AI Trends

Where AI genuinely helps legal research (and where it doesn’t yet)

PDSThe PDS Team·2026-04-28·7 min read
Glowing document pages dissolving into a soft network of light nodes, in gold and plum

An honest look at AI document review and research automation for professional practices.

AI can draft a research memo in the time it takes to make coffee, and for a busy Eastern Cape practice that speed is genuinely useful. The trouble is not what these tools do well. It is that the same model that summarises a 200-page bundle in seconds will also, with total confidence, invent a judgment that never existed and cite it to a real-sounding law report.

South African courts have already made an example of firms that trusted it too far. This is not a warning about some distant future. It is about how you use these tools this week: how to keep the speed without importing the risk, and where the line sits between a helpful assistant and a professional liability.

TL;DR: AI is a capable research assistant, not a legal authority. In Mavundla v MEC KZN, seven of the nine authorities a legal team cited simply did not exist, and the court referred the practitioners to the Legal Practice Council (The Conversation, 2025). Use AI to draft and summarise, verify every citation against the primary source, and never paste client data into public chatbots.

Where does AI genuinely help legal research today?

AI earns its place as a first-draft engine, not a final authority. It is fast and useful for the early, low-stakes stages of research: producing a rough memo to react to, summarising a long contract or bundle, translating dense clauses into plain language, and surfacing leads and search terms you then chase down yourself. The value is in the draft, not the verdict.

Think of it the way you would a bright but unsupervised candidate attorney. It reads quickly, writes fluently, and gives you a running start on a problem. It also states things it half-remembers as if they were settled law, misreads which court binds which, and cannot tell you when it is guessing. For document review and e-discovery, where you are narrowing a haystack rather than pronouncing on the law, that trade is often worth making. For an argument you put before a judge, it is not.

Used this way, AI genuinely saves hours. Our own view, from setting these tools up for professional firms, is that the productivity is real and the danger is specific and manageable. The firms that get burned are not the ones that use AI. They are the ones that mistake a first draft for a finished answer.

Why do AI tools invent case law that sounds real?

Because a language model predicts plausible text, not verified fact, so a citation that looks perfect can be completely fabricated. This is not a rare glitch. Legal researcher Damien Charlotin’s AI Hallucination Cases Database has logged more than 1,400 court decisions worldwide that dealt with AI-fabricated citations, still climbing by roughly 350 to 400 per quarter (Scientific American, 2026).

South Africa is on that list. In Mavundla v MEC: Department of Co-operative Governance and Traditional Affairs KZN, seven of the nine authorities the litigant’s team relied on did not exist, only two were genuine, and the Pietermaritzburg High Court referred the practitioners to the Legal Practice Council for citing unverified, AI-fabricated authority (The Conversation, 2025).

The second local case closes off the easy excuse. In Northbound Processing (Pty) Ltd v SA Diamond and Precious Metals Regulator, junior counsel used a tool called Legal Genius that claimed to be trained solely on South African legal sources, and it still produced fictitious citations. Acting Judge Smit held that a referral to the Legal Practice Council was warranted despite the apology, warning that admonishment alone is rarely enough when fake material is placed before a court (ITWeb, 2025).

If you think this is only a small-firm problem, consider that the South African government withdrew its own Draft National AI Policy in 2026 after News24 found at least six of the document’s 67 academic citations were AI-generated inventions, non-existent articles pinned to real journals and real authors (The Next Web, 2026). Hallucination is systemic. It is not a fringe user error.

What does the duty to verify actually require of you?

It requires you to check every citation, quote and holding against a primary source before it reaches a client or a court. Writing in De Rebus, the attorneys' journal, the point is put plainly: practitioners "remain accountable for the accuracy and ethical use of the AI tools they employ," and "human judgement remains paramount" (De Rebus, 2025).

In practice that means pulling up the actual judgment on SAFLII, reading the paragraph the AI quoted, and confirming the case exists, says what the model claims, and still binds. The duty is non-delegable. You cannot hand it to a machine, and "the AI said so" is not mitigation. The Northbound court made that explicit: an apology after the fact did not stop the referral.

This is the same professional standard that already governs a candidate’s work or a paralegal’s note. You would never sign off heads of argument on trust without checking the authorities. AI changes nothing about that duty. It only makes the temptation to skip the check more convenient, and the fabrication harder to spot, because a fake citation is formatted just as neatly as a real one.

