24 Mar 2026

by Linda Sheehan and Shaheen Solwa

On 18 March 2026, Oasis Security published ‘Claudy Day’. A vulnerability chain in Anthropic’s Claude AI. Unlike attacks on proprietary in-house tools requiring custom exploits, Claudy Day used only Claude’s built-in capabilities to silently exfiltrate user conversation history. They noted basic sessions were vulnerable and enterprise configurations faced higher risk.

In another recent incident, Codewall, an artificial intelligence (“AI“) security startup that deploys autonomous AI agents to stress-test corporate infrastructure, reportedly breached a global consulting firm’s internal AI platform in under two hours.

In both cases, the prompt injection vulnerabilities were responsibly disclosed and patched before public announcement. These incidents emphasise a sobering truth: even sophisticated organisations with substantial security investment remain vulnerable to AI-expedited attacks.

AI is now embedded in legal operations, compliance and investigations. Large Language Models (“LLMs”) process high volumes of sensitive content, including emails, documents and legal summaries. Their strengths also create vulnerabilities. Security researchers consistently rank prompt injection among the top ten risks for modern AI deployments, with the Open Worldwide Application Security Project (“OWASP”) listing it as the number one threat in its Top 10 for LLM Applications.

What is prompt injection?

Prompt injection is a social-engineering attack on an LLM in which an attacker causes the model to ignore its original instructions and execute malicious commands instead. The goal is often to extract private enterprise data or manipulate outputs. Common forms include direct injection (overriding live system prompts), indirect injection (embedding malicious instructions in training data), and stored injection (planting persistent instructions in application memory).

Why this matters for legal operations, compliance and investigation teams

When using AI, legal teams must manage risks such as hallucinations, bias, prompt manipulation and model-driven data exfiltration. In South Africa, AI-related breaches may trigger obligations under the Protection of Personal Information Act (“POPIA”) and could constitute offences under the Cybercrimes Act. Courts, regulators and other third parties increasingly address the challenges of AI-assisted workflows handling sensitive and privileged material in day-to-day legal work.

Prompt injection poses a concrete threat to the sensitive material that legal teams handle daily. The Codewall incident underscores the following key points:

• Escalating risk of deploying AI without securing critical endpoints: Codewall’s AI security research revealed basic security vulnerabilities that had persisted for over two years. Codewall’s offensive AI agent autonomously selected the consultancy firm as a target and exploited Application Programming Interface (“API”) vulnerabilities and Structured Query Language (“SQL”) injection flaws.

• Access to intellectual property, personal data and more: The breach allegedly exposed 46.5 million internal chat messages, more than 728,000 files (including PDFs, Excel and PowerPoint documents), 57,000 user accounts and 95 system prompts, all of which were read and writable.

• System prompts as “Crown Jewel” assets: Codewall’s agent gained access to the AI tool’s system prompts – the instructions that defined how it answered questions, what guardrails it followed, how it cited sources, and what it refused to do.

A successful prompt injection can directly undermine the security, accuracy and defensibility of in-house legal workflows. The consequences extend beyond data breach triggers: a tainted review could compromise litigation, trigger regulatory sanction or waive privilege protection.

Once a poisoned prompt is injected, it could instruct the model to:

• Misclassify documents by excluding key evidence, fabricating timelines, or ignoring bad actor behaviour, resulting in flawed case assessments and missed liability exposures.

• Taint privilege reviews by failing to identify legally privileged communications, inadvertently disclosing attorney-client advice to opposing counsel or regulators and potentially waiving privilege over entire subject matters.

• Extract confidential information from integrated systems (client data, deal terms, settlement figures, litigation positions) for exfiltration to third parties or use in insider trading and competitive intelligence.

• Manipulate communications sent externally or silently delete critical files and audit trails, potentially constituting obstruction of justice or professional conduct violations.

Real-world vulnerability

Early LLM users and cybersecurity researchers have attempted to manipulate ChatGPT in numerous ways to leak confidential and proprietary information. In one widely reported incident, a user convinced ChatGPT to reveal Windows activation keys through “emotional” manipulation, claiming his grandmother used to recite them as lullabies. While the example may seem whimsical, it illustrates a serious vulnerability: LLMs can be socially engineered to bypass their safeguards.

In December 2025, OpenAI acknowledged that prompt injection attacks against its ChatGPT Atlas (search feature) browser may never be fully solved. This stark acknowledgment from one of the world’s leading AI developers confirms that prompt injection must be treated as a persistent operational risk requiring layered defences, not a problem with a single technical fix.

Practical guidance for legal teams

The speed of AI adoption in legal work creates a paradox: the same integrations that boost productivity also expand the attack surface. As AI tools gain access to more data sources (emails, documents, matter management systems, regulatory databases) each new connection becomes a potential entry point for malicious instructions. The solution is not to abandon AI, but to deploy it with appropriate safeguards.

Below are practical steps that legal teams can take to protect their workflows.
When selecting or deploying AI-enabled legal technology, ensure the platform incorporates robust defensive features. Key technical safeguards to look for include:

• Input Validation – The use of prompt templates, delimiters and sanitation allow for input control, separation of user input from system instructions, isolation of data and stripping input data to remove executable code and hidden text.

• Advanced Filtering – Technologies and techniques such as prompt shields (real-time input detection of hidden instructions), spotlighting (distinguishing between trusted and untrusted instructions), and context-aware filters allow for further protection against prompt injections and execution of malicious instructions.

Beyond vendor selection, legal teams must establish operational protocols that reduce exposure to prompt injection risks.

Operational safeguards for legal workflows

• Privilege-safe configurations are foundational. When deploying AI tools, apply role-based access controls and data minimisation to carefully define which data sets the models can access. Scope access to what is strictly necessary for the matter at hand, limiting damage if an injection succeeds. Maintain immutable logs of prompts, tool calls and outputs to support audit, anomaly detection and incident response. These logs may prove critical when a court or regulator requires demonstrable defensibility.

• People remain the critical line of defence. Train staff to verify sources, treat unfamiliar files with caution, and simulate attacks to surface weaknesses proactively. Everyone using AI tools should know how to recognise suspicious model behaviour, when to escalate, and how to preserve relevant logs for investigation. Regular tabletop exercises help ensure these protocols become second nature.

• Constrain what AI tools can do by establishing clear usage policies. Define permissible actions for each workflow and prohibit blanket capabilities such as unrestricted file exports or bulk email forwarding. Ensure that users understand these boundaries. A single malicious instruction could escalate into a catastrophic, machine-executed outcome that unfolds before anyone notices.

Technical controls and compliance frameworks
Given the stochastic nature of generative AI, no single control can fully prevent prompt injection. OWASP and other authorities recommend a layered, defence in depth architecture combining input/output controls, access restrictions, monitoring, and system hardening:

• Input and output controls. Validation and sanitisation to filter adversarial patterns, structured prompts separating system instructions from user data, and output monitoring to detect prompt injection artefacts such as system-prompt leakage or policy violations.

• Access controls and human oversight. Least privilege access ensuring LLM applications operate with minimum permissions, human-in-the-loop controls for high-risk actions such as executing code or accessing sensitive datasets, and segregation of external content so RAG pipelines cannot override trusted instructions.

• Monitoring and testing (before Codewall does it for you). Adversarial testing and red team exercises against known attack patterns, behavioural and anomaly monitoring to detect unusual instruction structures or API call patterns that standard filters may miss.

• System hardening. Containment focused design with sandboxed execution environments and validated outputs, system prompt protection using read-only databases, comprehensive audit logging, and regular security posture evaluation prior to AI deployment.
Compliance frameworks such as NIST AI RMF 1.0, ISO/IEC 42001:2023 and GDPR Article 32 increasingly require organisations to address prompt injection risks through formal policies, risk assessments and appropriate technical safeguards. In South Africa, POPIA section 19 requires “appropriate, reasonable technical and organisational measures” for AI systems processing personal information, whilst the

Cybercrimes Act criminalises unlawful data interception (section 5) and cyber extortion (section 10).
Legal teams should work with information security and compliance functions to ensure these requirements are reflected in vendor contracts, internal policies and audit programmes.

