It has been roughly two years since the major general-purpose language models became capable of producing, from a sentence-length prompt, a document that looks substantially like a UK will. The output is grammatical, uses the correct vocabulary, includes the standard recitals, names executors in the correct form, and produces a residue clause that disposes of the entire estate. It is, on inspection, a plausible-looking will.
It is not, on the present law of England and Wales, a valid one. The reason is structural and worth understanding.
What the Wills Act actually requires
Section 9 of the Wills Act 1837, as amended, requires a will to be in writing, signed by the testator (or by another person in the testator's presence and at the testator's direction), and the signing must be made or acknowledged by the testator in the presence of two or more witnesses present at the same time, and each witness must attest and sign the will (or acknowledge their signature) in the presence of the testator. The provision has been in force, in materially this form, for one hundred and eighty-nine years. It applies to the document, the act of signing, and the witnesses.
A language-model output, downloaded as a Word document, printed, and signed without witnesses, is not a will. The presence of two witnesses, signing in the testator's presence and in each other's presence, is what produces validity. The drafting of the words is secondary.
This is, of course, equally true of a will downloaded from a template website, dictated to a typist, or written longhand on the back of an envelope. The validity criterion is the same regardless of the route to the words. The language-model case is interesting only because the output's plausibility is high enough to invite the assumption that something more than a draft has been produced.
Where the model's output actually misleads
Three patterns we have observed in language-model wills produce specific problems that are not immediately visible.
First, the model occasionally generates clauses that read as legitimate UK will language but reflect provisions of US or Commonwealth law that do not apply in England and Wales. A reference to 'community property' or 'tenancy by the entirety' will pass cursory reading and fail legal effect. The model is not adapting to jurisdiction reliably.
Second, the model produces named beneficiaries from prompt context but does not validate that the named persons exist, are capable of taking, are not predeceased, or are not under disabilities that affect their capacity to inherit. A human drafter would normally check; the model cannot.
Third, the model's residue clause is typically a single beneficiary or a small list of named beneficiaries, with no fallback provision for the case where the named beneficiaries predecease the testator. A solicitor's standard residue clause includes a chain of substitutional gifts. The model's does not, unless explicitly prompted.
What the model is genuinely useful for
The model is a competent drafting assistant for the questions a testator needs to think about before instructing a solicitor — what assets exist, who the beneficiaries are, who the executor should be, what specific gifts should be made, whether any charitable legacy is intended, whether substitutional gifts are needed. These are the conversations a solicitor would have at the first meeting, and the model can have them in advance, on the testator's own time, at no cost. The output of that conversation is not a will. It is, more usefully, a brief for the solicitor's appointment.
Used in that frame, the model saves an hour of solicitor time and produces a better-thought-through instruction. Used as a replacement for the solicitor, it produces a document that the testator believes is a will and which, when the moment comes, is not.