Is typing client details into ChatGPT a POPIA problem?

Yes, and it is the quieter risk that firms underestimate. Under POPIA, the most serious data-protection offences carry a maximum penalty of a R10 million fine and up to 10 years' imprisonment, and the Information Regulator can impose administrative fines of up to R10 million (Michalsons, 2025). Pasting privileged client facts into a public chatbot is not a theoretical breach.

When you type identifiable client information into ChatGPT, Copilot or Gemini, that data leaves for servers outside South Africa. That is a cross-border transfer and, potentially, an unauthorised disclosure of information you are legally bound to protect. De Rebus is clear that clients must give informed consent before their personal information is processed through an AI tool (De Rebus, 2025), and a client who handed you a matter in confidence did not agree to it being fed to a foreign model.

This is the AI-shaped version of a duty your firm already carries. Our pillar on POPIA for Eastern Cape law firms walks through the wider obligations, and the step-by-step requirements sit in our POPIA compliance guide for law firms. The short version: confidential client data and public AI tools do not mix.

How should an Eastern Cape firm use AI safely?

Neither banning AI nor trusting it blindly is the answer. The competitive edge now comes from disciplined, verified use, and that starts with a written AI-use policy so every person in the firm knows the rules before they open a chatbot. With hallucination running at hundreds of fresh court cases a quarter (Scientific American, 2026), a clear policy is cheap insurance.

A workable policy for a professional firm usually rests on a few hard rules:

  • Treat every AI output as an unverified first draft. Check each citation, quote and figure against the primary source before it goes anywhere.
  • Never paste identifiable client data, privileged facts or trust-account details into a public chatbot. Strip or anonymise before you type.
  • Prefer enterprise or closed tools that offer a data-processing agreement and keep your data out of public training sets, rather than free consumer accounts.
  • Get informed client consent before any personal information is processed through an AI system.
  • Keep a human accountable for every deliverable. The signature on the work is yours, not the model’s.

Setting that up well is mostly practical: choosing tools with proper data guarantees, configuring them so client data stays put, and training the people who will actually use them. That is the work behind our AI enablement services and our IT for legal practices, both built around confidentiality and POPIA-aligned handling rather than novelty for its own sake.

Frequently asked questions

Can I use ChatGPT for legal research in South Africa?

You can use it to draft, summarise and brainstorm, but not as a source of authority. It regularly fabricates cases: in Mavundla v MEC KZN, seven of nine cited authorities did not exist (The Conversation, 2025). Verify every citation on SAFLII, and never enter confidential client data into the public version.

Will a court accept that the AI made the mistake?

No. The duty to verify is non-delegable, and "the AI said so" is not mitigation. In Northbound Processing, an apology did not prevent the court referring counsel to the Legal Practice Council for filing fictitious citations (ITWeb, 2025). Accountability stays with the practitioner (De Rebus, 2025).

Is it safe to summarise a client contract with an AI tool?

Only if the client data is protected. Typing identifiable details into a public chatbot sends them offshore, a cross-border transfer that can carry POPIA penalties of up to R10 million (Michalsons, 2025). Use an enterprise tool with a data-processing agreement, anonymise where you can, and get informed consent first.

Does POPIA ban law firms from using AI?

No. POPIA does not prohibit AI. It requires that you protect personal information, obtain informed consent, and stay accountable for how the tool handles data (De Rebus, 2025). The practical answer is closed or enterprise tools with data guarantees, plus a firm rule against pasting client data into public models.

The bottom line for Eastern Cape practices

AI is a real productivity tool for first-draft research, document summaries and surfacing leads, and the firms that use it well will move faster than those that refuse to touch it. But it is an assistant, not an authority, and South African courts have shown zero tolerance for firms that forget the difference. Verify every citation, keep confidential data out of public tools, and stay the accountable human on every file.

If you want help drawing up a sensible AI-use policy and choosing tools that keep client data where it belongs, book a consultation or message us on WhatsApp. You’ll speak to the analyst who’d handle your account, not a script.

Let’s talk about your IT needs.

Connect with a principal architect to discuss your firm’s technology, security, and compliance — in the context of the Eastern Cape’s regulatory landscape.

No obligation. A response from a principal analyst, usually within one business day.