Can an Algorithm Put Your Marriage Under Suspicion? The Home Office, AI Risk Scoring and the Right to Family Life

1. The October 2026 investigation: what has been established

On 27 October 2026, Open Rights Group published Automated Discrimination: Unpacking the Digital Hostile Environment, a substantial research report by Fabio Chiusi. It draws on published research, expert interviews and Home Office documents obtained through freedom-of-information requests. It describes the expanding use of algorithms and AI in immigration case management, asylum procedures and border operations.

The report covers multiple systems. Its chapter on the sham-marriage triage tool is especially important for JSH Law because it addresses a very personal legal question: how the state decides whether to subject a proposed marriage or civil partnership to additional investigation.

The Home Office does not need a campaigning organisation to tell us that automated screening exists. Its own statutory guidance, last updated on 31 March 2025, confirms that referred proposed marriages are screened against immigration records and intelligence-based risk profiles using a mainly automated process. It then requires a trained decision-maker to assess whether reasonable grounds exist to suspect a sham.

The Open Rights Group report also revisits earlier reporting about the use of machine learning in the triage system from 2019 and historical disparities affecting certain nationalities. That is a significant research claim. However, the published statutory guidance does not disclose the current model architecture, version, validation dataset or detailed risk-profile criteria. We should not pretend to know those things.

Sources: Open Rights Group’s report and recommendations; full research PDF; Home Office statutory guidance.

2. What counts as a sham marriage — and what does not?

The expression ‘sham marriage’ is emotionally loaded. It is also a statutory concept. Under the relevant provisions of the Immigration and Asylum Act 1999, a sham marriage or civil partnership involves requirements including the absence of a genuine relationship and an intention by one or both parties to circumvent immigration controls. The precise statutory definition and procedural scheme must be read together.

A relationship is not a sham merely because the partners have different nationalities, a substantial age difference, an unconventional history, a language barrier, prior relationships or immigration difficulties. Equally, genuine concerns about deliberate attempts to evade immigration law may warrant lawful investigation. Neither a stereotype nor a blanket presumption resolves the question.

Marriage and civil partnership rights are not the same thing as an automatic entitlement to a visa, settlement or entry to the UK. A lawful decision about permission to marry must be distinguished from subsequent immigration decisions, which have their own rules and possible routes of challenge.

Who is ordinarily referred?

The referral scheme can apply where the couple includes a person who is not a ‘relevant national’ and the required evidence of an exemption, appropriate immigration status or specified visa has not been provided. It is not a scheme that automatically subjects every international couple or every non-British spouse to investigation.

The definition of a relevant national in the current guidance includes British and Irish citizens, people with qualifying EU Settlement Scheme status, and certain people with pending earlier EUSS applications. Exemptions and visa categories are technical and fact-sensitive. The published guidance explains the current categories; a person’s status should be checked rather than inferred from nationality alone.

If a referral is made, that is an administrative step. It is not itself a finding that the couple’s relationship is false.

3. The technical pipeline: from notice of marriage to an investigation

The scheme is governed principally by Part 4 of the Immigration Act 2014, including section 48. A simplified view helps show where automation enters and where legal judgment is supposed to begin.

The Home Office guidance expressly states that a decision to investigate will not be taken directly by the automated assessment tool. The trained official is required to review the whole of the available information, and the presence of one or more risk factors must not be the sole determinant.

That is a meaningful safeguard on paper. The unanswered technical question is how its performance is measured: what does the reviewer see, what contrary information is provided, and what audit trail records an independent judgment?

What can happen during investigation?

Depending on the circumstances, investigation may involve requests for information, evidence, photographs or interviews, including interviews at home, in Home Office premises or remotely. Compliance with lawful requirements matters. The consequences of non-compliance and any decision that a marriage is a sham depend on the statutory scheme, applicable notices and later decisions.

A 70-day extension is significant in practical terms, but it is not a 70-day automatic ban on marrying in every case: it is a statutory notice framework following a properly authorised investigation decision.