Ethical deployment and governance considerations

Organisations should be transparent about AI limitations, ensure human oversight of sensitive workflows and conduct proportionality assessments. The responsible use of AI in legal contexts demands not only technical safeguards but also governance frameworks that preserve professional accountability and employee and public trust. Board-level accountability for IT governance remains critical. Deploying AI without appropriate oversight creates legal, reputational, commercial and financial risk.

Conclusion: Security maturity must keep pace with AI adoption
AI adoption in legal operations should not be halted. The efficiency gains are significant and the competitive pressures real. But security maturity must accelerate to match the pace of deployment.

With disciplined architecture, least-privilege access, strict behavioural boundaries and vigilant monitoring, organisations can harness the power of LLMs while keeping attackers at bay.

How ENS intelligENS Can Help
ENS intelligENS is our specialist legal technology and AI division, comprising lawyers, data scientists and technologists who combine legal expertise with technical fluency. Our team brings a deep understanding of both AI systems and the legal frameworks governing their use. We advise on AI governance and assist clients in incorporating AI-assisted workflows into legal disputes, regulatory responses, and investigative workflows whilst preserving privilege. Reach out to ENS intelligENS to learn how we can support your organisation in navigating the evolving intersection of AI and legal practice.

Linda Sheehan
Executive and Head of intelligENS
lsheehan@ensafrica.com

Shaheen Solwa
Analyst | intelligENS
ssolwa@ensafrica.com

The post Prompt injection: Are legal-tech investigations safe during the AI boom? appeared first on intelligENS.

17 Mar 2026

by Linda Sheehan and Ettiene Beneke

Deepfakes are rapidly eroding the long‑held assumption that visual and audio evidence reflects reality. The sophistication of synthetic media now forces South Africa’s legal community to rethink how digital evidence is authenticated and how investigative processes must adapt.

Locally, reports already show rising AI‑driven identity fraud, biometric spoofing, voice cloning and impersonation scams. These trends reflect a global shift in which synthetic media is increasingly weaponised in both criminal and civil disputes. Internationally, courts are confronting the “liar’s dividend” – when individuals strategically claim genuine evidence is fake – signalling a future in which courts must evaluate both fabricated content and fabricated doubt.

This article explores:

1. How South African evidentiary law is adapting;
2. What the EU’s AI Act and its draft Deepfake Code of Practice mean for local practitioners; and
3. Practical, risk-based guidance for handling AI-generated content in litigation and investigations.

Authenticating evidence in the AI-era

South Africa’s evidentiary framework – the Consumer Protection Act (“CPA”), Cybercrimes Act, Electronic Communications and Transactions Act (“ECTA”), Protection of Personal Information Act (“POPIA”) and the common law – was designed for a world where digital records were assumed to be reliable  reflections of the real-world events. Deepfakes dismantle that assumption. They destabilise the notion of a trustworthy “original” and heighten the requirements for proving that a digital record is complete, unaltered and accurate.

The second draft of the European Commission’s voluntary deepfake code of practice published on 06 March 2026 sets out targeted obligations for when AI-generated content, such as:

• labelling standards, including use of an “AI” acronym, supported where appropriate by concise text such as “Generated with AI” or “Manipulated with AI”
• audio only content to include a brief spoken disclaimer
• notice appearing at the first point of interaction and expressed unambiguously, as required by Article 50(4) of the EU AI Act

1.1 Authenticity can no longer be sssumed

Courts can no longer accept authenticity based solely on appearance, a party’s assurance or supporting witness testimony. The key question now is whether the content is genuine at all. This requires more rigorous scrutiny, often including technical or forensic verification.

Courts will still prefer the most reliable – and ideally original – version of any digital file. This aligns with ECTA’s requirement that information must have “remained complete and unaltered,” and that the systems generating or storing it must be reliable.

Where AI reconstructs pixels, smooths audio, predicts frames or enhances quality, the result may no longer constitute an “unaltered” data message. These AI‑modified outputs should be treated as derivative evidence, with the untouched original remaining the primary source for integrity assessment.

1.2 The dual burden on practitioners

Deepfake‑era evidence imposes a dual burden:

• Authenticate the original digital artefact, and
• Validate the integrity of any AI-assisted analysis or enhancement performed on it.

This requires meticulous documentation of tools used, prompts or parameters applied, the purpose of each processing step and the preservation of both original artefacts and all AI-generated derivatives.

1.3 Integrity requirements Under ECTA

South African law already recognises that the integrity and evidential weight of a data message depend on more than the content itself. Courts must evaluate the reliability of the processes and systems through which data is generated, stored, communicated and preserved. Deepfakes directly undermine these assumptions by introducing avenues for subtle, nearly undetectable manipulation. This will require enhanced integrity assessments beyond those contemplated when ECTA was enacted, especially as AI‑mediated evidence becomes more common.

1.4 Legal remedies and their limitations

South Africa’s legal framework provides avenues for addressing malicious deepfakes, including:

• Cybercrimes Act (electronic fraud, impersonation and identity‑theft‑style offences),
• Common‑law claims (fraud, personality‑rights violations), and
• POPIA‑based claims (misuse of personal information.)

However, enforcement remains difficult. Deepfake creators often operate  anonymously, from foreign jurisdictions and with easily accessible tools that enable mass-production of synthetic media.

1.5 POPIA Section 71: Human oversight is mandatory

POPIA Section 71 prohibits decisions with legal or similarly significant consequences from being based solely on automated processing. This safeguard is critical in forensic and investigative contexts.

AI tools that identify individuals in suspected deepfakes, match facial features or classify subjects as high‑risk can influence prosecutorial, disciplinary, regulatory or employment outcomes. These determinations easily fall within Section 71’s “significant effect” threshold.

Human oversight must therefore be genuine and substantial – not a rubberstamp. Decision‑makers must interrogate the context, reliability and limitations of AI outputs.

This reinforces the importance of documented, independent human review throughout investigative workflows. AI may assist, but it cannot autonomously determine outcomes that affect legal rights, reputation or liberty.

1.6 The new evidentiary baseline

These legal frameworks collectively raise the authentication bar. Practitioners must now combine traditional evidence with:

• metadata analysis
• device forensics,
• cryptographic hashing,
• AI‑tool identification, and
• detailed logging of all analytical procedures.

Meeting this standard often requires specialised digital‑forensic expertise, adding cost and complexity – especially for smaller firms or legally aided matters. But these enhanced measures are becoming increasingly necessary environment where the line between genuine and fabricated digital evidence is blurred.

2. Lessons from courts abroad

Deepfakes have begun reshaping evidentiary disputes worldwide, signalling lessons and warnings for South Africa’s legal system:

2.1 Lack of awareness of deepfake risks: In a UK family law matter a “heavily doctored” audio recording portraying a father as violent was successfully challenged through digital forensic experts. The father’s attorney warned that it may not occur to most judges that deepfake material could be submitted as evidence.

2.2 Too much awareness: The “liars dividend” is a phenomenon where bad actors exploit deepfake awareness to cast doubt on genuine evidence. Tesla’s argument that it could not confirm the authenticity of video clips of Elon Musk because public figures are frequent deepfake targets was denied by the California Superior Court over concerns it would set a precedent that public figures could use to evade accountability.

2.3 AI-enhanced evidence rejected: A Washington court excluded cellphone footage enhanced using Topaz Video Enhance AI. The model’s opague, predictive methods introduced new pixels based on what the AI “thought” should appear. The court found this risked misleading the jury, and the evidence failed the Frye standard for scientific acceptance.

2.4 AI-generated content can mislead courts and harm society: The Minnesota’s Judicial Branch AI Response Committee noted judicial concern about making decisions based on AI-generated material. While early deepfakes displayed clear signs – monotone voices, repetitive expressions – these “tells” are rapidly diminishing.

2.5 Verification is resource-intensive and affects access to justice: In the US, a deepfake audio clip causing public unrest required intensive forensic investigation, including FBI involvement, to both confirm it was AI-generated and trace it back to the bad actor.

