonableness filter has real teeth. Courts consistently dismiss harassment claims where the complainant’s reaction, however genuine, is objectively unreasonable. The Employment Appeal Tribunal made this clear in Richmond Pharmacology v Dhaliwal: the statutory language prevents “trivial acts causing minor upset” from constituting harassment.
So what is the problem? The problem is what happens when institutions—universities above all—transpose the statutory language into their own conduct codes. They reproduce the trigger (”unwanted conduct” with the “effect” of “violating dignity” or “creating a hostile environment”) while discarding the constraint (the reasonableness test, the contextual assessment, the objective filter). A composite test requiring subjective perception, context, and objective reasonableness collapses into a single-element test requiring only subjective perception.
This is not a minor drafting oversight. It changes the operative decision rule at every stage of the process: at complaint initiation, at investigation, at adjudication, and at appeal.
The Three Stages
I identify a three-stage transition in how institutional “tolerance-language”—terms like respect, civility, dignity-violation, hostile environment—operates.
Stage 1 is act-constraint. Norms target interference, threats, coercion, and discrimination in access to goods, services, or opportunities. Evaluative judgment—forming and expressing opinions, including controversial ones—is presupposed as a normal feature of public discourse and academic inquiry. The University of Chicago’s 2015 Report of the Committee on Freedom of Expression captures this baseline perfectly: “Concerns about civility and mutual respect can never be used as a justification for closing off discussion of ideas.” Under this framework, you can say things other people find deeply offensive. What you cannot do is prevent them from participating, threaten them, or discriminate against them.
Stage 2 is the evaluative gateway. This is where institutions adopt language—typically drawn from the Equality Act—that extends violations to the subjective effects of evaluative speech on recipients’ experience of dignity. The critical feature is that a “respect” or “civility” norm is no longer merely aspirational (an ethos statement encouraging good manners) but becomes a trigger rule: it either conditions whether debate is protected (”debate is protected only if conducted respectfully”) or activates complaint-handling (”reports of disrespectful conduct initiate investigation”). The distinction matters enormously. An aspirational norm is advice. A trigger rule is a weapon.
I examined the conduct codes of five Russell Group universities—Oxford, Glasgow, Edinburgh, Cambridge, and Exeter—and all five had reached at least Stage 2. Oxford’s Code of Practice on Freedom of Speech states that “vigorous academic debate will not amount to harassment when it is conducted respectfully” but may constitute harassment when it “violates another’s dignity.” The “respectfully” qualifier conditions whether debate is protected, which makes it a gateway rather than an ethos statement. Glasgow replicates the s 26 harassment definition alongside its academic freedom provision without resolving the tension between them. Edinburgh established a Working Group in 2023 specifically to manage the conflict between its Dignity and Respect Policy and its academic freedom obligations—the very existence of the Working Group is evidence of the structural collision. Cambridge includes an interpretive disclaimer (”Nothing in this Policy should be interpreted as undermining freedom of thought and expression”) but does not alter the trigger clause. The disclaimer confirms the conflict by assertion rather than resolution.
Stage 3 is judgment-as-sanction-target. This is where evaluative judgment itself becomes disciplinable: individuals face institutional sanction not for acts of interference but for expressing positions that others experience as offensive or dignity-violating.
The University of Exeter’s disciplinary regulations provide a documented instance. The regulations state that disciplinary offences include “actions which cause actual or potential distress or harm to others irrespective of whether or not distress or harm was intended.” The trigger clause explicitly removes intent as a relevant consideration, leaving only the subjective effect on the recipient. The s 26(4)(c) reasonableness qualifier is absent from the operative language. The procedural scaffolding compounds the problem: investigations are conducted internally without independent adjudication, there is no formal disclosure obligation, no requirement to distinguish evaluative expression from targeted conduct, and no explicit defence for academic, philosophical, or political expression. The evidential threshold for initiating investigation is the filing of a complaint, not a preliminary assessment of whether the conduct meets an objective standard.
Trace a hypothetical through Exeter’s framework. An academic states in a seminar that biological sex is binary and immutable—a position the Employment Appeal Tribunal held to be a protected philosophical belief in Forstater. A student files a complaint alleging distress. The trigger clause is satisfied: the conduct caused “actual or potential distress” “irrespective of” intent. No preliminary reasonableness assessment is required. The complaint alone activates the machinery. At investigation, the decision rule is whether the conduct caused “distress or harm.” Reasonableness is not a specified element. At adjudication, the panel is not required to weigh the evaluative nature of the expression, apply proportionality, or consider Article 10 rights. In tribunal proceedings, all of these would be mandatory.
