Scratching the surface, checking the citations

A note on why I looked at this at all. Someone close to me is going through a difficult time in a corner of life that brushes up against the family justice system. I am not. My marriage is fine, I have no case, no axe, and no war story. But watching from the sidelines made me start reading in an area I would otherwise never have touched, and in June a report landed that made claims big enough to check. I have spent years around QED, the Skeptics Guide, Skeptics with a K, Cognitive Dissonance and The Scathing Atheist, and the one thing that community drills into you is that the size of a claim sets the size of the checking it deserves. I could not leave this alone. This post is the checking, so that anyone who goes looking can find it.

In June 2026 an organisation called Right to Equality published a report titled Scratching the Surface: Victim-Blaming and Bias in Family Court Judgments. It analysed 91 published family court judgments from England and Wales and concluded that 72.5% of them contained judicial victim-blaming, with 530 instances attributed to court professionals. It was launched in Parliament, covered by the Guardian and the Bureau of Investigative Journalism, and welcomed by major charities. The headline number is now loose in the world.

The claim is serious. If nearly three quarters of published family judgments contain victim-blaming by judges, that is a scandal. So I did what the podcasts train you to do. I pulled the report’s own sources. Every judgment it quotes is published and free to read on the National Archives Find Case Law service. Every statistic it cites has an underlying document. I checked 31 of the report’s quoted judgment-uses against the primary judgments, and its eight load-bearing statistics against their sources. I used AI tooling to fan the retrieval out and then verified the results against the primary documents, and every claim below carries a reference so you can do the same in minutes. That last part is the point. None of this required special access. It required only the willingness to click through.

Here is what I found.

Three checks anyone can run in five minutes

First check. The report’s “distraction” section quotes NT v LT [2020] EWHC 1903 (Fam) as an example of a judge diverting attention from abuse by praising a father: “The father is a loving father who will not abandon his son… he has a large home, businesses, an extended family.” Open the judgment and go to paragraph 74. It begins: “Mr Harrison submits that:”. The passage is not the judge speaking. It is counsel’s submission being recorded, as every judgment must record the parties’ arguments. Better still, Mr Harrison was acting for the mother, the father in the case was the abducting parent, and the mother won. A mother’s own barrister’s argument, in a case the mother won, is presented as evidence of judicial bias against mothers.

Second check. The report quotes ADK v ASI [2022] EWHC 2610 (Fam) at paragraph 88 for the phrase that a mother “was herself responsible for severing the children from their father”. The judgment ends at paragraph 86. There is no paragraph 88. The sentence exists, but it is a quotation within the judgment from a different Court of Appeal authority, and if you read that authority in full its thrust is protective of mothers who remove children from risk: the court’s focus must be on the children’s situation, not the source of the risk. A quotation attributed to a paragraph that does not exist, lifted from a case that says the opposite of what it is deployed to show.

Third check. The report cites a 2017 joint study by Cafcass and Women’s Aid for the statistic that 62% of private law cases involve domestic abuse allegations. That study is real and the figure is accurate. But the same study, in the same document, reports what courts actually did with those allegations: where abuse was alleged, courts ordered unsupervised contact less often (39% against 48%) and ordered no direct contact nearly twice as often (19% against 11%). In other words, the report’s own cited source contains outcome data showing courts responding more cautiously where abuse is raised, which runs directly against the report’s central thesis. The prevalence half of the study is quoted. The outcome half is absent.

I want to be fair about the base rate here. Across the 31 quoted judgment-uses I verified, one was fully defensible, seven were arguable, and twenty-three misrepresented the source in some material way: counsel’s submissions coded as judicial statements, a passage condemning a father coded as excusing him, judicial praise of a mother (“this was a conspicuously child centred decision”) coded as character assassination, and multiple cases where the mother substantially won, sometimes with her allegations proven and contact refused, presented as victim-blaming exhibits with the outcome unmentioned.

The methods problems underneath

The citation errors are symptoms. The design problems are the disease, and there are four.

