The Defamation Act 2013 and a Wrong AI Answer About Your Company: What Section 1 Says, and the Record to Keep

Section 1 of the Defamation Act 2013 is short, and it is the first thing to read when an assistant states something serious and wrong about your company. Subsection (1) reads: "A statement is not defamatory unless its publication has caused or is likely to cause serious harm to the reputation of the claimant." Subsection (2) adds a second limb for a business: "For the purposes of this section, harm to the reputation of a body that trades for profit is not 'serious harm' unless it has caused or is likely to cause the body serious financial loss." That is the whole of the section. It carries no figure, no table and no threshold in pounds, and whether a particular statement clears it is a question for a solicitor on your own facts. This page quotes the Act, reports what we have measured about AI answers, and stops there. Nothing here is legal advice.

What a director can do on the first day is different work. An answer from an assistant is not a page at a fixed address: it is produced on request, varies between sessions, and is gone when the window closes. So the first job is to turn it into a record another person can run again.

Reading section 1 for yourself takes about a minute

The text is published free at legislation.gov.uk, Defamation Act 2013, section 1. Four things are worth taking from the official page, which we read on 16 October 2026.

• The section has two subsections and nothing else. Subsection (1) sets the serious harm requirement for every claimant. Subsection (2) adds the serious financial loss limb and applies it to a body that trades for profit.
• There is no number in it. A pound figure quoted anywhere as "the section 1(2) threshold" is not in the Act.
• Section 1 took effect on 1 January 2014, under article 2 of S.I. 2013/3027.
• The extent marker on the section is England and Wales. Scotland has its own statute, the Defamation and Malicious Publication (Scotland) Act 2021, which the amendment note on section 15 of the 2013 Act records as having repealed that section for Scotland in 2022.

The two words the Act hands back to the general law

Section 15 is the interpretation section and it is unusually relevant here. It defines a "statement" as "words, pictures, visual images, gestures or any other method of signifying meaning", which is wide and deliberately medium neutral. For "publish" and "publication" it does something different: it says they "have the meaning they have for the purposes of the law of defamation generally". The Act therefore does not define publication itself, and nothing in section 1 or section 15 addresses a statement produced by a machine on request.

So the question most British directors arrive with, whether an assistant's answer is a publication and by whom, is not answered by reading the Act. That one goes to a solicitor. The useful part is that the facts a solicitor asks for are the facts a measurement produces anyway.

The citation numbers we hold, and the rate we do not

The dataset behind this is a run of 630 questions on 16 October 2026, put to ChatGPT and Gemini in three markets, which recorded 1,260 usable readings. Our own domain appeared in 6 readings of the 1,260, and the British count was 3 out of 210 on ChatGPT with none out of 210 on Gemini. The 20,143 citations recorded were dominated by self published pages: 82.0% of them, against 0.7% on review directories. The first audit, eleven days before and built on a different question design, had already landed in the same place: 3 of 597, with 85.6% of its citations on a vendor's own site. Two designs, one answer. Gemini is the starker half, with not one citation of our domain in 630 readings on the later run and none in the 291 of the earlier, 921 without a single hit. The questions and the counts are published in what AI assistants actually cite across 597 readings. The British slice, including the two public sector domains the American questions never produced, sits in the British target list.

Two limits belong beside those figures. Both runs asked only ChatGPT and Gemini. Perplexity, Claude and Google AI Overviews sit inside what the product reads and outside what either dataset measured, and the two are never merged into one sentence here. What neither run counted is the frequency of a false statement about a company, which is why no such rate appears with our name on it. What the source mix does say is useful here: when 82.0% of citations land on somebody's own website, the sentence contradicting you is usually on a page, and a page has an owner, a date and an address.

The record to build in the first hour

One row per reading, in one sheet. These are the fields that make a reading repeatable by somebody who was not there.

Field	Why it carries weight later
Date and time, with the timezone	Retrieval behaviour changes, so a reading describes one moment
The assistant and the mode, named exactly	Search on and search off are different products
Signed in or not, and memory on or off	A session that knows who you are answers differently
The prompt text, copied rather than described	A paraphrased prompt cannot be run again
The answer verbatim, with a screenshot of any citation list	The exact wording is the thing in issue
The one sentence carrying the claim, quoted on its own line	This is what a solicitor reads first
The sources it names when asked which pages it read	A claim about its own behaviour until you open each one
Two more runs in fresh sessions with memory off	One occurrence is not yet a pattern
The verified source, meaning the URL you opened yourself	Claimed and verified sources are different columns

A claim holding up across all three runs points at a source somewhere; one appearing in a single run points at the session. Keep the sheet monthly, because the dated series is what shows whether a correction held.

Then the part that changes the answer

Once the sentence is traced to a page, the work is ordinary and it belongs to you rather than to a legal department: correct the fact where it originates, publish your own version in plain text with the date you verified it, and take the reading again. The sequence, with the error classification that decides which fix applies, is in how to correct a wrong fact an AI states about you. Where the question is instead what a search engine will take out of its results and how long each route runs, the realistic timelines are in what can be removed from search results, what cannot, and how long each route takes.

Questions about the Defamation Act 2013 and wrong AI answers

An assistant said we are in liquidation. Does that meet the section 1(2) test?

Section 1(2) sets the test and puts no figure on it. Whether a particular statement clears it is for a solicitor who can see your accounts, your pipeline and the publication. What is in your hands today is the record: the verbatim answer, three repeats, the date, the engine, and the page the claim traces back to.

Is an answer from an assistant a publication?

The Act does not say. Section 15 gives "publish" and "publication" the meaning they have in the law of defamation generally rather than defining them, and section 1 does not mention machine generated statements. Take the question to a solicitor and bring the capture sheet.

Does section 1 apply to a Scottish company?

The extent marker on section 1 reads England and Wales. Scotland has the Defamation and Malicious Publication (Scotland) Act 2021, and the amendment note on section 15 of the 2013 Act records that section as repealed for Scotland in 2022. Jurisdiction is among the first things a solicitor settles.

How many readings before I treat a wrong statement as a pattern?

Three fresh sessions with memory off, logged individually, including the rows where the answer was fine. Monthly repeats on unchanged prompt text turn those rows into a series.

Do you publish a rate for how often assistants get a company badly wrong?

No, and we will quote one when we have run it. What we publish is the citation measurement above: 1,260 readings on 16 October 2026 across ChatGPT and Gemini, 6 citing our own domain, replicating a 597 reading run eleven days earlier.

Should I report the answer to the provider too?

Yes, and log what you sent and the date in the same sheet. The feedback route inside each answer is cheap to use, and the log entry makes the request checkable later.

The sentence is the asset, not the outrage

A director who finds an assistant saying their company is in administration wants it gone within the hour. What survives the week is a dated record of exactly what was said, in which session, traced to the page it came from, beside the correction and the date it went live. That record is what a solicitor works from, what a publisher acts on, and what tells you six weeks later whether anything moved. We built AI Knows Us to take those readings on a schedule across five surfaces and keep every raw answer, and we charge for it, which is the disclosure this page owes you before you weigh anything on it.