3. Practical guidance for deepfakes in litigation, investigations and regulatory matters

3.1 Ensuring evidence is accurate and reliable

Deepfakes demand a higher evidentiary standard. Investigators should work from forensic copies, record cryptographic hashes and maintain comprehensive logs of tools, prompts and settings. AI‑generated outputs must be treated as investigative leads, not conclusions – requiring corroboration, checking alternative explanations, and verifying claims from large language models. Clear, non‑technical explanation of methods is essential as courts increasingly scrutinise AI‑supported evidence.

3.2 Using generative AI to support evidence gathering

Generative AI can summarise material, flag anomalies and organise large datasets, but only with human oversight and full auditability.

Practitioners must:

• maintain transparent chain of custody,
• disclose when AI has generated or altered content, and
• distinguish clearly between machine-suggested insights and professional judgement.

Compliance with POPIA Section 71 requires decisions influenced by automated tools to be explainable and open to challenge.

3.3 Emerging judicial responses and the path forward

Internationally, courts are responding quickly. The US is considering burden‑shifting rules for suspected AI‑fabricated evidence; Louisiana now requires lawyers to verify authenticity and disclose AI‑generated material; and judges in England and Wales emphasise vigilance around synthetic media.

For South Africa, the imperative is clear – adopt enhanced verification protocols early, build judicial awareness and prepare for scenarios where litigants allege deepfakery even when evidence is genuine.

Conclusion

Deepfakes have reshaped the authentication burden. Success now depends on rigorous documentation, transparent and explainable AI use and sustained human oversight. Practitioners who modernise their evidentiary processes early will be best positioned to maintain credibility and meet the rising expectations of courts, clients and the public.

Linda Sheehan
Head of IntelligENS
lsheehan@ENSafrica.com

Ettiene Beneke
Practice Lead
ebeneke@ensafrica.com

The post Truth on Trial: Deepfakes and the new battleground for evidence integrity appeared first on intelligENS.

23 Feb 2026

by Megan Claassens

Lessons from a case where a defendant’s claim of legal privilege over AI documents failed.

In United States v. Heppner, Judge Jed S. Rakoff of the U.S. District Court for the Southern District of New York ruled that documents created by a defendant using a third-party AI tool are not protected by attorney-client privilege or the work-product doctrine, even when later shared with lawyers.

The case involved Bradley Heppner, who faced securities and wire fraud charges. Before his arrest, Heppner used Anthropic’s AI assistant Claude to generate approximately 31 documents analysing the government’s investigation. He then emailed these documents to his legal team. When federal agents later seized the materials, defence counsel argued they were privileged.

Judge Rakoff disagreed. Applying established legal privilege principles to an AI context, he concluded that:

• AI tools are not lawyers, so communications with them do not qualify as privileged;
• Heppner had no reasonable expectation of confidentiality given the AI tool’s data-collection practices and privacy terms;
• Sharing AI-generated content with a lawyer after the fact does not retroactively create privilege; and
• The work-product doctrine did not apply because the materials were not created at the direction of counsel.

The ruling serves as a clear warning: using consumer AI tools independently for legal purposes may expose sensitive materials to discovery.

How to protect legal privilege when using AI

Although this decision specifically addresses a defendant’s use of a publicly available AI platform in connection with a pending criminal investigation and is silent on the position of the use of enterprise AI platforms, it represents one of the first rulings on this issue and offers valuable guidance. Individuals and organisations using AI in legal contexts should consider the following precautions at a minimum:

• Avoid using public AI tools for confidential legal matters: Consumer-grade AI platforms with broad data-use terms pose significant risks.
• Use enterprise-grade or privacy-protected AI solutions: Use platforms with explicit contractual guarantees of confidentiality and restrictions on data use.
• Ensure counsel directs any AI use: Where AI is used in connection with privileged legal work, ensure that a suitably qualified lawyer provides explicit guidance on the appropriate use of the tool, that such use is conducted under the lawyer’s instructions, and that this direction is documented contemporaneously.
• Clearly mark privileged AI-assisted communications: When appropriate, note in prompts and logs when AI outputs or prompts are created at the instruction of counsel for legal strategy.
• Review privacy policies in advance: Understand and document the privacy and data-handling policies of any AI service before entering case-related information.

Why understanding the law and AI matters

As the Heppner ruling shows, the way AI is used and who supervises that use, has significant consequences in litigation and privilege assessments. As an emerging field, specialist expertise are required in order to:

• Select the right AI tools by distinguishing between consumer AI and enterprise platforms offering confidentiality protections.
• Integrate AI safely within workflows that align with legal privilege requirements and avoid inadvertent waiver.
• Document matters properly, ensuring records reflect lawyer direction and purpose where privilege is at stake.
• Anticipate discovery risks and develop strategies that minimize the risk that AI-generated materials will be used against a client.

With AI rapidly becoming part of legal workstreams all over the world, including the African continent, the Heppner ruling is a timely reminder on how traditional privilege doctrines intersect with modern technology. Thoughtful, well supervised use of AI can help preserve confidentiality in a world where digital tools are increasingly part of professional practice.

How ENS intelligENS Can Help

ENS intelligENS is our specialist legal technology and AI division, comprising lawyers, data scientists and technologists who combine legal expertise with technical fluency. Our team brings a deep understanding of both AI systems and the legal frameworks governing their use, including in complex matters such as criminal investigations, securities litigation, fraud proceedings, and regulatory inquiries. The very contexts at issue in the Heppner ruling. We advise on AI governance and assist clients in incorporating AI-generated work into legal disputes, regulatory responses, and investigative workflows whilst preserving privilege. Reach out to ENS intelligENS to learn how we can support your organisation in navigating the evolving intersection of artificial intelligence and legal practice.

Megan Claassens
Practice Lead | IntelligENS
mclaassens@ensafrica.com

The post AI as your legal sidekick? Court says think again appeared first on intelligENS.

3 Feb 2025

by Linda Sheehan, Shaheen Solwa and Megan Claassens

Digital transformation is creating exponential value in the mining industry.

From Artificial Intelligence (“AI”)‑enabled exploration and digital twins in processing plants to autonomous fleets and predictive maintenance, mining companies can run increasingly complex, connected systems at scale. The results are visible: safer operations, higher recoveries and improved capital discipline.

The risk is equally clear. A cyber incident can travel at machine speed across interconnected infrastructure, with consequences for safety, production and enterprise value. This elevates the importance of robust governance.

AI is reshaping the threat landscape

IBM estimates the global average cost of a cyber breach to approximately USD4.4 million in 2025.

AI has transformed both offence and defense in cybersecurity.

The early rush to deploy ChatGPT resulted in speed-to-market outpacing governance. This resulted in sensitive IP being leaked into the model and hallucinated cases being relied upon in court by humans.

Agentic AI does not just generates outputs, it acts. AI-agent social platforms show how agents can post and interact with each other and coordinate across systems. This has already resulted in a reported exposure of 1.5 million API keys and tens of thousands of emails. This validates concerns about the risks associated with vibe-coding (AI-assisted software development) and rapid AI adoption without proper governance.

Agents embedded in everyday workflows can significantly improve efficiency. Yet, the same autonomy introduces new risks. If a cybercriminal interferes with the information these agents receive, or with the systems and data they can access, an agent could rapidly carry out harmful actions faster than a human can, Agents can faithfully follow malicious instructions hidden in content and then cover up its tracks. Agents can expand a once-off breach into a coordinated multi-breach attack, while adding complexity to post-incident forensics.

Even a traditional business email compromise attaching a “poisoned” document can lead to data leakage or malicious actions. These attacks can drive poor decisions in models intended to manage real-world assets, introducing safety and operational risks. For example, misrouting autonomous haul trucks or degrading safety interlocks.

As critical minerals grow in strategic importance to the global energy transition, cyber-attacks are increasingly linked to state-backed groups. These actors plan carefully and operate over longer periods, ranging from espionage to funding illicit activities, adding to an already busy cyber-crime environment.

Supply chains threaten cyber-resilient organisations

Modern mining operations rely heavily on complex supply chains, including equipment manufacturers, software vendors and integration partners. Attackers increasingly  exploit this dependency by targeting the weakest security controls in the chain.