There is no comparator. The gender-bias claim needs a comparison: how do judges write about fathers in equivalent positions? The study codes only language directed at mothers who allege abuse. No father-directed language was coded, no male-complainant cases were examined, no baseline exists. Without that, the data cannot distinguish “judges are biased against mothers” from “judges test the account of whoever makes allegations”, which is what fact-finding on the balance of probabilities requires them to do.

The sample is not what it is described as. The pilot corpus was openly salted: 27% of its cases were selected because they were already known to contain victim-blaming, and another 23% because the organisation’s own founder had acted as counsel in them. For the main analysis, an AI tool pre-highlighted “potential victim-blaming” in 337 judgments, and the annotators then worked through them in platform order for as long as time allowed, reaching 91. The report later describes these 91 as “randomly selected”. A time-limited walk through a machine-pre-highlighted list is not a random sample, and pre-highlighting is a documented way to anchor human annotators. On top of that, published judgments are themselves a skewed sliver: fewer than one in ten judgments are published at all, and appeals are heavily over-represented, which means the corpus over-samples exactly the cases where something went wrong and was being corrected.

The reliability statistic is missing. The report says inter-annotator agreement was calculated using Fleiss’ kappa. The value appears nowhere in the document. In content-analysis research this is close to disqualifying on its own: the standard methods texts say plainly that unreported reliability leaves every downstream conclusion open to doubt. The annotators were the organisation’s own staff, including two of the report’s authors, and disagreements were settled in internal consensus meetings. If the kappa was good, publishing it would have cost one sentence.

The scheme cannot lose. A judge who notes an inconsistency in evidence is coded as “discrediting”. A judge who records a father’s submission is coded as victim-blaming by paraphrase. A judge who found, after four years of hearings and retracted coached allegations, that a mother had run a deliberate campaign against the father is coded as “role reversal”. And in my favourite example, a judge who explicitly warned against the “gold digger” trope, in a case where he then awarded the mother roughly ten million pounds, is coded as an obvious instance of victim-blaming on the theory that mentioning the stereotype implants it. Deploying a stereotype scores. Warning against it scores. Careful language in judgments is read as concealment, and the report says the true scale is “likely underestimated” for that reason. There is no observation the design could make that would count against the hypothesis. In this community we have a word for hypotheses like that.

What survives, because something does

Skepticism that only cuts one way is just another bias, so here is the strongest version of the report’s case. Its flagship example, JH v MF [2020] EWHC 86 (Fam), is genuine and it is awful: a judge reasoned that an allegation could not be rape because the complainant had not physically resisted, and the appeal court demolished that reasoning in terms, as it deserved. Re H-N in the Court of Appeal records further first-instance failures. Individual instances of judicial victim-blaming exist, are documented, and getting them corrected requires the money and stamina to appeal, which many people do not have. That is a real problem and nothing in this post argues otherwise.

The report’s transparency recommendations are also right. Fewer than a tenth of family judgments are published. Litigants are silenced by a 1960 contempt provision the Law Commission has already recommended repealing. Those reforms deserve support from everyone, whatever they believe about bias, precisely because more published judgments would let claims like this report’s be tested properly, in either direction. It is worth noticing that the strongest reforms the report asks for were all in motion before it was published.

The takeaway

I went looking for cherry-picking and found something more instructive: a closed loop. A campaigning organisation with a pre-committed position selected its own corpus, coded it under its own commercial partner’s unpublished taxonomy, used its own staff as annotators, reported no reliability figure, validated the result with the partner’s software, and arrived at its founding premise to one decimal place. The report’s central theory is that institutions immersed in a culture reproduce that culture’s assumptions in language they believe to be neutral. As a description of the report itself, that is the most accurate sentence in it.

The same trap is waiting on the other side of this argument, and I want to name that clearly. There is a fathers’ rights literature that runs the identical loop with the sign flipped, and the small number of proper case-file studies we have, the ones that examined outcomes rather than language, found little evidence of systemic gender bias in either direction. The boring, evidence-shaped conclusion is that the family courts get individual cases wrong in both directions, each side’s advocacy literature samples its own tail, and the single reform that would discipline everybody’s claims is the one everybody already agrees on: publish the judgments.

