Findings supplied to AfricaBusiness.com point to practical questions for employers and educators about verification, confidential information and the difference between using AI often and using it well.
Among 364 South African AI users surveyed, 83% said they used the technology several times a week or more. Yet fewer than half reported always taking precautions when interacting with AI services.
The findings, supplied by Kaspersky in response to questions from AfricaBusiness.com, offer a snapshot of how one group of users incorporates AI into work, study and personal life. They also raise a question for employers and universities: what support do people need once AI has become routine?
The South African figures cover users across different age groups, not Gen Z alone. They should not be read as a measure of AI adoption across the country’s population.
Frequent use spans work and everyday tasks
Kaspersky said the South African sample comprised 364 existing AI users aged 18–65. The research was conducted by its internal market research centre as a quantitative online survey in March 2026.
Of these respondents, 48% used AI daily or almost daily, 35% several times a week, 12% several times a month and 5% once a month or less.
Information search was the most frequently reported purpose, at 59%, followed by generating ideas at 51% and work at 50%. Studying, including respondents’ own studies or their children’s, accounted for 42%. These activities can overlap; they do not represent separate groups of users.
The work figure does not identify which tools people used, whether employers approved them or what information they entered. Likewise, the study figure is not a measure of university students’ AI use. For organisations, these distinctions matter when deciding what the research can tell them about their own staff or learners.
What the precautions question reveals
Asked whether they took protective measures when interacting with AI, 49% of South African respondents answered “always”, 41% “sometimes” and 10% “never”.
The question gave examples such as double-checking information and avoiding the disclosure of confidential information. It did not measure these behaviours separately. The results therefore cannot establish that 49% always verify AI answers, or that everyone else shares sensitive data. Kaspersky confirmed that the frequency of personal or sensitive information disclosure was not asked in the survey.
These are self-reported habits rather than an assessment of cybersecurity competence. The findings do, however, give institutions a reason to examine whether their guidance translates into repeatable practice.
“Familiarity with AI is not the same as understanding its limitations,” said Kim Grobbelaar, Consumer Channel Manager in the Middle East and Africa at Kaspersky, in written responses to AfricaBusiness.com.
“Someone may know how to prompt a chatbot effectively but still accept an inaccurate answer or share information they would normally keep private elsewhere online.”
Employers need rules people can apply
Grobbelaar argues that organisations should define approved services, set clear limits on confidential information and explain how employees can obtain suitable access when a task involves sensitive material.
“A conversational interface can feel informal, which makes it easy to forget that information submitted to an external service may be processed or stored outside the organisation’s direct control,” she said.
In practice, employers could translate an AI policy into examples drawn from ordinary work. Brainstorming a public campaign and uploading a customer database require different decisions. So do editing a published product description and summarising a confidential contract.
Employees need to know which combinations of tool, task and information are permitted, who can authorise exceptions and where to ask for help. A general instruction to “use AI responsibly” leaves those decisions unresolved.
Managed access can provide a way to establish controls, but an enterprise label alone is not an assessment of suitability. Organisations still need to examine the service’s terms, settings and access arrangements against the information involved.
For organisations seeking a structured starting point, the US National Institute of Standards and Technology provides a voluntary AI Risk Management Framework and a companion profile addressing generative AI risks. These offer resources for organising risk management; they are not a certification that a particular tool is safe. [3]
Verification belongs inside the workflow
Training should connect checking requirements to the consequences of an error. A generated list of possible workshop themes has a different purpose from a factual statement going into a client report.
For example, a team preparing a market briefing could require staff to trace important figures to original sources, check dates and definitions, and retain those references in the finished work. A fluent explanation would then be a starting point for analysis rather than its final evidence.
Managers could test these habits with short exercises containing an unsupported claim or an incorrect reference. The purpose would be to see whether staff notice the problem and know how to resolve it, rather than simply whether they have completed a training module.
These are possible organisational responses to the issues raised by the research, not interventions whose effectiveness the survey tested.
Education should preserve the learner’s reasoning
The education question is similarly about how the tool is used. Grobbelaar cautions against treating frequent use alone as evidence that skills are deteriorating.
“A good boundary to consider is whether AI is supporting the learner’s thinking or substituting for it,” she said. “Students should still be able to explain the reasoning in their own words and independently check important information.”
Institutions could make that distinction explicit in assignments. Where AI assistance is permitted, learners might explain how they used it, identify an output they corrected and discuss the evidence behind their final answer. Requirements should reflect the skill being assessed.
The survey does not measure learning outcomes. It cannot determine whether respondents understood more, retained less or performed better because they used AI. Those questions require evidence beyond reported usage frequency.
A personal conversation still involves a digital service
AI use also extended beyond practical tasks. Among the South African respondents, 26% said they discussed questions or issues they did not want to share with other people, while 22% used AI as a friend, including for casual conversation or emotional support.
These responses should not be equated with a diagnosis of dependency or proof that confidential information was disclosed. They do indicate that some users approach conversational systems in a personal way.
Grobbelaar’s concern is that a familiar conversational style may encourage assumptions about privacy and reliability. Whether a service retains information, uses it for other purposes or offers particular protections depends on its policies and configuration, rather than how reassuring the conversation feels.
For South African employers and educators, the practical task is to turn familiar use into informed use: clear boundaries for information, appropriate checks on outputs and opportunities for people to demonstrate their own judgement.
About the research
The South African findings were supplied in a written Q&A sent to AfricaBusiness.com on 17 September 2026, with methodology clarifications on 1 October. Kaspersky identified 364 existing AI users aged 18–65, surveyed online in March 2026 by its internal market research centre. Detailed recruitment and weighting information was not supplied. The findings describe this surveyed group and are not presented as nationally representative. The original international release’s Gen Z statistics are separate from the South African figures used here.
Sources
- Kaspersky, Africa Business Media Q&A, responses attributed to Kim Grobbelaar, supplied through INK&Co. on 17 September 2026. Unpublished written responses.
- Kaspersky methodology clarifications supplied through INK&Co. on 1 October 2026. Unpublished correspondence.
- National Institute of Standards and Technology, AI Risk Management Framework and linked Generative Artificial Intelligence Profile.