Mergers and acquisitions pose particular risk due to shared systems, temporary access arrangements and inconsistent security practices. Cyber risk should therefore be treated like environmental or financial due diligence that is measured, factored into deal value and addressed before a transaction closes. This requires clear contractual rights around monitoring, cooperation and incident notification that are tested in advance rather than during a crisis.

Effective outcomes depend on aligning technological ambition with operational reality and regulatory expectations through collaboration between operators, original equipment manufacturers (“OEM”s), data providers, law‑enforcement and regulators.

Models and AI agents should be treated like assets: inventoried, version-controlled, access-restricted and subject to the same change controls applied to critical operationa technology (“OT”) assets. They influence operational decisions and risk.

Regulatory realities require evidential readiness

Cross‑border mining operations straddle multiple privacy, cyber security and sector-specific laws. Disclosure triggers vary and timelines are tight. After a cyber  incident, defensibility (what can be proven quickly to regulators, courts, counterparties and communities) matters as much as technical recovery.

In a double‑extortion ransomware incident, the ability to credibly establish what was accessed, what was exfiltrated and whose data is implicated will shape negotiations, regulatory posture and litigation exposure.

That makes logs, not narratives, the strategic asset.

Build forensic and litigation readiness into architecture, including immutable logging, role‑based access for humans and agents, data minimisation and anonymisation. Enterprise data mapping should include evidence of what is held, where it resides, who has access and the lawful basis for processing.

Knowing what was taken – fast: rapid content intelligence

When extortion groups dump or threaten to publish terabytes of mixed data, the decisive question is: what, precisely, was taken and what is the exposure risk?

A privileged, forensically sound methodology can turn messy breached data sets into classified content by business harm and legal sensitivity, for instance, personnel information, customer and supplier information, negotiation materials, board papers, intellectual property such as block models, drill logs and processing parameters and safety or OT‑adjacent documentation.

Critically, the process should move beyond labelling files to quantifying and mapping exposure so executives can act quickly on reputational and legal risk.

Organisations that maintain current data maps and retention schedules before an incident can answer “what was taken” and “whose data” materially faster. Counsel can move from raw data to a prioritised risk picture quickly enough to brief boards with specificity and sequence external communications before the narrative hardens.

This enables legal, cyber, operations and communications teams to work from the same, quantified picture of harm.

Incident response that reflects operational reality

Incident response plans must be AI‑aware and OT‑literate. A staged, achievable path forward for operators:

• Mandate and rehearse AI‑aware incident response that spans IT and OT, with clear decision rights on pausing automation and validating data integrity.
• Harden identity and session management across cloud and vendor ecosystems; regularly review controls during M&A and major projects.
• Uplift board reporting to include AI agent governance and site‑relevant resilience metrics.
• Invest in forensic readiness and surge review capability that can produce regulator‑ready outputs at speed without sacrificing quality.
• Train high‑risk functions against deepfakes and AI‑enabled social engineering using realistic simulations.

Boards need site‑relevant oversight and AI governance mechanisms

Boards need site-relevant oversight and information and AI governance mechanisms.. Directors should understand how technologies are governed, which actions require human approval, how quickly AI agents can be isolated, the scope of red‑teaming and safety testing, and whether evidential logging captures AI tool use across IT and OT environments.

Cyber readiness has shifted from a discretionary IT function to a mandatory, board-level, and legally auditable requirement. Auditable decision‑making pre and post-incident are critical factors in mitigating regulatory and reputational damage.

“Together we are better”: collaboration that changes outcomes

When operators prioritise security, regulators articulate consistent expectations around notification and evidence and when law‑enforcement engages early in high‑impact incidents, outcomes improve.

Inside the enterprise, legal, cyber, operations and communications teams must train together. Externally, industry platforms can accelerate uplift by promoting minimum standards for agent governance, logging and incident cooperation.

Applied carefully under human supervision, AI can also accelerate the post-incident plan, closing the gap between detection and defensible response.

Resilience is becoming a competitive capability in mining operations. Companies that can demonstrate how AI agents act, how models are governed, and how information governance and post-response meet cross‑border standards will modernise with confidence while protecting people, production and reputation.

Cybersecurity, done well, is not a brake on innovation. It is the discipline that lets you adopt AI at speed without losing control of your risk, your data or your licence to operate.

Mining plays a critical role in Africa’s economic future, but progress depends on the right legal, regulatory and partnership frameworks. We curated a compilation of insight and perspectives on the trends, challenges and opportunities shaping Africa’s mining sector. Click here to explore the page: https://bit.ly/4sZov1T

Linda Sheehan
Head of intelligENS
lsheehan@ENSafrica.com 

Megan Claassens
Practice Lead | intelligENS
mclaassens@ensafrica.com

Shaheen Solwa
Analyst | intelligENS
ssolwa@ensafrica.com

The post AI, cybersecurity and operational resilience in mining appeared first on intelligENS.

4 Dec 2025

by Linda Sheehan, Ettiene Beneke and Megan Claassen

Accelerating legal disputes with GenAI-enhanced data review

As eDiscovery day closely coincides with ChatGPT’s third birthday, we reflect on how generative AI (“genAI”) has gone from a novelty to an everyday tool for legal teams. It sits alongside proven machine learning in standard eDiscovery work – from early case assessment and scoping, to issue analysis, privilege review, disclosure and trial preparation. Courts and regulators expect technology-assisted review to be used responsibly and to be auditable. The question for in-house legal teams and law firms is not whether to use AI, but how to use it in a way that is defensible, cost-effective and clearly improves legal outcomes.

Through this article, we discuss how GenAI and intelligENS, the specialist legal technology division of ENS, can help legal teams get to the facts faster, protect privilege and raise quality – without risking defensibility. We ensure that internal and external standards on AI governance is met, both locally and aboard. We help inhouse legal teams and law firms augment their internal capabilities with specialised AI, data and investigation support, helping them reach the facts sooner while maintaining full legal integrity.

What GenAI does in in legal disputes data review

Modern eDiscovery combines proven machine learning with genAI to help lawyers surface key evidence from the outset, whether the data set contains hundreds or millions of documents. Based on known legal issues and human training, active learning quickly separates likely relevant from irrelevant documents. Guided by a matter brief and clear prompts, genAI summarises the relevant documents, groups them by issues in dispute and flags key (helpful or harmful) evidence immediately. The result is a strategic head start. In hours and days, not weeks or months, legal teams can see the themes, links and the factual shape of the case, so pleadings, settlement negotiations and investigation reports are grounded in evidence.

Active learning tools provide explainable, consistent and repeatable classifications and genAI explains the relevance of every document in plain language and produces a concise report of the key passages. Senior lawyers gain early access to the key documents they need to shape strategy and focus on the most material points without wading through pages of irrelevancy. In parallel, genAI enables the discovery review to produce more consistent relevance and privilege decision and logs, faster.

Responsible AI is an operating model, not a button

“Trust but verify” should be the rule. Responsible AI governs the ways of working across people, process and technology. It starts with sensible guardrails before anything is deployed This may include agreeing what acceptable accuracy looks like with internal stakeholders, the opposing parties and the court. It relies on transparent workflows, ongoing monitoring and iteration and human oversight that focuses on legal judgement.

Key disciplines include: selecting the right dataset; strict access controls aligned to local privacy rules, prompts designed carefully around the agreed legal issues and continuous testing with documented settings and checks. GenAI’s reasoning for decisions should be visible through short explanations rationalising its decision and in-text links back to the source document or piece of evidence. If the genAI’s explanations or in-text links cannot be found in the document, or it is simply echoing the prompt rather than the evidence, the methodology should automatically flag it for human review. Verification, sampling and targeted quality checks keep performance of the genAI within agreed limits and human oversight is essential. Conversely, the genAI output can be used to quality check the human review decisions too.

Built‑in privacy, security, data minimisation and defensibility

When we use AI in eDiscovery, we are handling clients’ most sensitive information. The strict ground rules: keep data on secure, jurisdiction‑appropriate platforms, do not allow it to train external models and analyse only what is necessary for the specific legal purpose – nothing more.

