Who is liable when an AI tool gives bad legal information? In South Africa, the answer depends on more than whether the chatbot displayed a warning. It may involve the service provider, the platform presenting the answer, a professional who relied on it and the person who acted on it.
A wrong answer does not automatically establish a claim for compensation. But a disclaimer does not automatically defeat one either.
The important questions concern what was promised, which duties applied, how the information was used and whether the error caused a legally recoverable loss. Those questions require attention to the facts, even when the mistake originated in software.
A disclaimer cannot answer every liability question
Many AI tools warn that their output may be inaccurate and should not be treated as legal advice. Those warnings matter. They can help explain the service’s limits and influence whether reliance on an answer was reasonable.
However, the surrounding presentation matters too.
A general writing assistant that clearly identifies its limitations creates different expectations from a subscription service advertised as providing dependable answers to South African legal problems.
If prominent marketing promises expertise while a less visible notice disclaims responsibility, the relationship between those statements deserves examination.
The promise and the warning must be considered together
Where the Consumer Protection Act applies, section 41 addresses misleading representations. Sections 48 and 49 deal with unfair terms and notices limiting liability, while section 51 prohibits certain exclusions, including liability attributable to gross negligence. Their application depends on the transaction.
This does not mean that every disclaimer is invalid. Nor does it mean that every inaccurate answer amounts to misleading marketing.
It means that the provider’s conduct and contractual terms need to be assessed together. A warning is part of that assessment, rather than the end of it.
Several parties may be involved
An AI answer can reach a user through a chain of businesses and professionals.
One company may develop the underlying model. Another may build a legal interface around it. A platform may decide which documents it searches, how it answers questions and when it recommends professional assistance.
A law firm may then incorporate the system into its work.
These roles should not be treated as interchangeable. Being involved in the chain does not automatically make every participant liable.
The developer’s role
Questions about a developer might concern the capabilities it represented, the limitations it disclosed and its response to known problems.
A developer supplying a general tool may have different obligations from one contracted to deliver a system for a defined legal task.
The assessment should identify what the developer undertook to provide and what the person bringing the claim can establish.
The platform’s role
A platform controls important parts of the user’s experience.
It may label the service as a legal assistant, select its source material and decide whether answers appear alongside qualifications or warnings. It may also determine whether users can contact a human.
Those choices can influence expectations and reliance. They deserve separate consideration from how the underlying model was developed.
The professional’s role
A professional who uses AI in client work adds another relationship.
The client may never see the tool or its disclaimer. They receive an explanation, document or recommendation from the professional they instructed.
The question then includes whether that professional performed the service to the required standard.
An attorney cannot outsource professional judgment
AI may assist an attorney with research, drafting or organising a matter. It does not remove the need to understand the client’s instructions and check the work.
The Legal Practice Council’s Code of Conduct requires legal practitioners to carry out work competently and promptly, protect client confidentiality and provide appropriate professional advice.
Applied to AI use, those duties make verification particularly important. A plausible case citation is of little value if the case does not exist. A correct statement of foreign law may be unsuitable for a South African dispute.
Verification must go beyond reading the draft
Professional review should test the substance of the output.
That may include:
- Finding the original legislation or judgment.
- Checking that the authority says what the draft claims.
- Confirming the applicable jurisdiction.
- Identifying missing facts or qualifications.
- Checking dates, calculations and procedural requirements.
- Reviewing whether the advice fits the client’s circumstances.
An attorney who presents unchecked output as their own advice cannot assume that the software’s warning resolves their responsibility.
Nevertheless, professional liability still requires assessment. An error, disciplinary concern and damages claim are related questions, but they are not identical.
A wrong answer is not automatically a successful damages claim
Consider someone who asks a chatbot about a legal deadline and receives an incorrect answer. They act too late and believe they have lost an opportunity to bring a claim.
The inaccuracy is important. It does not, by itself, establish who must compensate them or how much is recoverable.
The legal basis could involve a contract, a delictual claim or another applicable remedy. Each has requirements that must be proved.