4. The risk-scoring problem: a red light is not evidence of an insincere relationship

A screening model can be technically good at finding patterns in historical records and still be poor at identifying the present legal question. That is not a paradox. It depends on what the system was trained or designed to detect, what labels were used, which populations are represented and whether the historical records accurately reflect genuine behaviour.

If historical immigration enforcement focused more intensively on particular communities, a system trained on those outcomes may learn patterns of enforcement attention rather than patterns of fraudulent marriages. That is a potential feedback-loop and proxy-discrimination problem, not a finding about the current code or the outcome of any particular case.

The Bureau of Investigative Journalism reported in April 2021 that the triage system had adopted machine learning and that an internal assessment showed very different historical red-flag rates for some nationality groups. The Home Office disputed that its triage system was indirectly biased and pointed to equality assessment as a means of identifying and justifying risk differences. Those opposing positions must be reported accurately.

Five questions an independent audit should answer

1. Target variable: Was the tool evaluated against an independently verified finding of a sham relationship, or against referrals, historical refusals or enforcement outcomes?
2. False-positive rate: How often was a genuine couple flagged for investigation, and did that rate vary by group or by location?
3. Selection effects: Were past referrals shaped by officers’ discretion, data quality or enforcement priorities, producing biased labels?
4. Version history: What changed when the system was updated, and was the effect on affected groups assessed?
5. Human override: How often did officials reject a system flag after reviewing relevant evidence, and were reasons recorded?

A simple overall ‘accuracy’ percentage cannot answer these questions. The public interest lies in how false flags are generated, whether they are corrected and whether the system is fit for the statutory threshold it supports.

5. The human decision-maker: safeguard or rubber stamp?

The Home Office’s published decision rule is explicit: the model does not itself take the section 48 investigation decision. An official must have reasonable grounds to suspect a sham and must assess the available information as a whole. This matters legally, but the phrase ‘human in the loop’ is not a complete governance description.

Meaningful review requires access to adequate source information, time and authority to challenge the automated output, and an opportunity to take account of evidence against the risk indicator. A reviewer whose only options are to accept a red flag or send the case onward is not necessarily performing the judgment that statutory guidance envisages.

The distinction between meaningful review and superficial sign-off is also relevant to the amended UK GDPR. Under the Data (Use and Access) Act 2025, section 80, the new Article 22A framework asks whether a significant decision was based solely on automated processing, defined by the absence of meaningful human involvement. Article 22C provides safeguards for certain significant solely automated decisions, including information, representations, human intervention and the ability to contest the outcome.

That does not automatically mean this screening tool breaches Article 22C. First, automated triage and the final investigation decision are different acts. Secondly, data-protection processing regimes and exemptions depend on the legal purpose and nature of the controller’s processing. Thirdly, where there is genuinely meaningful human involvement, the special solely automated safeguards may not apply in the same way, while wider public-law, equality and data-protection duties remain relevant.

The ICO confirms that the Act’s data-protection changes were fully in force by 19 June 2026. Older online summaries that refer only to the previous Article 22 should therefore be checked against the current framework. ICO guidance on the amended law.

6. Nationality, proxy discrimination and ministerial authorisation

The Home Office guidance states that risk profiles will not take account of nationality or other protected characteristics unless specific ministerial authorisation is in place for that purpose. It refers to the exception in Schedule 3, Part 4, paragraph 17 of the Equality Act 2010. That exception is legally technical; it is not a general exemption from equality scrutiny or a blank cheque to discriminate.

The same published guidance says that adding such a factor would require consideration of whether ministerial authorisation was necessary. It does not publicly establish that the current model directly uses nationality as a scoring input. Direct use, indirect correlation and disproportionate impact are different propositions.

Even a model that excludes nationality as a named field may produce different outcomes among nationality groups if other variables correlate strongly with it. That possibility requires evidence and testing; statistical disparity alone does not identify its cause or settle whether the legal treatment is justified.