Choose tools and providers that let you prove where data lives, who accessed it and when. Maintain clear records, including DPIAs, transfer risk assessments and data‑flow maps where applicable. Build in purpose limitation, segregation of duties and meaningful human oversight. Privilege, confidentiality and proportionality duties cannot be delegated to a model.

Defensibility is about disciplined process as much as technology. Your protocols should explain the approach in plain language, including key settings and validation checks, so every decision is traceable back to the documents. Preserve chain of custody. Record prompts and model versions. Keep a live register of privacy and procedural obligations. Done well, AI improves quality and lowers total cost by focusing review effort where it matters and cutting wasted reviewer hours, allowing budget to shift to strategy and negotiation.

How in-house legal teams and law firms work with intelligENS

intelligENS equips lawyers with leading technology-driven methodologies without the need to build an internal eDiscovery function. intelligENS designs, digitises and governs end-to-end methodologies that are auditable and privacy first, with our eDiscovery expertise and your legal expertise in the loop at the right moments. We set up a matter‑specific workflow under privilege, validate it with you, and then scale. Results come from combining our globally leading technology with platform fluency, forensic discipline and legal judgement. Our team has hands‑on experience of running genAI reviews successfully. We are experts in designing and translating genAI review methodologies into defensible legal outcomes and recognised at the forefront of global innovation in this space.

What in‑house legal teams and law firms gain with intelligENS

You gain earlier sight of key facts and themes, tighter privilege control, and a documented method that courts and regulators can follow. Reviews are faster and higher quality, delivering predictable cost control. Most importantly, you keep accountability and strategic control: we make the evidence available, defensible and ready, so your team can focus on winning the legal arguments.

Conclusion: relevant innovation, measured results

This is not experimentation for its own sake. It is innovation that accelerates time to facts while preserving accountability. Trust the technology – but verify it. Keep humans in the loop. Make every decision traceable back to the record. That is how you cut noise, surface the signal, and progress your matter with confidence.

Linda Sheehan
Head of intelligENS
lsheehan@ENSafrica.com

Ettiene Beneke
Practice Lead
ebeneke@ensafrica.com

Megan Claassen
Practice Lead
mclaassens@ensafrica.com

The post Accelerating legal disputes with GenAI-enhanced data review appeared first on intelligENS.

14 Oct 2025

by Linda Sheehan, Shaheen Solwa, Ettiene Beneke and Bridget Zitha

At the J20 summit, Justice Mlambo, Deputy Chief Justice posed a critical question: in a multilingual society, are litigants still excluded if they do not speak the colonial language of record? South Africa’s legal system must confront a dual reality. On the one hand, the Constitution guarantees parity of esteem and treatment for all twelve official languages. On the other, the courts have, since 2017, adopted English as the language of record for reasons of practicality and uniformity.

South Africa recognises 12 official languages, namely Sepedi, Sesotho, Setswana, siSwati, Tshivenda, Xitsonga, Afrikaans, South African Sign Language, English, isiNdebele, isiXhosa and isiZulu. While isiZulu and isiXhosa are the most widely spoken at home, English remains the pre-eminent language of business and legal record. In practice, this means accused persons and witnesses may testify in the language of their choice, but the official record is in English. The intent is administrative clarity; the effect is a translation burden and the risk of exclusion for those whose legal realities are expressed most precisely in another tongue.

This balancing act is not unique to South Africa. Jurisdictions worldwide grapple with ensuring access to justice for non-native speakers as they standardise court records for efficiency and interoperability. The immediate challenge is operational: producing reliable, timely translations without losing nuance. The strategic opportunity is technological: using responsible artificial intelligence (“AI“) to narrow, rather than widen the language gap.

Are indigenous languages linguistically excluded from AI transcription and translation?

Justice Mlambo’s intervention at the J20 Summit highlighted a pivotal question for access to justice: do AI tools enable courts to operate credibly across languages, particularly for indigenous language speakers?

According to UNESCO, the African continent is home to a third of the world’s languages with, “an estimated 1,500 to 3,000 languages, Africa is a true hub of linguistic and cultural wealth”.

The “Mind the (Language) Gap” white paper notes that most major large language models (“LLMs”) underperform for non-English, and especially low-resource languages, and are not attuned to relevant cultural contexts. Many African indigenous languages are considered low-resource because they lack sufficient digital data, annotated datasets and computational tools needed for effective AI-powered translation and speech recognition. This scarcity makes it difficult to train and deploy language technologies, resulting in digital underrepresentation for speakers of these languages. Conversely, the so-called “curse of multilinguality” means that as more languages are added to a model, performance for each may decline, especially for those with limited digital resources.

LLMs trained predominantly on English data are often biased in favour of native English speakers, meaning low-resource language communities are at risk of being left behind in the AI revolution.

Local is lekker

A growing body of international research highlights the transformative potential of AI in supporting and revitalising indigenous languages. As noted by Forus International in May 2023, “Chat GPT for example knows the capital of Kenya but when being asked in Kinyarwanda (one of the official languages of Rwanda), it does not understand the question”. Whilst AI models are continuously improving, Forus highlighted the importance of domain specific data sets and the challenge posed for those languages that are spoken but not written.

The NGLUEni project benchmarks and improves the performance of pretrained language models for South African Nguni languages, isiXhosa, isiZulu, isiNdebele and siSiswati, four of South Africa’s official languages. By creating a unified evaluation framework and adapting models specifically for those linguistically related, under-resourced languages, the study demonstrates significant gains in accuracy and cross-lingual transfer. This work supports greater linguistic inclusion in digital tools and AI systems.

African Innovation and Collaboration

Recent developments on the continent demonstrate African innovation and collaboration to address the digital language divide and overcome the challenges in linguistic underrepresentation in AI. As reported by iAfrica in September 2025, a “landmark AI dataset” has been created by African researchers to “close the language gap and boost digital inclusion.”

Linguists, computer scientists and AI experts across Kenya, Nigeria and South Africa, funded by a USD2.2 million Gates Foundation grant, created AI-ready open-access datasets representative of real African speech patterns. This critical resource will enable translation, transcription and conversational AI tools to better understand and process the way people live, speak and interact on the African continent.

The world’s largest language barrier lifted by AI

At the J20 summit, Luís Roberto Barroso, Justice of the Supreme Federal Court of Brazil, considered that language is one of the most important benefits that artificial intelligence will bring to the world.

This holds particular importance when handling cross-border legal matters. It is essential to incorporate a framework for the responsible and effective use and integration of AI and language technologies into everyday legal practice.

Crossing linguistic borders in litigations and investigations with AI

Language is no longer a hindrance in litigation, investigations, regulatory responses and any other document-heavy matters.

Generative AI tools can analyse millions of documents involving multiple languages in a matter of days. Specialised tools, such as Relativity’s aiR for Review, when prompted with case-specific criteria can summarise, classify and pull together the evidence, regardless of language.

This approach not only accelerates the identification of critical evidence but also provides transparent reasoning and direct links to supporting evidence, which is invaluable for validation of the generative AI review as well as organising the findings for cross-examination and report writing. Further, it removes the human bias inherent in the search phase of document reviews when, in multilingual matters, search criteria for certain languages is often curated by non-native speakers creating gaps in the identification of crucial information.

By combining the speed and scalability of AI with human oversight, legal teams can manage cross-border, multilingual matters more efficiently, reduce risk and ensure that critical evidence is not lost in translation.

Co-piloting in-house legal practice with AI translation and transcription

In-house legal teams across Africa routinely operate in English, French, Portuguese and a multitude of local languages. AI translation and transcription can accelerate contract review and drafting, localise compliance policies, enable multilingual knowledge bases and streamline communications across jurisdictions. Employees are more likely to internalise policies and training when they are conveyed in their first language, increasing the likelihood of operational compliance rather than “paper compliance”.

As organisations build translation memories and vetted glossaries, quality and speed improve over time, creating a resilient multilingual knowledge infrastructure. However, adoption requires robust governance. Legal exposure can be created through mistranslations or omissions in high-stakes documents, confidentiality risks from unvetted tools and inconsistent terminology resulting in misalignment across offices. The prudent approach is a structured, dual-pass process: machine translation for scale, followed by subject-matter expert review, supported by controls for privacy, privilege, confidentiality and data security.

