UK immigration law is not a forgiving subject. The rules are dense, frequently amended, and riddled with cross-references. An incorrect answer is not just unhelpful — it can lead someone to submit a doomed application, miss a critical deadline, or worse, believe their status is secure when it is not.
So we ran an experiment.
We designed a scenario so legally complex that even an experienced caseworker might miss one or two of its traps. We gave it simultaneously to three AI tools: Immigranta, Google AI Mode, and ChatGPT. We then evaluated each response against a detailed model answer constructed from official gov.uk sources.
The results were illuminating — not just for what each tool got right, but for why each got things wrong.
The Test Scenario: Dr Priya Chandrasekaran
Meet Priya. She is a Nigerian-Indian academic researcher who entered the UK on a Tier 4 Student visa in October 2018 to complete a PhD. She switched to a Skilled Worker visa in September 2022 and has been working at a Russell Group university ever since.
On the surface, she looks like a straightforward ILR applicant. In practice, her history contains at least seven legally material events, each of which engages a different part of the Immigration Rules — and several of which interact with each other in ways that are not immediately obvious.
Here is the compressed version of what happened:
- She got a promotion in January 2024, which triggered a new Certificate of Sponsorship with the same employer and occupation code, just weeks before major salary reform rules came into force in April 2024.
- Her university’s sponsor licence was downgraded from A-rated to B-rated in August 2024 following a Home Office compliance audit.
- She took four months of unpaid parental leave from November 2024.
- In June 2025, her department was absorbed into a new “Life Sciences Institute.” Her contract was reissued under the new name. TUPE applied. Nobody told the Home Office.
- In October 2025, she enrolled on a part-time evening MBA at a different institution.
- In November 2025, she submitted an ILR application, citing five years of continuous lawful residence, and included a B1 English language certificate taken in October 2023.
The question to each AI was precise: interpret how the Immigration Rules apply to each material fact. Identify every point at which the rules engage, the specific provision engaged, and how it operates on these facts.
The Scorecard
Before we get into the detail, here is how each tool performed when assessed across four dimensions: issue identification, rule precision, ambiguity handling, and structure.
| Dimension | Immigranta | Google AI Mode | ChatGPT |
|---|---|---|---|
| Issue Identification | 8/10 | 4/10 | 7/10 |
| Rule Precision | 9/10 | 2/10 | 5/10 |
| Ambiguity Handling | 9/10 | 2/10 | 8/10 |
| Structure and Synthesis | 7/10 | 3/10 | 7/10 |
| Overall | 78/100 | 28/100 | 68/100 |
The gap between Immigranta and the others is not marginal. It reflects something structural about how each tool was built.
The Central Trap: Did Anyone Spot It?
Before diving into each tool’s answer, it is worth explaining what the single most important legal issue in the scenario was — because it is the issue that determines whether the entire ILR application succeeds or fails before any other factor is even considered.
Priya entered the UK in 2018. She submitted her ILR application in November 2025, claiming five years of continuous lawful residence. Seven years in the UK. Sounds fine, doesn’t it?
Here is the problem.
Under Appendix Skilled Worker (SW 21.1 and SW 21.2), the five-year qualifying period for settlement must consist of time spent on specific routes. Skilled Worker is one of them. Tier 4 Student is not. Priya was a student until September 2022. Her qualifying Skilled Worker clock therefore started then — not in 2018.
By November 2025, she has approximately three years and two months of qualifying residence. Not five. Her ILR application fails at the very first hurdle, regardless of everything else in the scenario.
Any AI tool that misses this has failed the fundamental test. Every other issue in the scenario — the B-rated sponsor, the TUPE reorganisation, the English language certificate — becomes secondary.
How Each Tool Handled It
Immigranta: Grounded in the Rules Themselves
Immigranta identified the qualifying period failure immediately, anchored it to the specific provisions (SW 21.1 and SW 21.2), and stated the consequence clearly: the application fails. It then went further and correctly identified a second independent ground for refusal that the other tools either missed or dismissed.
