Across the different ways AI can support legal marketing, from briefing it properly to planning content and drafting social posts, the same handful of mistakes tend to recur. None of them are exotic. Most come down to asking AI to make a judgement it has no basis for, rather than asking it to carry out a task with real information already supplied.
Hallucinations come from asking AI to know things it does not know. Overstatement comes from asking it to sound impressive rather than accurate. Over-automation comes from asking it to decide what is worth publishing.
This guide brings those mistakes together in one place, as a practical checklist of what to watch for.
1. Hallucinations: confident and wrong
AI can state something false with exactly the same confidence as something true, and nothing in its tone signals the difference. This is usually described as “hallucination”, and it covers fabricated case citations, invented statistics, misattributed quotes, and details that sound plausible but do not check out.
The legal profession already has well-known cautionary examples. In 2023, lawyers in a New York personal injury case, Mata v Avianca, submitted a court filing containing six fabricated case citations generated by ChatGPT. When challenged, they asked the tool itself to confirm the cases were real, and it did, wrongly. The court fined the lawyers and required them to notify every judge falsely named as the author of the fake opinions.
The courts of England and Wales have since faced the same problem. In June 2025, the Divisional Court heard two cases together, R (Ayinde) v London Borough of Haringey and Al-Haroun v Qatar National Bank, in which non-existent authorities had been put before the court. It referred the lawyers involved to their regulators. In Al-Haroun, the solicitor said he had relied on research his client had produced using publicly available AI tools.
These cases involved legal research rather than marketing, but the lesson transfers directly. A fabricated award, an invented statistic about a practice area, or a misremembered case outcome in a piece of marketing content is the same failure in a lower-stakes setting. The safeguard is the same too: verify facts against a real source, not by asking AI whether its own claim is correct.
2. Overstatement: claims that outrun the evidence
AI writing tends toward confident, superlative language: “leading”, “trusted”, “comprehensive”, “unparalleled”. In most sectors this is merely bland. In legal marketing, it is a compliance risk. For solicitors, paragraph 8.8 of the SRA Code of Conduct for Solicitors requires publicity to be accurate and not misleading, and paragraph 7.1(c) of the Code of Conduct for Firms applies the same duty to firms. For barristers, rule rC19 of the BSB Handbook prohibits misleading clients about their services, and the BSB’s guidance at gC57 warns that publishing inaccurate or misleading advertising could breach it, as the Bar Council’s guidance on advertising and website profiles sets out.
As covered in how to brief AI for legal marketers, AI defaults to this kind of generic, unsubstantiated language when it is not given real, specific detail to work from. The fix is to replace vague superlatives with specific, checkable facts: the actual practice areas covered, a real case outcome, a genuine ranking or accreditation, rather than a claim invented to sound impressive.
3. Tone drift: sounding less measured than intended
Tone can shift gradually and almost invisibly as AI edits, condenses, or rewrites a piece of content, particularly when asked to make something “punchier” or “more engaging”. A paragraph that starts out measured and factual can end up sounding closer to a sales pitch, without any single edit looking wrong in isolation.
This is a particular risk in short-form content, such as social posts, where there is less room for nuance to survive editing. Checking condensed or rewritten content specifically for tone, separately from checking it for accuracy, is worth building into the review process. A factually correct draft is not automatically a well-toned one.
4. Confidentiality: what should never go into an AI tool
Client names, unpublished case details, draft settlement terms, and other confidential material should not be entered into general AI tools as part of drafting or editing content. Some platforms, particularly free or consumer versions, may use submitted content to train their models unless settings say otherwise. Even where that risk is managed, submitting confidential information to a third-party tool without client consent raises its own professional conduct questions.
Unlike the other mistakes on this list, a confidentiality lapse cannot be fixed at the review stage. Once information has been entered into a tool, it has already left the organisation’s control. This risk has to be managed before anything is typed in.
The practical approach is to keep AI input generic and anonymised wherever possible: describe a scenario without naming the client, use placeholder details, and add real specifics back in manually once the drafting is done. Enterprise tools used under a firm’s own data processing agreement are a different matter, governed by that firm’s policies rather than by assumption.
5. Publishing without review: the mistake that makes the others worse
Most mistakes on this list become far more serious without a human review step before publication. Hallucinated facts, overstated claims and tone drift can all be caught at the review stage, and are all much harder to undo once something has gone out under the chambers or firm’s name.
AI-assisted content should be reviewed with the same rigour as any other public-facing material, and arguably more. AI can produce something that reads as polished and confident while being wrong in ways that are not obvious on a quick read. Treating AI output as a draft that needs checking, rather than a finished product, is the most important habit on this list.
6. Over-automation: efficiency at the cost of substance
AI can make content production faster at almost every stage: drafting, planning, condensing, scheduling. The risk is treating that speed as the goal in itself, rather than a means to producing content that is still worth publishing.
This shows up as calendars filled with interchangeable, generic topics, social posts that could belong to any chambers or firm, and a steady stream of output that meets a publishing schedule without saying anything distinctive. AI should reduce the effort required to produce good content, not replace the judgement about what is worth saying in the first place.
Frequently asked questions
What is an AI “hallucination” and why does it matter for legal marketing?
A hallucination is when AI states something false with the same confidence and fluency as something true, with no signal that it is uncertain or wrong. In legal marketing, this might mean a fabricated statistic, a misattributed quote, or an invented case detail. AI-generated errors are often well written and plausible, which makes them harder to spot than an obvious mistake. Verifying facts against a real source is essential rather than optional.
Is it ever safe to put client information into an AI tool?
Not into a general or consumer AI tool, unless the information has been made generic or anonymised first. Some platforms may retain or use submitted content, and entering client details without consent raises professional conduct questions. Describing a scenario without identifying details, then adding real specifics back in manually, is the safer approach. Enterprise tools used under a firm’s own data processing agreement are a separate matter, governed by that firm’s policies.
How can a legal marketing team tell if AI-generated content has drifted in tone?
The most reliable way is to compare the AI-edited version against the original source material specifically for tone, not just for facts. Look for language that has become more enthusiastic, more superlative, or more promotional than the source, particularly after asking AI to shorten or “punch up” a piece of writing. A dedicated tone check, separate from a fact check, catches drift that a general read-through often misses.
Does using AI for legal marketing content increase compliance risk?
It can, if content is published without adequate review. AI has no inherent understanding of which claims are properly evidenced or which phrases risk misleading a reader, and it defaults to confident, superlative language. Both the SRA (paragraph 8.8 of the Code of Conduct for Solicitors) and the BSB (rule rC19 of the Handbook) require marketing not to mislead. With proper checking before publication, AI-assisted content should carry no more risk than any other reviewed material.
What is the single most important safeguard when using AI in legal marketing?
Human review before publication. Most of the other mistakes (hallucinated facts, overstated claims, tone drift) are recoverable if caught before content goes live, and far more damaging once published. Confidentiality is the exception: it has to be managed before anything is entered into an AI tool, because review cannot undo it.
Sources
- Mata v Avianca, sanctions order (S.D.N.Y., 22 June 2023)
- R (Ayinde) v London Borough of Haringey; Al-Haroun v Qatar National Bank [2025] EWHC 1383 (Admin)
- SRA: Marketing your services to members of the public (warning notice quoting paragraph 8.8)
- Bar Council Ethics: Advertising and website profiles (November 2025, quoting rC19 and gC57)
- Square Eye: How to brief AI, for legal marketers