Call to action: legal standards for linguistic inclusion

AI is reshaping legal work, from court transcription to document review. When designed for local contexts and deployed responsibly, AI-powered transcription and translation can turn linguistic barriers into vectors for inclusion and more equitable outcomes. Sustainable progress depends on legal professionals who are AI-literate, insist on equality of treatment across languages and support ongoing investment in local-language datasets, domain-specific fine-tuning and governance frameworks that protect confidentiality and accuracy.

Legal professionals should set the tone: embrace innovation, insist on equality.

Linda Sheehan
Head | intelligENS
lsheehan@ENSafrica.com

Ettiene Beneke
Practice Lead | intelligENS
ebeneke@ENSafrica.com

Shaheen Solwa
Analyst | intelligENS
ssolwa@ENSafrica.com

Bridget Zitha
Analyst | intelligENS
bzitha@ENSafrica.com

The post From exclusion to inclusion: AI, language and access to justice in South Africa appeared first on intelligENS.

06 Oct 2025

by Linda Sheehan

At the recent Constitutional Courts and Supreme Courts Summit (“J20 Summit”), the Chief Justice of South Africa, Madam Justice Mandisa Maya, together with heads of judiciaries, chief justices and court presidents, shared their insights on artificial intelligence (“AI”) in legal practice.

The J20 Summit, hosted in South Africa in September 2025, brought together judicial leaders from G20 nations and across Africa under the theme, “Justice in a time of change: Independence, innovation and cooperation”. Central discussions focused on the transformative role of AI and legal technology in strengthening judicial independence, improving access to justice and addressing cross-border challenges such as climate change and cybercrime.

This article explores the topic raised during the J20 summit that:

“AI, when designed for local contexts, can transform linguistic exclusion into linguistic inclusion, an advancement of enormous significance for justice in Africa”.

Is ChatGPT linguistically inclusive for legal matters?

When used responsibly, AI can make court proceedings more accessible. It can deliver near real-time speech-to-text transcription, translate documents into multiple languages and enable multilingual searches across court judgments, enabling litigants to read court documents in their home language and legal teams to find precedents faster.

However, as with the well-publicised issue of AI hallucinations in legal research, AI transcription and translation require a strict “human-in-the-loop” approach. Lawyers must remain alert to hallucinations, omissions and confidently presented inaccuracies.

To improve access to justice, it is essential to address limitations, risks and potential biases in AI outputs that can compromise fairness. Cultural and technical subtleties, including dialects, idiomatic usage and specialised legal terminology are often misunderstood by AI models (and by non-expert human translators). Inadequate or skewed training data can further disadvantage speakers of indigenous languages. Unregulated use may also lead to inadvertent disclosure of privileged or personal information, raising significant confidentiality and intellectual property concerns.

Large language mishaps

When asked which AI translation method is best, the lawyer’s answer applies: “It depends”.

Do you prioritise readability or reliability?

While large language models (“LLMs”) can accelerate translation, improve fluency, and cover broad topics, they are not a replacement for purpose-built machine translation (“MT”), particularly where scale, completeness, accuracy and consistency are required. A human-machine hybrid remains best practice.

• Generative AI (e.g., ChatGPT) vs Cognitive MT

It is critical to benchmark tools on the relevant languages and dialects before deployment. For example:

• Generative AI (LLM-based) excels at summarising and creating content. It predicts likely word sequences and tends to produce more readable translations. Whilst user-friendly, this can create a false sense of security in legal work.
• Cognitive/MT tools rely on predefined linguistic rules or statistical correlations from parallel corpora (large datasets of texts that are translations of each other, with the original source text and its translations aligned at the sentence or phrase level). They are often more predictable and transparent, and easier to audit and control for legal use.

According to this article, generative AI is better suited to creating and localising content for specific languages, whereas MT is better suited for specialised documents, (such as legal contracts) and instant translation (such as, contemporaneous interpretation of court proceedings).

• AI – ChatGPT vs Human

A Finish study, in the social and health care sector found that GPT-4 demonstrated promising productivity in translation, but confirmed the need for a dual-pass process; AI generates a draft and a human reviewer edits the output. In legal contexts, the reviewer should be a professional translator with subject-matter expertise, supplemented by back-translation (translating back into the source language to detect any changes in meaning).

Another study comparing human and ChatGPT-generated translations of Arabic legal documents, (including contracts) found human translators more accurate on legal terminology, complex constructs and cultural nuance. The study concluded that AI translations are not “inherently flawed” but take a different approach; it found that AI translations are not yet suitable for official use in legal contexts.

• Validating generative AI output

To mitigate risks in generative AI translation:

• Be aware of the “lost in the middle” effect (bias towards translating information at the beginning and end of documents) and context-window limits, which can cause omitted or skipped information in translations.
• Expect hallucinations, (plausible but incorrect names, citations, or legal terms, formatting errors, and inconsistent terminology).

Call to action: Setting legal standards for linguistic inclusion

AI is reshaping legal work, from court transcription to document review. AI-powered translation and transcription – often integrated as co-pilots – are now commonplace. As these tools become embedded in court systems and legal workflows, the profession stands at a critical juncture.

Designed for local contexts and deployed responsibly, AI can convert linguistic barriers into opportunities for inclusion, promoting more equitable legal outcomes.

To achieve sustainable linguistic inclusion, lawyers must develop and maintain AI literacy, backed by investment in local-language datasets, fine-tuned models for low resource languages and domain expertise.

Legal professionals should set the tone: embrace innovation, insist on equality.

Linda Sheehan
Head | IntelligENS
lsheehan@ENSafrica.com

The post AI in legal practice: Opportunities and risks for linguistic inclusion appeared first on intelligENS.

8 Sep 2025

by Era Gunning and Rakhee Dullabh

In the recent case of European Data Protection Supervisor v Single Resolution Board (“EDPS v SRB“), the Court of Justice of the European Union (“CJEU”) addressed the legal status of pseudonymised data and the obligations of data controllers when sharing such data with third parties. The case arose after the Single Resolution Board (“SRB”), an EU agency, shared pseudonymised comments from affected shareholders and creditors with Deloitte, an external consultant, during the resolution of Banco Popular Español. The European Data Protection Supervisor (“EDPS”) found that the SRB had failed to inform data subjects that their data would be shared with Deloitte, in breach of transparency obligations under Regulation (EU) 2018/1725. The General Court initially annulled the EDPS’s decision, holding that the assessment of whether data is personal should be made from the recipient’s perspective, if the recipient cannot reasonably re-identify individuals, the data may not be personal data in their hands. On appeal, the CJEU clarified that the context and risk of re-identification are crucial in determining whether pseudonymised data remains personal data, and that controllers must assess and document these risks when sharing data.

Although delivered within the EU legal framework, the CJEU’s reasoning in EDPS v SRB offers valuable guidance to South African organisations that qualify as “responsible parties” under the Protection of Personal Information Act 4 of 2013 (“POPIA”). In particular, the judgment illustrates how a context-specific, risk-based approach to pseudonymisation and data sharing can operate in practice. Against that backdrop, the following implications arise for POPIA-regulated entities:

Context-specific identifiability and pseudonymisation

POPIA’s exclusions clause removes “de-identified” information from the Act’s scope only where it cannot be re-identified by a reasonably foreseeable method. The EDPS v SRB judgment reinforces a core lesson: identifiability must be assessed from the perspective of the party in possession of the data at any given moment. A responsible party that retains the re-identification key will continue to process “personal information” and must therefore comply with all eight POPIA conditions. Conversely, once the data is disclosed – without the key – to a recipient that has no foreseeable means of re-identification, the dataset will generally be non-personal information in that recipient’s hands. Responsible parties should:

• maintain a robust internal pseudonymisation protocol (allocation of unique codes, secure segregation of the linking dataset, and strict access controls);
• document the technical and organisational measures demonstrating that a recipient cannot reasonably re-identify the data subjects; and
• periodically test whether new technologies or additional data sources have altered the re-identification risk.