The English language certificate was taken in October 2023. The application was submitted in November 2025. Under Appendix KoLL, an English language test must have been taken no more than two years before the date of application. October 2023 plus two years equals October 2025. The application was in November 2025. The certificate was expired by approximately one month.
Immigranta caught this. It worked through the date arithmetic precisely, identified the alternative route under Appendix KoLL (where leave was previously granted on the basis of a B1 qualification), and correctly noted that the alternative route could not be assumed to be satisfied on the stated facts because the scenario did not confirm the required precondition.
Two independent refusal grounds, both identified, both correctly analysed, both anchored to specific provisions.
On the more complex issues — the January 2024 CoS and the transitional arrangements question, the B-rated sponsor’s effect specifically on a settlement application under SW 24.1, the TUPE corporate structure question — Immigranta handled ambiguity in exactly the right way. It did not paper over the gaps with false confidence. It identified precisely where the rules engage, precisely where they are silent, and precisely what additional facts would be needed before a definitive view could be formed.
This is what rigorous legal analysis looks like. It is also, not coincidentally, what a RAG system built on authoritative official sources produces when the retrieval is working correctly.
Google AI Mode: The Danger of Confident Wrongness
Google AI Mode’s response is the most concerning of the three — not because it was simply wrong, but because it was confidently and structurally wrong in ways that could mislead a user into false security.
The response was formatted as a compliance audit with emoji-coded status badges: green ticks for compliant items, red alerts for violations. It looked authoritative. The actual legal analysis underneath it did not hold up.
The ILR qualifying period. Google’s timeline showed ILR eligibility as September 2027, which correctly implies that the five-year Skilled Worker clock started in 2022. But the response never confronted the actual submitted application in November 2025. The entire answer was structured as a forward-looking compliance plan rather than an assessment of the application Priya had already submitted. The question asked for an assessment of a live application. Google answered a different question.
The MBA enrolment. Google marked this as fully compliant with no authority cited. The statement “Skilled Worker visa holders are legally permitted to undertake supplementary study without needing a separate Student Visa” was presented as settled fact. It is not that simple — the conditions of leave attached to a Skilled Worker visa contain study provisions that require analysis, not assertion. No provision was cited. No qualification was made.
The sources. This is where the architecture of Google AI Mode becomes visible. The citations in the response included a YouTube short, a Reddit thread, and multiple third-party law firm blogs. These are research starting points for a human lawyer. They are not authoritative sources for a legal analysis. The Immigration Rules, Appendix Skilled Worker, Appendix KoLL, the sponsor guidance on gov.uk — none of these were directly cited with specific paragraph references.
The English language certificate. Not addressed at all.
The overall picture. Google AI Mode produced an answer that looks like a professional legal assessment. The structure, the formatting, the action plan sections — they create an impression of thoroughness. But the core analysis was wrong, the sources were not authoritative, and the most critical issue in the scenario was answered as a planning observation rather than an assessment of a live application. A person relying on this answer to proceed with their ILR application would be misinformed.
This is the specific risk that AI tools drawing on general web content — rather than the Immigration Rules themselves — carry in a legal context.
ChatGPT: Knowledgeable but Imprecise
ChatGPT’s response was considerably better than Google’s. The qualifying period failure was correctly identified. The Tier 4 exclusion was stated clearly. The TUPE reorganisation was handled with appropriate nuance — rather than asserting a breach, ChatGPT correctly identified that the consequence depends on whether the Life Sciences Institute is a legally distinct entity, and held the analysis open on that basis. The ambiguity taxonomy was disciplined throughout.
But it fell short in two specific ways that matter.
Rule citation precision. ChatGPT consistently identified the right appendix and framework without descending to specific paragraph numbers. “Appendix Skilled Worker settlement provisions” is named but SW 21.1, SW 21.2, SW 24.1, SW 24.4(b) are not cited. “Appendix KoLL” is referenced but paragraph 2.2 and the specific routes within it are not identified. At the level of a complex application or a tribunal proceeding, it is the specific paragraph that governs — not the appendix name.
