It is now quite easy for a client to take a lawyer’s opinion, put it into a chatbot and receive, within seconds, an answer that appears to say something different. The temptation is to treat that answer as a “second opinion”. It is not. A second opinion comes from another lawyer who has reviewed the facts, documents, governing law, procedure and commercial objectives, and who stands behind the advice. A consumer AI response is only a generated answer to a prompt – including all the prompt’s assumptions and omissions.
This is no longer a theoretical point. The Solicitors Regulation Authority in England and Wales has noted that technology-literate consumers will increasingly expect firms to use tools that improve service, not least because they use similar tools themselves. A 2023 LexisNexis survey found that, among consumers already aware of generative AI, 48% had used it for legal advice or help with a legal question. Those figures are not Hong Kong-specific, but they are a useful warning. Some clients will already have asked the machine before they ask the lawyer. Others will ask it afterwards.
There is a commercial edge to this as well. The Law Society of Hong Kong has recognised that barristers and solicitors need to show not only legal skill and judgment, but also cost-effectiveness where technology speeds up routine work. Clients are beginning to press the point directly. Financial Times reporting headlined “Wall Street banks push Big Law to cut fees because of AI” says Goldman Sachs, Morgan Stanley and Citigroup are telling outside counsel that bills should fall where AI has made work faster. The reported focus is on firms quantifying AI-driven savings, bidding work on that basis and sharing the benefit with clients, particularly in research, document review, due diligence and contract analysis.
That is the pressure point. Clients may still accept that lawyers are essential for judgment, strategy and accountability. They may be less willing to accept bills that assume every task still takes the same amount of junior-lawyer time. Firms will have to explain, in ordinary language, where AI has saved cost, where human review remains indispensable, and how the fee reflects both.
Why an AI answer can look more persuasive than it is
Part of the difficulty is tone. AI answers often read well. They are orderly, confident and quick. They may include apparent statutory references, case names and a conclusion that sounds more definite than the law allows. Good legal advice is often less tidy. It points out assumptions, evidential gaps, procedural risks and the questions that still need answering. To a worried client, caution can look like uncertainty, while the machine’s certainty looks like clarity.
That is where the false equivalence arises. The retained lawyer will usually have seen the contract, the correspondence, the parties’ roles, the governing-law clause, the available evidence and the client’s appetite for risk. The chatbot may have seen only a short paraphrase of one conclusion. It may not know whether the matter is governed by Hong Kong law, another common-law system, PRC law or the law chosen by the contract. It may blend jurisdictions, rely on obsolete material, invent sources or state a correct principle that does not fit the facts. The danger is not that every AI answer is wrong. The danger is that the client may not be able to tell when it is wrong.
Confidential information: the risk is more serious than a wrong answer
The first risk is often not the wrong answer, but the disclosure made to obtain it. A client may paste an advice note, draft contract, board paper, witness statement, correspondence, tax analysis or personal data into a public chatbot simply to get reassurance. In a Hong Kong matter, the material may also contain settlement communications, regulatory information, commercially sensitive material or third-party documents provided under express restrictions.
It would be too broad to say that every prompt is automatically public. Much depends on the provider, product, terms, settings, processing location and safeguards. A properly configured enterprise tool is not the same as a free public chatbot. Still, once information leaves the controlled lawyer–client channel, the client may lose practical control over it. The content may be stored, retained, accessed for service operation or used to improve the tool. It may also be difficult to trace where it has gone or ensure deletion.
A recent English and Welsh regulatory warning makes the point directly. The SRA referred to an Upper Tribunal observation that putting client letters and Home Office decision letters into an open-source tool such as ChatGPT is to place that information on the internet in the public domain. The SRA warned that this is likely to breach confidentiality and may result in the permanent waiver of legal professional privilege. That is not a statement of Hong Kong law, but it is a useful caution against treating a chatbot as a private search engine.
Privilege is especially sensitive. Whether and to what extent privilege is affected by a disclosure will depend on the applicable law and facts. Yet a client who discloses legal advice or litigation material outside the confidential lawyer–client setting may create a serious argument that confidentiality has been compromised. The harm may extend to another party’s data, confidential commercial information or personal data.
Hong Kong’s Privacy Commissioner for Personal Data has provided a local framework for thinking about this risk. Its 2025 checklist on the use of generative AI by employees is intended to help organisations develop internal AI policies in compliance with the Personal Data (Privacy) Ordinance. Its 2024 model framework similarly gives practical recommendations for procuring and using AI while safeguarding personal data privacy. The logic is directly relevant: before confidential matter information is entered into any system, the data flow, authorised access, retention, contractual protections, training settings and deletion arrangements should be understood. If those matters are uncertain, the information should not be uploaded.
