Case study
Which news sources does AI trust in South Africa?
Six AI models answered two trust questions across six reader segments: which South African news sources are most trustworthy, and which journalists or public personas to follow. The answers show a compact AI trust map: News24 and Daily Maverick dominate reach, while public-interest outlets supply much of the trust vocabulary.
- AI models
- 6
- audience segments
- 6
- responses
- 1,452
- sources tracked
- 12
The headline
AI trust clusters around scale plus visible editorial standards
AI systems do not answer the trustworthy-news question with one source. They build a shortlist. News24 and Daily Maverick are the strongest source names, while Mail & Guardian anchors the second tier and public-interest outlets explain much of the trust language.
The trust map favours scale plus visible editorial standards. News24 and Daily Maverick dominate the source ranking, but specialist public-interest sources such as GroundUp, amaBhungane, and Africa Check give the field its credibility cues.
Source ranking
The top two own the broad recommendation space
News24 has the widest mention share at 25.1% and is the first named source in 60.1% of responses. Daily Maverick follows at 12.4% share of voice and 27.2% top-of-mind recall, close enough to read as a co-anchor rather than a distant challenger.
Below the top two, the category changes shape. eNCA, GroundUp, SABC News, and Mail & Guardian each hold meaningful share, but their strategic roles differ: national reach, local accountability, public broadcaster familiarity, or editorial-depth trust.
Trust vocabulary
Public-interest cues explain why sources are trusted
The models use public-interest cues to justify the ranking. Investigative work, editorial independence, fact checking, and accountability appear as explicit reasons for trusting a source; that benefits outlets whose public role is easy for AI systems to retrieve and summarise.
These are extracted entities from the responses, not brand claims written into the case study. They show the language AI systems already use when they explain trustworthy news in South Africa.
Risk language
Low-sentiment visibility is a warning signal
Source visibility is not always positive. IOL appears in the source set, but with materially weaker sentiment than the other focus sources. That means the brand often appears beside concerns about editorial independence, credibility, or industry disputes.
The negative vocabulary is mostly about information integrity: misinformation, paywalled content, sensationalism, clickbait, and political agenda. For publishers, AI trust depends on visible proof: standards, corrections, ownership transparency, bylined expertise, and verifiable public-interest work need to be machine-readable.
The human layer
AI turns institutional trust into people to follow
The second question moves from institutions to people: which journalist or public persona should someone follow for news? The answer set is more fragmented than the source ranking, but the leading names are still clear.
Karyn Maughan and Ferial Haffajee lead this layer. The next tier mixes investigative reporters, broadcast anchors, political analysts, and public intellectuals rather than one single kind of news personality. The people layer shows how AI systems make trust practical for a user: a source gives an institution to visit; a journalist gives a feed, beat, or voice to follow.
Audience segments
The same leaders recur, but the rationale changes
The campaign included six audience segments plus a no-segment baseline. Segment prompts did not create a different universe of trusted sources, but they changed which source types were easiest for AI systems to justify.
Community and local focusers make local accountability more salient. Sceptical scrutinisers make verification and independence more salient. Social and mobile grazers make accessibility and followability more salient. Loyal legacy trusters make institutional familiarity more salient.
Method
How this study was run
Two questions ran through six completed AI model sets across six audience segments plus a no-audience baseline: trusted news sources and people to follow. Entity variants were resolved before source, journalist, trust-driver, and risk metrics were calculated.
Perplexity responses were excluded after the provider quota was exhausted and the circuit breaker activated. That left 1,452 completed responses for the final metrics.
Run this study on your market.
Book a demo to run the same research design on your category.