The public sector equality duty in section 149 of the Equality Act 2010 imposes duties of due regard when exercising relevant public functions. Proper evaluation is not established merely by attaching an ‘equality checked’ label to a system. Decision-makers need to understand relevant impacts, evidence and mitigations.

In 2023, Public Law Project publicly announced a pre-action legal challenge raising concerns over indirect nationality discrimination, transparency, the public sector equality duty and human review of Home Office sham-marriage screening. That announcement records allegations and the initiation of proceedings: it is not a court judgment establishing every alleged breach. We have not verified a final determination of those pleaded claims, and this article does not invent one.

7. The right to marry and family life: Articles 8 and 12

The right to marry is protected by Article 12 of the European Convention on Human Rights, given domestic effect through the Human Rights Act 1998. Article 8 protects respect for private and family life. Article 14 may also be relevant where discrimination affects the enjoyment of Convention rights. These rights are important, but none means that every proposed marriage investigation is automatically unlawful.

States may regulate the formalities of marriage and take proportionate steps against abuse of immigration law. The legal questions include whether the measure is authorised by law, pursues a legitimate objective, operates without unlawful discrimination and respects the substance of the protected right.

There is a practical reason for scrutinising delays. A couple may have booked a ceremony, arranged family travel, taken unpaid leave or made plans affecting children. A 70-day notice period can impose costs even if the eventual investigation finds no sham. That does not make every delay unlawful; it means the legal assessment must not ignore the real effect on the people concerned.

A right to marry is not a right to residence

A valid genuine marriage does not, by itself, entitle a spouse to enter, remain in or settle in the United Kingdom. Immigration applications have separate substantive requirements. A couple challenging the investigation of a proposed marriage may therefore need to distinguish the registration decision from any visa, leave or enforcement decision. The remedies and deadlines may differ.

The important danger of automated screening is not simply a slow decision; it is the possibility that an inaccurate or improperly understood risk indicator makes a lawful relationship appear suspicious before an official applies the correct legal test.

The Convention text is available in Schedule 1 to the Human Rights Act 1998.

8. What should a couple do if their proposed marriage is investigated?

The first step is not to accuse the Home Office of using illegal AI. It is to understand the exact notice received, the statutory basis of the investigation, the applicable period and any information requested. A risk flag is not necessarily the reason recorded in the eventual decision.

1. Preserve the notices. Keep copies of the original registration documents, all Home Office correspondence and the dates on which the parties received them.
2. Check status and exemption evidence. Confirm whether the Home Office has the correct immigration records and whether any relevant visa or status documentation was omitted.
3. Identify the legal stage. Are you dealing with a referral, an investigation decision, requests for evidence, compliance determination or a separate immigration refusal?
4. Respond to lawful requirements promptly. Follow the actual notice, seek clarification where needed and request appropriate adjustments or interpretation if relevant.
5. Keep evidence authentic. Preserve genuine relationship records in chronological order; do not create or embellish correspondence to satisfy an imagined algorithm.
6. Identify errors specifically. If a record states an incorrect date, relationship, immigration status or identity, locate reliable documentary evidence of the mistake.
7. Ask proportionate questions. Where lawful and appropriate, seek information about the process and whether automated screening materially contributed, without assuming entitlement to confidential source code or third-party data.
8. Seek specialist help promptly. A regulated immigration adviser or solicitor can identify the correct route for challenging an investigation, adverse compliance finding or separate immigration decision. Some challenges have short time limits.

A neutral request for clarification

Re: Marriage/civil partnership investigation — [reference]

Please confirm the statutory decision and the date on which it was made, the current deadline and the requirements with which each party must comply. If any personal record or immigration-status information has materially informed the decision and is disputed, please identify the appropriate process for requesting its correction. Please also confirm how the parties may seek clarification or raise relevant accessibility and safeguarding concerns.

We will comply with lawful requests for information and would be grateful for confirmation of the correct contact details and response method.

