Using ChatGPT, Claude, Grok or Perplexity for Your Employment Claim: What Generic AI Gets Wrong

AI is genuinely useful when you are facing a problem at work. It explains jargon, it is available at 2am, and it does not judge you. It is easy to see why people reach for ChatGPT, Claude, Grok, Gemini or Perplexity to help with a grievance, a settlement offer, or a tribunal claim.

But there is a difference between a general-purpose AI tool and one built for a specific, high-stakes legal job. On UK employment tribunal claims, that difference can cost you your case. This guide explains exactly where generic AI falls down - not to say "never use AI" (this platform is an AI tool), but to show what to trust it with and what not to.

Why this matters: an employment tribunal claim runs on strict deadlines and precise figures. A confident but wrong answer - the wrong time limit, an out-of-date cap, a made-up case - is worse than no answer, because you act on it without knowing it is wrong. The problem with generic AI is not that it is always wrong. It is that it is confidently wrong in ways you cannot see.

Which AI are we talking about?

People reach for several very different tools, and it is worth separating them:

• Chatbots - ChatGPT (OpenAI), Claude (Anthropic), Grok (xAI) and Gemini (Google). These generate answers from what they absorbed in training. They are the most prone to inventing case law and quoting figures that are years out of date.
• Answer engines - Perplexity, and the "search" or browsing modes of the chatbots above. These search the web and cite sources as they answer, so they are usually more current and less likely to fabricate a case outright.

The distinction matters, but it does not make the second group safe for a tribunal claim. An answer engine is only as good as the pages it happens to find - it can cite an outdated blog, a US website, or a marketing article with the same confidence as an authoritative source. And whichever type you use, none is scoped to England and Wales law, none calculates your specific deadline, and none has a human checking the result. The rest of this guide applies to all of them; where a point is specific to one type, it says so.

The core problem: confident, and sometimes wrong

Large language models are built to produce fluent, plausible text. They are not built to be correct, and they have no built-in way to tell you when they are unsure. That combination - high fluency, no reliable uncertainty signal - is exactly the wrong shape for legal information, where being 90% right and 10% confidently wrong can be worse than being unsure throughout.

Here are the specific failure modes that matter for a UK employment claim.

1. It can miss the deadline that bars most late claims

This is the single biggest risk. In England and Wales you generally have 3 months less 1 day from the dismissal or the act you are complaining about to start Acas Early Conciliation, and missing it almost always bars the claim entirely (section 18A, Employment Tribunals Act 1996). The exact date depends on your facts, how early conciliation pauses the clock, and which claim type applies.

A general AI tool does not know your dates, does not calculate the deadline reliably, and will not chase you as it approaches. It may even state the limit as a flat "3 months", which is subtly wrong. Deadlines are the one thing that should never be estimated by a language model - they should be calculated in code from your dates. See the deadlines guide for how the timing really works.

2. It quotes out-of-date figures

Compensation caps, the week's pay limit, and the Vento injury-to-feelings bands change every April. A model trained on older data will happily quote last year's cap, or a figure from three years ago, as though it were current. For claims presented on or after 6 April 2026, for example, the week's pay cap is £751 and the compensatory award cap is £123,543 - but a generic tool has no reliable way to know that, and no reason to flag that its number might be stale.

For anything involving money, the figures must come from a source that is deliberately kept current - not from a model's training data of unknown vintage.

3. It blends US law into UK answers

The training data behind mainstream AI models is dominated by US content. UK and US employment law are fundamentally different: the US has "at-will employment" (you can be fired for almost any reason), no equivalent of the statutory unfair dismissal regime, different discrimination frameworks, and no Acas. Ask a general model an employment question and you can get an answer that quietly imports American concepts - "right to work" doctrine, "wrongful termination" as understood in the US, or references to bodies that do not exist here.

The scope restriction matters: a tool built for this should be locked to England and Wales employment law and should say so.

4. It invents case law and citations

There are now well-documented instances, including in reported court cases, of AI tools generating fake legal citations - realistic-looking case names and references that do not exist. Courts and tribunals have warned litigants about relying on AI-generated authorities for exactly this reason. If you cite a case in your claim that turns out to be invented, you damage your credibility with the tribunal at the worst possible moment.

This is mainly a chatbot problem (ChatGPT, Claude, Grok, Gemini working from memory). Answer engines like Perplexity that cite sources are less likely to fabricate a case outright - but they can still present an outdated, US, or low-quality page as authority. Either way, verify every case, section number and figure against the official record on legislation.gov.uk or the tribunal's own published decisions before you rely on it.

5. There is no human checking the output

This is the difference that a chatbot cannot close on its own. A general AI tool gives you an answer and stops. No one reviews whether it fits your facts, whether the deadline is right, or whether the document reads well to a judge. For a decision as consequential as a tribunal claim, the absence of a human quality check is the gap that matters most.

6. Your sensitive data may not be private

Employment disputes involve highly sensitive information - medical details, allegations, other people's names. With consumer AI tools, what you paste in may be stored outside the UK or used to improve future models, depending on the settings and the terms you agreed to. Before you paste evidence into any tool, it is worth reading how it handles your data, and preferring one that tells you plainly.

What generic AI is genuinely good for

None of this means AI is useless for your situation - the opposite. Used well, it is a real help:

• Understanding concepts - what "constructive dismissal" or "protected disclosure" means in plain English
• Getting your thoughts in order before a grievance or a meeting
• Preparing questions to ask a specialist, Acas, or a support service
• Drafting a first version of something you will then check carefully

The line is simple: use it to learn and prepare, not as the final word on anything that carries a deadline, a figure, or a consequence.

What "purpose-built" actually changes

A tool designed for UK employment claims is not just a chatbot with a legal skin. The things that make generic AI risky are the things it is built to fix:

• Deadlines are calculated in code from your actual dates, not estimated by a model
• Figures are kept current and drawn from the statutory source, not from training data of unknown age
• Scope is locked to England and Wales employment law, so US concepts do not leak in
• Answers are grounded in a maintained knowledge base rather than generated from memory
• A human checks the output that matters before you rely on it

That is what Aricase is built to do. AI builds your case. A human checks it.

That last point - the human check - is the one no general model can replicate on its own. The help options compared guide sets out how the different routes - going it alone, generic AI, a purpose-built tool, and a solicitor - stack up, and the self-representation guide covers what representing yourself really involves.

Key takeaway

Generic AI - ChatGPT, Claude, Grok, Gemini and answer engines like Perplexity - is a powerful way to understand your situation and prepare, and it is fine to use it that way. What it is not is a safe substitute for a tool built for the job. It can miss your deadline, quote stale figures, blend in US law, invent cases, and hand you all of it with total confidence and no human checking. For an employment tribunal claim, where a single wrong date or figure can end an otherwise strong case, the safest approach is to let AI help you learn - and to rely on correct, current, human-checked information for the decisions that count.

_This article describes documented failure modes of general-purpose AI in the context of UK employment claims, based on publicly reported findings. It is legal information, not legal advice; check current figures and deadlines against the official sources linked above._