When the client’s AI becomes the lawyer’s professional problem
This also becomes the lawyer’s problem very quickly. A client may say that an AI-generated authority supports a different argument, ask for it to be used in correspondence or pleadings, or use AI to prepare evidence. The lawyer cannot simply dismiss the concern. Nor can the lawyer treat the material as reliable because the client produced it. It has to be checked.
The UK High Court’s joined decision in Ayinde v London Borough of Haringey and Al-Haroun v Qatar National Bank QPSC illustrates the risk. The Court said freely available generative-AI tools are not capable of reliable legal research. They may give coherent answers that are entirely wrong, cite cases that do not exist and attribute quotations to real sources that say no such thing. If AI is used for legal research, the output must be verified against authoritative sources before anyone advises on it or puts it before a court.
In Ayinde, fictitious cases appeared in grounds for judicial review, together with a material misstatement of the relevant statutory provision. Wasted-costs orders were made against the barrister and the law centre, and both were referred to their regulators. The Divisional Court did not finally determine whether the barrister had used generative AI, but stressed that lawyers who use unchecked AI research risk severe sanction.
Al-Haroun is particularly relevant where the first draft comes from the client. The client accepted that inaccurate and fictitious material in his witness statement had been generated using publicly available AI tools, legal search engines and online sources. His solicitor relied on that material without independently verifying the authorities. The Court held that the client’s mistakes did not absolve the legal representatives: a lawyer was not entitled to rely on a lay client for the accuracy of citations or quotations filed with the court. That lesson travels readily. Client-supplied AI may be useful as a prompt for discussion. It is not a substitute for the lawyer’s own verification.
The problem is not limited to research. In R v FGD, an appeal from the Crown Court in a jury trial, the Court of Appeal considered a stay for abuse of process after the complainant used AI to prepare to give evidence. The AI had generated cross-examination-style questions and suggested answers about the witness’s own account. That went beyond familiarisation and into coaching. The distinction matters. In English criminal and civil proceedings, witness coaching is prohibited. Witness familiarisation is different: it prepares a witness for the process in the abstract, and may include mock questioning, but it must not rehearse the facts of the actual case or the witness’s evidence.
The US decision in Mata v Avianca, Inc. is another reminder. Lawyers filed a submission containing non-existent authorities generated through ChatGPT. The court imposed a US$5,000 penalty and corrective measures, noting that false authorities waste the resources of the opposing party and the court, and may deprive the client of arguments based on real precedent. Hong Kong courts may not take the same procedural route, but the professional point is the same: unchecked AI can harm the client it is meant to help.
A sensible response for Hong Kong practices and clients
A sensible response is to deal with the issue openly at the start. Engagement terms, onboarding notes and client portals can tell clients that AI-generated questions are welcome, but that privileged, confidential or personal information should not be put into public tools. Firms can also say that any AI output will be assessed against the actual facts and authoritative legal sources. That is not an attempt to shield advice from scrutiny. It is a reminder that important decisions should be made on professional advice, not on a chatbot’s prose.
Internally, firms need rules that people will actually use. The policy should identify approved systems and prohibited data, require source checks for material legal propositions, allocate supervision and sign-off, and explain how AI-assisted work is recorded and charged. It should apply whether the AI output comes from a colleague, vendor, in-house lawyer, witness or lay client. As Al-Haroun and R v FGD show, responsibility cannot simply be passed to the person, or the tool, that supplied the text.
Pricing needs the same candour. If AI has materially reduced routine work, clients will expect to see that reflected in fees. Law firms cannot credibly market efficiency and then price as if nothing has changed. Equally, they should not pretend that technology removes the need for legal supervision. The better answer is to distinguish commoditised work from work requiring judgment, record how technology was used, and price the matter in a way that shares efficiency without giving up independence, verification or responsibility.
For the client, the practical rule is simple: use AI to formulate questions, not to disclose the file, rehearse evidence or make the decision. Asking a lawyer, “Why does this output differ from your advice?” can be useful. Asking a public chatbot to analyse confidential advice, or to script answers about contested facts, may create a more expensive problem than the one the client was trying to solve.
AI does not reduce the value of a properly reasoned second opinion. If anything, it shows why one matters. A real second opinion requires legal expertise, access to the relevant material, verifiable sources, confidentiality safeguards, transparent pricing and professional accountability. Lawyers who explain that difference clearly will be better placed to protect clients, preserve trust and show the value that a confident machine-generated answer cannot provide.
The article was originally published on Hong Kong Lawyer