Transparency duties remain with the responsible party
Even where onward disclosure renders the dataset non-identifiable for the recipient, the responsible party must still satisfy POPIA’s openness and notification requirements. Privacy notices should:

• name or describe all third-party recipients (including consultants and service providers) who will receive the pseudonymised dataset;
• explain that the recipient will receive only de-identified information and lacks any means of re‑identification; and
• set out the purposes of the disclosure in clear, specific terms, alongside the applicable legal basis.

Operator contracts and due diligence

When a responsible party appoints an “operator” (processor) to perform services on pseudonymised data, POPIA requires a written agreement mandating security safeguards equivalent to those imposed on the responsible party. Building on the EDPS v SRB rationale, the contract should also: prohibit the operator from attempting to obtain, or otherwise gaining access to, the re-identification key; include warranties that the operator lacks – and will not seek – additional datasets that could reasonably facilitate re-identification; and impose an obligation to notify the responsible party immediately if any circumstance arises that could change the identifiability risk profile.

Risk assessments and record-keeping

POPIA obliges responsible parties to implement “appropriate, reasonable” measures to identify foreseeable risks. Prior to any disclosure of pseudonymised data, entities should conduct and document a risk assessment addressing: the nature of the data elements remaining in the dataset; the likelihood that the recipient (or anyone to whom the recipient could grant access) could combine the dataset with other information to re-identify data subjects; and the legal, contractual, and technical barriers preventing such re-identification. These assessments should be retained as part of the responsible party’s processing records.

Data subject participation and exercise of rights

Because the responsible party retains the re-identification key, it alone remains capable of responding meaningfully to data subject requests (access, correction, deletion). Internal procedures must ensure that: pseudonymisation keys are retrievable by authorised personnel only when necessary to fulfil a data subject request; and any deletions or corrections made at the data subject’s direction are propagated to both the identifiable and pseudonymised versions of the dataset, or that appropriate technical measures are in place to prevent re-linkage.

Cross-border transfers

POPIA restricts transfers of personal information outside South Africa. Where only a pseudonymised dataset (without the key) is transferred, responsible parties may argue that POPIA not engaged because the recipient processes non-personal information. Nonetheless, prudent organisations should analyse and document why the dataset, in the recipient’s context, is no longer personal information; and if any residual risk of re-identification exists, comply with POPIA (among other things) by ensuring that the foreign recipient is subject to adequate protection, contractual safeguards, or binding corporate rules.

Era Gunning
Executive | Banking and Finance
egunning@ENSafrica.com

Rakhee Dullabh
Executive | Technology, Media and Telecommunicaitons
rdullabh@ENSafrica.com

The post Unlocking pseudonymisation: What EDPS v SRB means for South African responsible parties under POPIA appeared first on intelligENS.

21 Aug 2025

by Isaivan Naidoo and Alexander Powell

Artificial Intelligence (“AI”) is no longer a theoretical buzzword; it has rapidly morphed into an integral part of many organisations’ operations, production, and supply chains. While organisations have embraced AI tools for productivity and efficiency, the reality is that internal governance and legal frameworks have lagged behind. Many organisations do not have an internal governance function dedicated to governing and regulating the approval and use of AI systems within the organisation. A failure to govern internal staff use of AI systems can lead to a situation known as shadow AI, where an organisation has no oversight, control, and management over (i) which AI systems are used; (ii) how these AI systems are used internally; and (iii) which data and information is shared with such AI systems. Failure to implement proper governance processes could expose the organisation to both reputational and legal risks.

Many existing organisational contracts and policies were drafted before AI became widespread, and in today’s landscape, this anachronism presents growing legal and operational risks for organisations that cannot be overlooked or left to chance.

Outdated internal policies: A blind spot in AI governance

Internal policies that were drafted and implemented prior to the AI ‘boom’ often fail to address the unique challenges posed by modern AI technologies, creating significant internal governance risks. Organisations must ensure that they continuously review and amend their internal policies to remain up to date with evolving AI standards, best practices, ethical and other guidelines and frameworks. Corporate governance frameworks, including IT policies, acceptable use policies, data protection guidelines, and cybersecurity protocols, which were developed without AI in mind, often fail to address, inter alia:

• the approval, deployment and use of generative AI tools by employees;
• the risk posed by data and (personal) information inputted into AI systems and the risk of leaking confidential, proprietary, and personal information to AI models;
• use of AI-generated outputs in business-critical functions; and
• liability for decisions influenced by AI recommendations or outputs.

Why organisations need a bespoke AI Policy

A one-size-fits-all approach to AI governance is no longer sufficient. Each organisation operates within a unique context defined by its industry, data practices, risk tolerance, and ethical values. A bespoke AI Policy ensures that AI systems are aligned with specific business goals, regulatory requirements, and stakeholder expectations. It also empowers teams to manage AI responsibly, fostering innovation while minimising unintended consequences and risks. Tailored policies are not just a safeguard but serve as a strategic asset in the age of AI systems. AI governance policies should not replace existing governance documents but cross-reference and supplement them, integrating with acceptable use, cybersecurity, procurement, and other relevant policies. It is also important to ensure that existing policies are amended to reference the AI governance policy.

Reviewing legacy contracts: MSAs and AI risk allocation

Many Master Services Agreements (“MSAs”), procurement contracts, and technology service agreements in use today were negotiated before the rise of  AI and do not contain contractual provisions addressing the key risks introduced by the use of AI systems. This creates a material gap in how legal risk is allocated when service providers integrate AI into their service offerings (often without disclosure to their customers).

Some of the additional key risks that legacy contracts need to be updated with include:

• Hallucination: AI-generated outputs can be incorrect, misleading, or biased. Customers must ensure that their contracts with service providers include contractual provisions requiring service providers to, inter alia, (i) notify them of their use of AI systems; and (ii) implement human-in-the-loop reviews to fact-check and vet any outputs generated by AI systems.
• Automated decision-making: Where services involve algorithmic decisions (for example, in HR, credit scoring, or fraud detection), organisations must comply with laws governing automated processing, such as the Protection of Personal Information Act, 2013.
• Undisclosed use of AI to cut costs: Some service providers may quietly adopt AI tools to automate service delivery, reducing internal costs significantly, but continue to invoice clients at legacy rates premised on manual work or human expertise, resulting in clients overpaying for AI-delivered services.
• Bias and discrimination risks: If a service provider’s AI systems are trained on flawed or biased data, outputs could perpetuate discrimination or inequity, especially in sectors such as recruitment, insurance, or lending. This could expose the client to reputational damage or legal liability.

Reviewing contractual templates: Are you contracting in the AI age?

Standard templates, including non-disclosure agreements, Master Services Agreements, software development agreements, Software-as-a-Service contracts, and even employment contracts, should be updated to reflect:

• Approval of use of specific AI tools;
• Ownership of AI-generated content and outputs;
• Data usage rights, especially where client data is inputted into service provider AI systems;
• Limitation of liability, limiting the customer’s liability;
• Service provider indemnities and warranties, addressing, inter alia, compliance with applicable laws, human-in-the-loop reviews, and bias identification and mitigation; and
• Service provider’s obligation to disclose AI use in delivering deliverables or services.

Organisations should audit existing contractual templates to ensure alignment with business expectations for the use and deployment of AI systems, whilst mitigating the risks introduced by such AI systems.

Are customers aware of what they are paying for?

Many organisations may not realise that their service providers are already using AI and incorporating outputs generated by such AI systems into deliverables and services. AI-driven solutions can significantly reduce operational costs, but where these efficiencies are not passed on to the customer, the customer may be overcharged for services without realising the value of AI in the services. As such, customers should be aware of any service provider’s use of AI systems in the rendering of services, which should be governed by appropriate contractual terms in the relevant service provider agreements. This is another critical reason why legacy contracts need to be updated.

AI will not wait for your policies and contracts to catch up

AI adoption is moving rapidly, far faster than internal governance structures can manage it. To stay legally compliant, competitive, and in control, organisations must ensure that they (i) review and amend their existing policies and implement an AI Governance Policy; and (ii) review and amend their service provider agreements to include AI-specific clauses. A proactive legal strategy in the AI space is no longer a nice-to-have; it is a competitive and regulatory necessity.