The Office for Students’ own Regulatory Advice 24 guidance confirms this is not an isolated phenomenon. The regulator itself identified university IT policies stating that “users must not transmit offensive material” and blanket pronoun-usage policies that “in intent or effect, prohibit the expression of a lawful viewpoint,” concluding both required amendment.
What the Courts Have Said
The courts have confronted the act–judgment boundary repeatedly, and the pattern is consistent: institutions treat evaluative judgment as equivalent to misconduct, and courts correct that conflation—but only retrospectively, at enormous cost, and in ways that do not prevent the next institution from making the same error.
In Forstater v CGD Europe, the initial Employment Tribunal held that gender-critical beliefs were “not worthy of respect in a democratic society” and therefore did not qualify as a protected philosophical belief. The tribunal did not find that Forstater had committed any act of harassment, intimidation, or discrimination. It found that her evaluative judgment—a philosophical position on the relationship between sex and gender—was itself unworthy of protection. The EAT reversed, and the subsequent full merits hearing confirmed direct discrimination. But it took multi-year litigation across three proceedings to establish that holding a philosophical position is not equivalent to harassment. For most people, the rational response to that kind of risk is silence.
In Miller v College of Policing, the Court of Appeal held that the police recording of lawful evaluative speech as a “non-crime hate incident”—combined with a police visit to Mr Miller’s workplace—constituted a disproportionate interference with his Article 10 rights. Knowles J at first instance found there was “not a shred of evidence” that Miller was at risk of committing a criminal offence, and that his tweets were “no more than expressions of opinion on a topic of current controversy.” This is Stage 3 operating through administrative rather than disciplinary mechanisms: judgment itself becomes the target of institutional action—through recording, retention, and a workplace visit—even where no offence has been committed.
In Lee v Ashers Baking Company, the Supreme Court held unanimously that anti-discrimination legislation cannot compel the expression of views that a person does not hold. The bakery’s objection was to the message (”Support Gay Marriage”), not to the customer’s sexual orientation. This draws a clean line between discrimination in the provision of services (an act) and refusal to endorse an evaluative position (a judgment).
In Ngole v University of Sheffield, the Court of Appeal held that the university had failed to consider a social work student’s Article 9 and Article 10 rights when expelling him for posting Biblical views about homosexuality on his personal Facebook page. The university had treated the expression of a religious evaluative position as equivalent to an inability to practise non-discriminatorily—conflating judgment with conduct without evidence that Ngole had ever actually discriminated against anyone.
Each of these cases was eventually resolved correctly. But each required years of litigation, significant personal cost, and considerable legal uncertainty. For every Forstater, Miller, or Ngole who fights, there are many more who simply stop speaking.
The Criminal Scaffold
Below the Equality Act, a network of criminal statutes creates parallel evaluative gateways that compound the uncertainty. Section 5 of the Public Order Act 1986 criminalises “threatening or abusive” words “likely to cause harassment, alarm or distress.” Section 127 of the Communications Act 2003 creates an offence of sending a “grossly offensive” message. The Protection from Harassment Act 1997 prohibits “a course of conduct amounting to harassment” and, because it requires only two instances, is capable of capturing repeated evaluative speech.
These thresholds are doctrinally distinct from s 26—they operate under different standards with different interpretive traditions. But the point is cumulative uncertainty. Where multiple overlapping evaluative standards apply to the same speech, the rational response is not careful legal analysis of each provision but blanket self-censorship.
The case law illustrates the inconsistency starkly. In Hammond v DPP, a street preacher was convicted under s 5 for displaying a sign reading “Stop Immorality, Stop Homosexuality”—the court treated evaluative expression as an act of public disorder. In Scottow v CPS, the High Court quashed a conviction under s 127, holding that persistent misgendering on social media did not meet the “grossly offensive” threshold. And in Redmond-Bate v DPP, Sedley LJ held that a street preacher’s expression of religious views was protected by Article 10 even where it provoked a hostile audience reaction. Hammond treats evaluative expression as disorder; Scottow and Redmond-Bate resist that characterisation. The inconsistency is not aberrational—it reflects the absence of a clear statutory distinction between act-constraint and judgment-constraint in the criminal context.