This is illustrative wording, not a prescribed request or a claim to disclosure of security-sensitive risk rules. Individual advice is particularly important where an investigation interacts with immigration enforcement, domestic abuse, refugee protection or a child’s immediate welfare.

9. Can a couple obtain the algorithm, profile or data behind the flag?

The answer depends on the type of information, how it is held, the legal basis for withholding it and the process used. No single ‘send an SAR’ response resolves every issue.

The earlier PLP litigation and information-rights proceedings illustrate that demands for internal scoring criteria may encounter legitimate disputes over secrecy, fairness and public interest. Conversely, an explanation of what the model does, who reviewed the individual case and how an error can be corrected may often be possible without publishing the full technical criteria.

If a child or another vulnerable person is affected, an otherwise justified information request still needs to respect relevant confidentiality and data-protection rights. Do not make public accusations against a named officer or disclose protected records merely to highlight concerns.

10. Why JSH Law is writing about an immigration screening tool

JSH Law’s principal work concerns Family Court support, evidence integrity, safeguarding and access to justice. This is an immigration and public-law story with an unmistakable human-rights and family-life dimension. But its direct legal consequences should not be confused with an English private-law children case.

A marriage-screening algorithm does not decide where children live, whether a parent has contact or how a Family Court must assess domestic abuse. Immigration enforcement decisions and child arrangements orders are separate legal acts. The family-justice relevance is the broader need to preserve accurate information and maintain accountable decisions where government systems affect the lives of parents and children.

A carefully labelled hypothetical

Imagine a couple who have lived together, care for a child and plan to marry. Their notice is referred; an administrative system flags a record because of an old, incorrectly linked immigration application. A trained officer then has to decide whether there are reasonable grounds to suspect a sham.

If the officer checks the underlying records, resolves the identity error and reviews the genuine relationship evidence, the process could work as intended. If the officer treats the automated flag as conclusive, the couple might experience avoidable scrutiny. We do not know which occurs in any particular case without examining the evidence.

A person whose records are inaccurate may benefit from careful documentary organisation. That work — identifying original sources, comparing competing records and preparing a concise factual chronology — is familiar in JSH Law’s non-reserved support work. Where the issue concerns immigration rights or remedies, the substantive advice must come from an appropriately regulated specialist.

11. A technical accountability standard for high-impact relationship screening

What should a government system be able to demonstrate if its output can trigger a legally significant investigation into an intimate relationship? JSH Law proposes the following seven checks as a governance framework, not an existing statutory test.

1. Defined decision boundary: document exactly what the model can flag, which decisions it cannot make and the legal threshold reserved to the human official.
2. Source traceability: retain reliable provenance for the immigration records, reports and historical data used by the system.
3. Bias testing: evaluate rates of flags, false positives and outcomes by relevant groups, while accounting for lawful restrictions on sensitive information.
4. Independent human assessment: show that a trained official can inspect contrary evidence and reject a machine recommendation without procedural pressure.
5. Clear audit logs: record the tool version, key inputs, output, human response, overrides, and any corrections without keeping personal data for longer than justified.
6. Effective correction and challenge: provide clear and accessible routes to correct inaccurate underlying data and raise concerns about an individual decision.
7. Independent evaluation and public explanation: publish sufficient methodology and aggregate evidence to support democratic scrutiny, with proportionate redaction of genuinely sensitive information.

None of this requires revealing every live enforcement threshold to the world. A system can protect information whose disclosure would enable evasion and still explain its legal purpose, governance, validation, limitations and the process by which an individual decision can be reviewed.

Public authorities and suppliers should address these questions at procurement and design stage, not wait until an affected couple must litigate to discover how the decision-making workflow works.

12. Difficult questions this investigation leaves unanswered

Is the system finding shams, or finding the kinds of couples officials historically suspected?

A predictive system’s performance depends on its target labels. If a dataset treats historical referrals or investigations as a proxy for sham marriages, the output may inherit the selection choices of previous officials. Independent outcome validation is vital.

Can a ‘red flag’ remain influential even after human review?