The English language certificate. ChatGPT noted the October 2023 certificate and identified ancillary uncertainties around provider approval, but did not perform the obvious date arithmetic. October 2023. November 2025. Two-year window. Expired. A senior practitioner should catch this as a straightforward calculation before moving to the more nuanced questions. ChatGPT did not.
This pattern reflects what training-data-based AI does: it gives you the shape of the law without always giving you the precise text. For many questions that is adequate. For complex immigration analysis — where a single month’s difference in a certificate date can determine an application’s outcome — it is not.
Why the Difference Exists
The gap between these three responses is not a gap in intelligence. All three tools are sophisticated. The gap is architectural.
ChatGPT draws on its training data — a vast corpus of text that includes immigration law, commentary, and guidance. It has absorbed a great deal of knowledge about UK immigration. But it cannot always distinguish between the rule as it was, the rule as it is, and the commentary about the rule. It also cannot cite the precise current text of a provision because it does not retrieve it — it reconstructs it from memory. In a domain where rules change via Statements of Changes published on gov.uk, sometimes several times a year, this matters enormously.
Google AI Mode draws on live web content, which solves the currency problem in theory. But web content about immigration law includes law firm blogs, community forums, visa guidance from universities, YouTube videos, and Reddit threads alongside the actual Rules. A system that weights these equally — or that cannot distinguish an authoritative provision from a secondary commentary about it — will produce citations that look like research but do not constitute a reliable legal foundation.
Immigranta is built differently. It is a Retrieval-Augmented Generation system whose knowledge base consists of authoritative official sources: the Immigration Rules themselves, Statements of Changes as they are published, Home Office caseworker guidance, and gov.uk policy documents. When it answers a question, it retrieves from that corpus and generates a response grounded in the retrieved text. It does not reconstruct rules from memory. It does not pull from law blogs or Reddit. The citations point to the actual provisions.
This is not a small difference. In the Dr Chandrasekaran scenario, the decisive issues — the Tier 4 exclusion from SW 21.2, the English language two-year window under Appendix KoLL paragraph 2.2, the B-rated sponsor effect under SW 24.1 — are answerable only by reading the specific text of those provisions. A system that retrieves that text directly will almost always outperform one that reconstructs it from secondhand sources.
What This Means for Caseworkers and Applicants
For a caseworker, the lesson is specific: the tool you use is only as reliable as the sources it draws from. An answer that cites Appendix Skilled Worker without citing the paragraph number is giving you a map without coordinates. In a casework decision, that gap is the gap between a correct determination and a judicial review.
For an applicant navigating their own case, the stakes are different but not lower. Immigration decisions are rarely reversible quickly. An ILR refusal does not simply delay settlement — it can trigger cascading consequences for leave, for dependants, for employment. Getting a confident wrong answer from an AI tool that drew on a blog post from 2023 is arguably worse than getting no answer at all, because it removes the instinct to seek further verification.
The Bottom Line
Three AI tools. One complex scenario. One clear winner — not because Immigranta is cleverer, but because it is built on the right foundations.
The scenario was designed to surface exactly the kind of compound complexity that real ILR applications contain: layered transitional arrangements, sponsor compliance issues, ambiguous corporate restructures, ancillary requirements with precise date-sensitivity. These are not edge cases. They appear in thousands of applications every year. And they are precisely the issues that general-purpose AI tools — trained on the internet’s general knowledge of immigration law — struggle to handle with the precision that real legal analysis requires.
Want to test Immigranta on your own immigration question?
Ask Immigranta at immigranta.co.uk
Immigranta is a RAG-based tool built specifically for UK immigration information, drawing exclusively from authoritative official sources including the Immigration Rules, Statements of Changes, and Home Office guidance. It provides cited, source-grounded answers — not reconstructed summaries, not blog commentary, not confident assertions without a provision to back them up.
Immigranta provides immigration information, not legal advice. For complex personal circumstances, always consult a qualified and regulated immigration adviser.