Eric Kaufmann’s large-scale survey found that one-third of conservative academics and doctoral students reported being threatened with discipline for their views, and over half self-censored in research and teaching. Frederick Schauer’s work on chilling effects explains the mechanism: legal uncertainty deters speakers from protected activity even where the law would not sanction their conduct. The criminal scaffold does not independently prove the s 26 mechanism, but it reinforces the broader environment of evaluative risk that makes self-censorship the rational choice.
The Comparative Picture
The three-stage transition is not an inevitable feature of rights-protective legal systems. Comparative analysis shows it is strongest where legal architectures provide evaluative gateways and weakest where strict act–speech distinctions are maintained.
The foundational ECHR authority on freedom of expression, Handyside v United Kingdom, established that Article 10 protects information and ideas “that offend, shock or disturb the State or any sector of the population.” This formulation, in principle, resists the evaluative gateway. But Article 10(2) permits restrictions “prescribed by law” and “necessary in a democratic society,” and Article 17 excludes activities aimed at destroying Convention rights. In Norwood v United Kingdom, the Court held that a poster stating “Islam out of Britain” fell outside Article 10 protection entirely—a categorical exclusion, not a proportionality assessment. The margin of appreciation permits significant variation in how strictly different member states maintain the act–judgment distinction.
US constitutional doctrine illustrates a different mechanism entirely. The First Amendment’s act–speech distinction, as articulated in Brandenburg v Ohio, limits the sanctionability of advocacy to speech “directed to inciting imminent lawless action and likely to produce such action.” That imminence requirement categorically excludes evaluative speech from criminal or civil sanction in public discourse. But even within the US system, hostile-environment standards exist in educational settings: the Supreme Court’s Davis v Monroe County decision established Title IX harassment requirements of conduct “so severe, pervasive, and objectively offensive” that it effectively denies equal educational access. Those threshold requirements—severity, pervasiveness, objective offensiveness—are not explicit conditions under s 26. Yet even robust constitutional speech protection does not eliminate evaluative gateways within institutional anti-discrimination frameworks.
The comparative evidence supports a precise claim: different legal systems employ different mechanisms to maintain the act–judgment boundary. US First Amendment doctrine maintains it through explicit threshold requirements. ECHR jurisprudence maintains it partially through proportionality analysis. The UK Equality Act architecture lacks an equivalent explicit mechanism, relying instead on the composite s 26(4) test—which works in judicial hands but is vulnerable to attenuation when transposed into institutional settings. That absence is the structural vulnerability enabling the three-stage transition.
The Dignity Problem
The most sophisticated justification for the evaluative gateway is Jeremy Waldron’s dignity rationale, which treats dignity not as a subjective feeling but as an objective social status—the basic social standing of persons as full members of society. Waldron insists on distinguishing dignity from offence: the law should protect social standing, not feelings. This is an important distinction. The problem is that section 26 does not maintain it. The statutory text asks whether conduct had the “effect” of “violating” dignity—a formulation that collapses back into subjective perception unless the reasonableness filter is rigorously applied.
Alexander Brown argues that Waldron’s dignity–offence distinction, while analytically powerful, is unworkable in practice: the boundary is extremely difficult to maintain. The s 26 evaluative gateway instantiates Brown’s concern perfectly. Ronald Dworkin identified the democratic cost: criminalising evaluative speech “spoils the democratic justification we have for insisting that everyone obey” anti-discrimination legislation. Eric Heinze goes furthest: viewpoint-selective speech restrictions are never democratically legitimate in stable democracies, because free expression is a constitutive attribute of democratic citizenship—one of the “legitimising expressive conditions” without which democratic authority is compromised.
If Heinze is right, the evaluative gateway is not merely an operational inconvenience but a structural threat to democratic governance.
There is a further compounding factor that makes the institutional slide predictable. The public sector equality duty under s 149 of the Equality Act requires public authorities to demonstrate conscious consideration of equality implications in everything they do. This interacts with complaints procedures to create compliance cultures that prioritise the avoidance of equality-related complaints above all else. Universities adopt broad conduct codes not because the Equality Act requires them to sanction evaluative speech, but because the PSED incentivises risk-aversion in which expansive prohibition is perceived as the safer institutional strategy. Paul Yowell identifies the resulting dynamic: the PSED, in combination with subjective harassment definitions, “incentivises ideological uniformity” and creates institutional pressure to exclude debate. The PSED does not independently create the vulnerability, but it systematically amplifies it.