For more information or advice on AI governance, please contact our experts:

Ridwaan Boda
Executive | Technology, Media and Telecommunications
rboda@ensafrica.com

Alexander Powell
Associate | Technology, Media and Telecommunications
apowell@ensafrica.com

Aobakwe Motebe
Candidate Legal Practitioner | Technology, Media and Telecommunications
amotebe@ensafrica.com

The post How to draft precise Statements of Work: Lessons learned from real-world deals appeared first on intelligENS.

04 Aug 2025

by Janine Thomas and Craig Shapiro

The rapid advancement of generative artificial intelligence (“AI”) is transforming industries across the globe, and South Africa is no exception. From creative content generation to business process automation, generative AI is being adopted by South African enterprises, startups, and individuals alike. However, as with any disruptive technology, the legal landscape, particularly in relation to copyright, presents both challenges and opportunities. This article discusses the intersection of generative AI and South African copyright law, highlights key legal risks, and offers practical guidance for navigating this evolving field.

Territorial nature of copyright and enforcement

Copyright is inherently territorial, meaning that protection is granted by the laws of each individual country. There is no such thing as a “global copyright.” Instead, international treaties such as the Berne Convention require member countries (including South Africa) to provide certain minimum standards of protection to works originating from other member countries. However, the scope, duration, and exceptions to copyright protection can vary significantly from one country to another.

Enforcement of copyright is also territorial. If infringement occurs in a particular country, the copyright owner must rely on the laws and courts of that country to enforce their rights. For example, a South African copyright owner whose work is infringed in the United States would need to bring an action in the US courts under US copyright law. Similarly, if a work is infringed in South Africa, South African law and courts would apply.

This territoriality means that a work may be protected in some countries but not others, and the outcome of copyright disputes can differ depending on the jurisdiction. It also means that compliance with copyright law in one country does not guarantee compliance elsewhere, especially given differences in exceptions (such as “fair use” in the US versus “fair dealing” in South Africa) and in the recognition of rights in AI-generated works.

Training vs. Generative AI

Generative AI refers to systems capable of producing new content—such as text, images, video, music, or code—based on training data.

The training of the AI on the other hand refers to the process by which an AI system learns to perform specific tasks or make decisions by analysing data. The training process varies depending on the type of AI and the task at hand, but follows a set of common steps, which include collecting the data and exposing the model to the data.

The distinction between using materials to train the AI model and the AI-generated outputs is quite important, and both have significant IP risks.

While AI models offer efficiency and creativity, they also raise important questions about intellectual property rights, particularly when AI systems are trained on or generate content that may be protected by copyright.

Two Key U.S. Cases

Two recent U.S. cases provide insight into how courts are grappling with the use of copyright works in the training of AI models:

• Bartz v. Anthropic PBC: In this case, three authors sued Anthropic for using their books to train its Claude AI model. The court ruled that using purchased (authorised) books for training was fair use due to the transformative nature of the use. However, using pirated (unauthorised) books was not fair use, and a trial on damages will follow.
• Kadrey v. Meta Platforms: Thirteen authors, including Sarah Silverman, sued Meta for using their books to train its Llama AI model. The court found Meta’s use highly transformative and ruled it as fair use.

Both above cases related to the materials used to train the AI models and not AI-generated content. What is important to note is that the use of the copyright works to train Clause and Llama would have been infringement save for the “fair use” defence. This defence is not available under South African law and if the above cases were decided in South Africa (under South African law) the outcome would have been different.

Fair use vs. fair dealing: A comparative perspective

The U.S. doctrine of “fair use” allows for a broad, flexible defence to copyright infringement, considering factors such as purpose, nature, amount, and effect on the market. This enables courts to adapt the law to modern technologies, including generative AI.

In contrast, South Africa’s “fair dealing” is more limited, permitting use only for specific purposes enumerated in the Copyright Act. This restrictiveness means that many uses of copyrighted material for AI training or generation may not qualify as fair dealing, increasing legal risk for South African users and developers.

South African Copyright Law and Generative AI

South African copyright law is primarily governed by the Copyright Act 98 of 1978 (“Copyright Act”). The law grants exclusive rights to the copyright owner, including the right to reproduce, adapt, and distribute the work.

When it comes to generative AI, two key legal questions arise:

1. Training Data: Is it lawful to use copyrighted works as training data for AI models?
2. AI-Generated Outputs: Who owns the copyright in works generated by AI, and can these outputs infringe existing copyrights?

Ownership of AI-Generated Content

Copyright arises automatically in original works such as literary, artistic, and musical works, provided certain requirements are met (such as originality and fixation in a material form).

The owner of copyright is generally the author of the work, who must be a natural person for most categories of works. However, for “computer-generated works,” the law recognises the person who made the necessary arrangements for the creation of the work as the author. This is particularly relevant for AI-generated content: if a human exercises sufficient control or input in generating the work (i.e. the prompt used), they may be considered the author and thus the copyright owner. If no human author can be identified, the work may not qualify for copyright protection under South African law.

Why is this important you may ask. The Terms of Service of most AI platforms provide that the user is the owner of copyright in the AI-generated works. However, it is not that simple. Copyright ownership is a factual question and the platform owner cannot grant you ownership contrary to the law, nor can it transfer rights to you that it does not own.

Legal defences against copyright infringement claims

The rights conferred by copyright are exclusive and include the right to reproduce, publish, perform, broadcast, adapt, and distribute the work. Only the copyright owner (or someone authorised by them) may exercise these rights.

If a generative AI system produces content that is substantially similar to a copyrighted work, the user or developer could face infringement claims. However, several legal defences may be available under South African law:

• Fair Dealing: South Africa recognises the doctrine of “fair dealing,” which permits limited use of copyrighted works for specific purposes such as research, private study, criticism, review, or reporting current events. Unlike the broader U.S. concept of “fair use,” fair dealing is more restrictive and purpose-specific.
• Lack of Substantial Similarity: If the AI-generated work is sufficiently transformative or does not reproduce a substantial part of the original, infringement may not be established.

Risks of copyright infringement and practical strategies

The use of generative AI carries several copyright risks:

• Unknowing infringement: Be aware of the extent that your suppliers use AI to generate content and materials for you or your business. You may unknowingly use content that infringes copyright and would still be liable for the consequences thereof (including removal of the materials from your online and physical properties).
• Unintentional copying: AI models may inadvertently reproduce protected elements from their training data.
• Unclear ownership: The absence of clear authorship for AI-generated works can complicate licensing and enforcement.
• Third-party claims: Users and developers may both face infringement claims from rights holders if AI outputs are deemed infringing.

To mitigate these risks, consider the following strategies:

1. Use licensed or public domain data: Ensure that the database is curated and is either licensed for such use or in the public domain.
2. Implement technical safeguards: Use tools that detect and prevent the reproduction of protected content.
3. Maintain transparency: Keep records of data sources and referenced materials used.
4. Obtain legal advice: For high-risk applications, seek guidance on compliance with copyright law.

Other legal and practical considerations

• Contractual arrangements: Clearly regulate the extent to which content may be AI-generated and define ownership and liability in contracts with your vendors and suppliers.
• Data protection: Ensure compliance with the Protection of Personal Information Act (“POPIA”) when using personal data in AI systems.
• Policy Developments: Monitor legislative and policy developments, as South Africa may update its copyright framework to address AI-specific issues.

Conclusion

Generative AI offers immense potential for innovation in South Africa, but it also raises complex copyright issues. By understanding the legal landscape, adopting risk mitigation strategies, and staying informed about global developments, South African businesses and creators can harness the power of AI while respecting intellectual property rights. As the law evolves, ongoing dialogue between technologists, legal professionals, and policymakers will be essential to ensure a balanced and forward-looking approach.

Craig Shapiro
Executive | Intellectual Property
cshapiro@ENSafrica.com

Janine Thomas
Executive | Intellectual Property
jthomas@ENSafrica.com

The post Generative AI and IP in South Africa: Navigating legal complexities and opportunities appeared first on intelligENS.