South Africa Elections, An AI Case Study
Microsoft Elections Communications Hub worked with election authorities and campaigns. Here's what we learned.
The first US presidential debate mentioned AI, zero times. And crypto zero times. An observation only, amid everything else…
2024 is a major year for elections, with over 60 countries heading to the polls. That’s nearly half the world’s population getting a say in how their countries are run, according to Statista.
Election Benefits of AI:
Direct voter engagement - AI can interact with voters through more personalized messaging
Transparency and accessibility - chatbots provide instant info on polling locations, candidate platforms, and voting procedures.
Better Data Management - efficient collection, storage, and analysis of electoral data can help officials make quick decisions and spot trends easier
Efficient election administration - AI can handle large datasets at rapid speed, reducing human error and ensuring accurate and speedy results.
But …
AI algorithms can be manipulated to favor certain candidates or parties, either intentionally or unintentionally. Biases in AI systems can arise from the data used to train them or the algorithms’ design. This can reinforce existing forms of discrimination, prejudice and stereotyping. Biased AI algorithms could unfairly influence voter behaviour, compromising the fairness of the electoral process - UN
Election Risks
Disinformation - AI makes disinfo more believable and emotionally impactful, in turn making it harder to detect and counter.
Deepfakes - highly realistic but fake audio, video, and images, misleading voters by making people appear to do or say things they never did.
Sophisticated Tools - it’s getting harder and harder to tell between real and fake as the tech gets better.
Privacy Concerns - collecting large amounts of personal data means there needs to be strong data protection measures to maintain voter trust and comply with privacy laws.
So…
How do we make sure AI systems are free from bias and do not unfairly influence election outcomes?
Who will be responsible for overseeing AI use in electoral processes to ensure transparency and accountability?
Case Study: Microsoft & South Africa’s Election
Microsoft Elections Communications Hub worked with election authorities and campaigns in South Africa by using Microsoft security and support teams leading up to the election.
Microsoft hosted a Combating the Deceptive Use of AI in Elections capacity building workshop in mid-May which focussed on combatting AI generated mis/disinformation and strengthening election integrity in South Africa. The workshop was attended by Political Campaigns, IEC officials, regulators, Fact Check Africa and the tech industry - Microsoft
Three Takeaways
Local actors did attempt to sway the vote through disinformation and influence-for-hire schemes on social media.
The Daily Maverick had been monitoring trends running up to the election, and revealed that buying and selling influence on the platform X is becoming a significant trend that needs close monitoring. Often many of these influencers are anonymous and operate from the political fringes to inject highly emotive and polarizing content, driving their engagement, but at a huge cost to social cohesion. You can read more about it here.
Investigations showed that paid influencers used automated accounts, known as sock puppets, and were driving social media movements as early as 2021. AI will make this even easier.
While this election featured reported deepfakes, none were high-impact, owing, in part, to the work of fact-checking organizations.
Microsoft highlighted how Africa Check were able to quickly analyze and debunk a deepfake video of Eminem supposedly criticizing the ANC party. And the video got hundreds of thousands of views. We covered it at the time on the substack, here.
How it was spotted by Africa Check:
Africa Check did a simple reverse image search from screenshots of the Eminem deepfake came back with longer video clips, with more context showing it was actually recorded in 2009 on a British talk show.
They also noticed that the lips and audio in the deepfake weren’t syncing up, and could tell there was some alteration in the voice.
You can read more from The Rundown about how to spot deepfakes here.
Kudos to The Real 411 initiative, crowdsourcing info
The South African Independent Electoral Commission (IEC) launched Real 411, a platform to combat election disinformation. At Real411, users can report anything that looks like suspected info manipulation, which the IEC investigates. The IEC created this with Media Monitoring Africa (MMA), which uses data tools for great quality media monitoring. The goal is to protect electoral integrity by involving the public in identifying and reporting potential misinfo.
How it worked:
1. Submit complaint via Real411 website
2. IEC's Directorate for Electoral Offences reviews with experts
3. Reviews based on guidelines and make recommendations to Electoral Commissioners
4. Electoral Commissioners make ruling on the complaint
5. Decision is communicated
6. Anyone unhappy with the decision can approach the Electoral Court
Global Outlook
The Alan Turing Institute looked at 112 national elections since January 2023 and found something pretty interesting: AI was only used in 19 of them. AI didn't seem to change election results much then, but they have a warning for the years ahead:
Things such as an erosion of trust in governments and information sources, deepened political polarization, the kind of consolidation of echo chambers. All of these things, I think, deepfakes and AI threats are just enhancing
Finally, when it comes to AI and elections, we should not be reactive - need to be proactive here. It’s also the only way we will keep up with growth in technology.