The Higher Education (Freedom of Speech) Act 2023: An Incomplete Fix
Parliament’s response to this problem in higher education was the Higher Education (Freedom of Speech) Act 2023, which strengthens the duty on universities to “secure” and “promote” freedom of speech. The OfS’s Regulatory Advice 24 states that “free speech includes lawful speech that may be offensive or hurtful to some” and that there should be a “very strong presumption in favour of permitting lawful speech.”
But the guidance also states that “speech that amounts to unlawful harassment is not protected.” And here is the gap the three-stage model exposes: the operational question for institutions is how “unlawful harassment” is determined. If institutions treat the filing of a complaint as sufficient evidence of unlawfulness—rather than undertaking the contextual, reasonableness-qualified assessment that s 26(4) requires—then the distinction between protected and unprotected speech collapses in practice.
The Act also faces a more fundamental structural limitation: it operates alongside the Equality Act without resolving the tension between the two statutory regimes. Where an academic expresses a view that a student experiences as unwanted conduct creating a hostile environment, the university simultaneously owes a duty to secure that speech under the 2023 Act and a duty to investigate a harassment complaint under the Equality Act. The Act does not establish a hierarchy, and the OfS guidance explicitly declines to create one, framing the issue as a matter for “case-by-case” institutional judgment. The government’s subsequent decision to pause implementation added further uncertainty, though a June 2025 policy paper reaffirmed commitment to robust protection of academic freedom.
Critically, the Act applies only to registered higher education providers. It does not address the three-stage transition as it operates through employment, professional regulation, or the criminal scaffold. The legislative response is domain-specific, addressing the symptom without addressing the cause in s 26.
What Would Fix This
Three reforms would address the structural vulnerability.
First, amend s 26 to include an explicit carve-out: the expression of a philosophical, political, religious, or academic position does not constitute harassment merely because it has the effect of causing offence to a person who disagrees with it. This should operate as a definitional exclusion (the expression falls outside the definition of harassment), not a defence (the expression constitutes harassment but can be justified). The distinction matters: a definitional exclusion prevents the disciplinary machinery from being activated at all; a defence permits activation and places the burden on the respondent, preserving the chilling effect. Protection against targeted, sustained personal abuse, threats, intimidation, and exclusionary conduct would be entirely unaffected.
Second, strengthen the reasonableness qualifier in s 26(4)(c) by requiring particular weight to be given to whether the conduct constituted evaluative expression on a matter of public, academic, or professional concern.
Third, establish an explicit hierarchy within the Higher Education (Freedom of Speech) Act providing that the duty to promote free speech takes precedence where the conduct constitutes evaluative expression rather than targeted interference.
Under all three reforms, the “purpose” limb of s 26 would continue to capture conduct intended to harass, and the “effect” limb would continue to apply to sustained personal abuse, threats, and exclusionary conduct. No one is proposing to legalise bullying. The proposal is to stop pretending that expressing a philosophical view is the same thing as bullying someone.
The Bottom Line
The Equality Act 2010, properly applied by tribunals with the full composite test of s 26(4), does not equate evaluative judgment with harassment. That must be stated clearly because the argument is not that the law itself is broken in judicial hands. The argument is that the statutory text creates a liability pathway that institutional actors—operating under compliance incentives, procedural informality, and risk-averse cultures amplified by the public sector equality duty—predictably extend beyond its intended scope. The three-stage transition from act-constraint through evaluative gateway to judgment-as-sanction-target is not an inevitable consequence of protecting dignity. It is a consequence of statutory imprecision interacting with institutional incentive structures in ways that can and should be corrected.
A normative architecture in which evaluative judgment on matters of public concern is perpetually liable to institutional sanction through subjective-effect tests administered without adequate procedural safeguards creates a documented risk of viewpoint-discriminatory enforcement and chilling. That is difficult to reconcile with the conditions of open inquiry that academic freedom and democratic governance both require. The law needs to distinguish between acts and judgments—explicitly, in the statute, not by relying on tribunals to clean up the mess after years of litigation and thousands of pounds in legal costs. It needs to distinguish them because the failure to do so has real consequences for real people, and because a legal system that cannot tell the difference between harassing someone and disagreeing with them has a problem that no amount of institutional guidance will solve.