Financial loss raises particular questions
In Cape Empowerment Trust Ltd v Fisher Hoffman Sithole, the Supreme Court of Appeal considered a negligent misstatement claim involving financial loss. The claim failed because wrongfulness and legal causation were not established.
The case did not concern AI. Its relevance lies in showing that negligent information and resulting financial loss do not automatically establish delictual liability.
For an AI dispute, the claimant would need to identify the duty relied on, the relevant conduct and the connection to the loss.
Causation needs evidence
A claim should establish what the person would have done if they had received correct information.
Would they have acted in time? Was the underlying legal claim viable? Did another event contribute to the outcome? Was there an opportunity to correct the mistake?
These questions can be difficult, but they are central to assessing compensation.
The cost of the subscription may be straightforward to prove. The value of an allegedly lost legal claim may require a much more detailed investigation.
Reasonable reliance depends on the circumstances
Not all reliance on an AI answer occurs in the same setting.
A person casually exploring a topic through a free general chatbot is in a different position from someone paying for a service expressly presented as answering their particular legal problem.
Relevant circumstances may include the wording of the answer, the questions the system asked and the warnings shown before the person acted.
A confident instruction can create a different impression from a general explanation that identifies uncertainty and recommends checking a deadline.
The user’s conduct also matters
Users should preserve the facts honestly, including warnings they received and any independent advice available to them.
Did they provide the information the tool requested? Did they omit an important date? Did the answer expressly say that it could not determine the deadline?
Those details may affect the assessment. They should not be used to assume that the user is always responsible, just as an inaccurate output should not be used to assume that the provider is always liable.
Legal platforms should explain what service they provide
Public legal tools can help people understand terminology, prepare questions and recognise when they need professional assistance.
Problems arise when the service’s boundaries are unclear.
A platform describing itself as providing general information should examine whether its actual responses give personalised instructions that users are likely to treat as advice.
Explain limitations at the point of use
As a practical design measure, significant limitations should be visible when they matter.
A user asking about an urgent deadline needs more than a general warning elsewhere on the website. The response should avoid unjustified certainty and explain when the available facts are insufficient.
Useful safeguards may include:
- Asking for essential missing information.
- Identifying the jurisdiction addressed.
- Linking to original sources where appropriate.
- Distinguishing general explanations from professional advice.
- Directing urgent or complex matters to a qualified person.
- Providing a way to report inaccurate answers.
These are recommended safeguards, rather than a claim that one prescribed design applies to every legal platform.
Privacy creates a separate area of responsibility
A legal question may contain names, identity numbers, employment records or sensitive details about a dispute.
Even if the answer is accurate, the handling of that information can raise a separate legal issue.
POPIA addresses lawful processing, security measures and the handling of information by operators. Transfers outside South Africa may also require assessment under section 72.
Users and professionals should understand where information goes before entering confidential material. Platforms should explain their processing arrangements accurately.
A disclaimer about inaccurate answers does not answer a complaint about how personal information was handled.
Preserve the evidence when an answer causes harm
AI conversations can be difficult to reconstruct later. The same question may produce a different answer, and the platform’s terms or presentation may change.
Keep a record of:
- The question and information supplied.
- The complete answer and follow up exchanges.
- The date and time.
- Warnings displayed during the interaction.
- Advertising or promises relevant to the service.
- The applicable terms and payment records.
- The action taken and evidence of the resulting loss.
Screenshots can help, but a complete export may preserve more context.
If a deadline or live legal matter is involved, obtaining correct advice should take priority over pursuing a complaint. Prompt action may prevent further harm.
Accountability depends on the role and the evidence
AI can make responsibility harder to trace because several parties may influence one answer. It does not follow that nobody can be held accountable.
The developer’s promises, the platform’s presentation, the professional’s review and the user’s reliance may each matter. Whether more than one party is liable requires a legal basis for each claim.
The sound approach is to examine the actual service, the duties attached to it and the evidence of what happened.
A warning deserves consideration. So does a promise of expertise. Neither should replace the careful assessment needed when bad legal information causes harm.