Yes, as an analytical possibility: initial risk labels can anchor attention or frame what evidence is sought. Whether that occurs in this system requires actual workflow and behavioural evidence, which the public guidance does not provide.

Can legitimate immigration enforcement justify disproportionate consequences for genuine couples?

Lawful regulation can be justified; proportionality and non-discrimination remain questions requiring evidence. The fact that a system pursues an important public goal does not establish that every criterion or outcome is defensible.

If private source code stays confidential, is meaningful accountability impossible?

Not necessarily. There are degrees of disclosure. An accurate explanation of the system’s purpose, testing, limitations, human review and correction rights may be possible without releasing exploitable operational detail.

What evidence should the public expect in 2026?

At minimum, a serious evaluation of whether flagged cases correlate with independently established shams, how false positives are detected, what group differences arise, and whether trained officials actually exercise statutory judgment. These are proposed standards for accountable governance, not claims that the current Home Office has or has not satisfied each one.

13. Frequently asked questions

Will the Home Office investigate every couple where one partner is not British?

No. Referral depends on the statutory scheme and status or visa exemptions. Even where a referral is made, it does not automatically lead to an investigation. The official must consider whether the legal conditions are met.

Does a machine automatically extend the notice period to 70 days?

No. Home Office statutory guidance says a trained decision-maker makes the decision to investigate. A properly authorised investigation results in the extension. The machine is used for preliminary screening.

Does the new report establish that the system is currently unlawful?

No. Open Rights Group raises serious concerns supported by earlier investigations and litigation history. Specific findings of unlawful discrimination or improper delegation require evidence and legal determination. We have not verified a judgment conclusively establishing the current system’s unlawfulness.

Can someone marry if an investigation concludes that their relationship is genuine?

The outcome depends on statutory notice, compliance and other legal requirements. A genuine relationship does not override all formalities, but a risk flag alone is not a permanent bar to marrying.

Are all marriage screenings carried out using generative AI?

No such conclusion follows from these sources. The Home Office describes mainly automated risk-profile assessment, while historical reporting says the triage system adopted machine learning. That is different from claiming ChatGPT or another generative model independently makes marriage decisions.

Can the couple challenge incorrect data?

Potentially, through applicable personal-data rights, administrative procedures and specific legal remedies. The appropriate route and any exemptions depend on the information, controller, decision and procedural stage. Specialist advice may be needed.

Does the right to marry also give a spouse immigration status?

No. The right to marry and immigration entitlement are legally distinct. An immigration application must satisfy the applicable requirements.

14. Primary authorities, original research and further reading

Original report and reporting

• Open Rights Group, Fabio Chiusi: Automated Discrimination: Unpacking the Digital Hostile Environment (27 October 2026)
• Open Rights Group: full 105-page research report, especially the sham-marriage triage section around printed pages 37–38
• The Bureau of Investigative Journalism: Alice Milliken, Home Office algorithm to detect sham marriages may contain built-in discrimination (19 April 2021)
• Public Law Project: Legal action launched over sham marriage screening algorithm (17 February 2023)

Primary law and official guidance

• Home Office: Marriage and civil partnership referral and investigation scheme — statutory guidance (31 March 2025)
• Home Office: Marriage investigations — operational guidance
• Immigration Act 2014, section 48
• Immigration Act 2014, section 50
• Immigration and Asylum Act 1999, section 24
• Human Rights Act 1998, Schedule 1 — Articles 8, 12 and 14
• Equality Act 2010, Schedule 3, Part 4, paragraph 17
• Equality Act 2010, section 149
• Data (Use and Access) Act 2025, section 80
• ICO: changes to data protection and automated-decision law

Related JSH Law analysis

• The Invisible Algorithm: Why People Cannot Challenge Public Decisions They Don’t Know AI Helped Make
• The Next Frontier in Family Justice Is Decision Integrity
• Family Court disclosure and records guide
• School safeguarding records are wrong: challenging inaccuracies