For a worked example from the other direction, here is a post deriving a 56 to 60% false-allegation rate by comparing allegation rates in court cases with population prevalence. The trouble is that a private law case sits behind three selection gates: the relationship ended, the arrangements were disputed, and the dispute reached court rather than settling. Each gate changes the mix in ways nobody has measured, and plausibly in both directions at once, because abuse ends relationships and drives contested contact, and high-conflict litigation is also where tactical allegations thrive. The comparison cannot tell you the false-allegation rate, high or low, and the within-court studies we do have support neither “vanishingly rare” nor “mostly false”. Same loop, opposite sign, same missing denominator. Both sides of this argument are doing arithmetic with a denominator nobody has.

A thought

Bias exists, and no one has more trouble seeing it than the biased. Every perpetrator thinks they are a victim; no one is the villain of their own story. Some may realise it at 2am when alone, but mostly, even after being shown to be wrong, people cling to the belief that they were right and the system is unjust.

Of the problems that are real and clear, delay does the most damage, and I will defend the strong version of this: delay in examining evidence, or ignoring it, is functionally child abuse. It can detach children from valid, important relationships, leave them with abusers, or keep victims under stress that transfers straight to their children. The attachment research says severed parental relationships harm children in measurable ways, and the appellate courts have seen cases where years of delay put the damage beyond any court’s power to repair. Courts cannot be weaponised in the same way if they act fast.

Control of access to children is coercive control. But spend an hour on social media and practically everything gets that label, including clear, well-defined boundaries. The politics of division, so visible in left/right social media, is applied just as readily to gender and to every other line we can find to divide ourselves along, and this is not entirely organic. The US Senate Intelligence Committee documented Russian influence operations deliberately amplifying divisive social issues, gender and sexuality debates among them. One peer-reviewed study found ten state-linked accounts drove over a third of a sampled feminism conversation on Twitter, arguing both sides at once, #MeToo and #FeminismIsCancer from the same operation. Simulation work suggests a small percentage of bots in a network can tip what people perceive the majority view to be, and the majority illusion research shows we misread opinion prevalence from our feeds even without interference. People believe their views are held far more widely than they are, partly because someone is paid to make it look that way. The UK’s own Russia Report described the astroturfing and declined to measure its impact, which is itself a data gap.

So what would help? Quicker resolution from social services, police and family courts, ahead of everything else. More investment where the reviews keep pointing: training, capacity, and better oversight. And consequences with real weight where an allegation is proven fabricated, because that protects genuine victims most of all; the honest inherit the credibility the fabricators burn. But be careful with the words here, because I nearly got this wrong myself. The family court runs a binary system: the burden sits with the accuser, and an allegation not proved on the balance of probabilities is treated in law as if the event had not happened (Re B [2008] UKHL 35). That is not the same as a finding that the allegation was false, which is a separate finding with its own consequences, and the difference matters. Much of the unfairness people perceive lives in the gap between the two: the interim months in which an untested allegation does the work of a proven one.

On evidence, my practical advice to anyone in this position is to record, openly. The Family Justice Council published guidance on exactly this in May 2025, and the line it draws is the line I would draw. There is no legal reason you cannot record your own meetings with professionals; Cafcass’s own position is that there is no legal reason a parent should not record their own interview. So ask, announce it, and expect it noted. Openly recording high-conflict interactions such as handovers protects everyone, including you, from claims about what was said, and I have seen accounts of statements supposedly made to children that a recording would have settled in seconds. What the courts condemn, correctly, is the covert end: in one reported case a father sewed recording devices into his child’s clothing, and the recordings were admitted, as evidence of his parenting. Do not tape children’s questioning, do not run a surveillance campaign against the other parent, which can itself amount to harassment even when announced, and never make a child the device. Openness is necessary but not sufficient: record openly, proportionately, and of your own interactions.

One more observation from the research, offered carefully because it is an inference rather than a measured result. The documented reality is that identical conduct reads differently depending on who you are. The Harm Panel recorded that mothers who resist contact risk the implacable-hostility and alienation labels, while male victims report being read against a “real victim” stereotype and recast as perpetrators, and the credibility-discounting literature describes the same effect in other jurisdictions. Meanwhile court process varies enormously by geography; private law application rates range from 9 to 183 per 10,000 families across England and Wales. I take something narrower from this than the usual advocacy claims on either side: advice calibrated to the average case is unreliable in a system with this much variance and this much position-dependence. And hold me to my own standard here, because the case-file studies I cited approvingly above found little gender bias in outcomes. The asymmetry I am describing is documented in the labels and the process. It is not documented in the outcomes, and I am not claiming it is.

For my part there is little I can legitimately do beyond listening. The restrictions on sharing information about live proceedings are comprehensive, and testing them is one of the easiest ways to damage a case you care about. Good luck to everyone out there. I am glad I am not in that position.

The cheapest fix

If one thing from all of this deserves to survive contact with your scepticism, it is this. Nearly every failure documented above, on every side, is a data-publication failure before it is a bias failure. The report has no comparator corpus. Its mirror-image critics have no measured denominator. The reliability statistic was computed and not published. Fewer than one in ten judgments are published at all, fact-finding outcomes are not collated anywhere, and the UK’s intelligence oversight committee described an influence operation and declined to measure it. Everyone in this argument is reasoning from samples they cannot see the edges of, and filling the gaps with their priors.

Training programmes are expensive and contested. Court reform is slow. But publishing the data, judgments, outcomes, fact-finding results, disaggregated and routine, is cheap, and it is the one reform every side of this argument already endorses. Publish better data. Then everything can be examined, this post included.

Check my working. Every source is below.

Sources

Report: Right to Equality, Scratching the Surface (June 2026), righttoequality.org.

Judgments (all free on caselaw.nationalarchives.gov.uk): NT v LT [2020] EWHC 1903 (Fam) at [74]; ADK v ASI [2022] EWHC 2610 (Fam); JH v MF [2020] EWHC 86 (Fam); Re H-N [2021] EWCA Civ 448; DN v UD [2020] EWHC 627 (Fam) at [67]; A v B [2023] EWHC 1864 (Fam); FJ v LT [2023] EWHC 1783 (Fam) at [76]; H v N [2020] EWHC 1863 (Fam); X v Y [2025] EWFC 62 (B); Re B (Children) [2008] UKHL 35; M v F (Covert Recording of Children) [2016] EWFC 29; Re C (A Child) [2015] EWCA Civ 1096.

Statistics: Cafcass and Women’s Aid, Allegations of Domestic Abuse in Child Contact Cases (2017); JCIO Annual Report 2024-25; Lowe and Stevens, HCCH Global Report (2023); Burton and Hunter, Everyday Business (Domestic Abuse Commissioner, 2025).

Outcome studies: Hunt and Macleod, Outcomes of Applications to Court for Contact Orders (2008); Harding and Newnham, How Do County Courts Share the Care of Children (Nuffield, 2015); Trinder et al, Enforcing Contact Orders (Nuffield, 2013).

Methods: Lombard, Snyder-Duch and Bracken, Human Communication Research 28(4) (2002); McHugh, Biochemia Medica 22(3) (2012); Krippendorff, Content Analysis.

Transparency and guidance: Law Commission, Contempt of Court, Law Com No 423 (2025); McFarlane P, Confidence and Confidentiality (2021); Family Justice Council, Covert Recordings in Family Law Proceedings Concerning Children (May 2025); Hunter, Burton and Trinder, Assessing Risk of Harm (MoJ, 2020); Epstein and Goodman, Discounting Women, 167 U. Pa. L. Rev. 399 (2019); Hine, Bates and Wallace, Journal of Interpersonal Violence (2022).

Influence operations: US Senate Select Committee on Intelligence, Russian Active Measures, Volume 2 (2019); Bradshaw and Henle, International Journal of Communication 15 (2021); Intelligence and Security Committee, Russia Report, HC 632 (2020); Lerman, Yan and Wu, The Majority Illusion, PLOS ONE (2016); Ross et al, European Journal of Information Systems 28(4) (2019).

Worked example (other direction): empathygap.uk, False Allegations in the Family Courts (September 2020), discussed in the takeaway above.

James Davies