Predicting political misinformation on X — lessons from Kenya's digital landscape, by Doreen Nkirote

I did not start studying misinformation because it was a fashionable research topic. I started because I watched a false claim about election results travel from a single account to half a million people in under four hours during Kenya's 2022 general election — and I could not stop thinking about the mechanics of how that happens, and whether a machine could learn to see it coming.

That question sits at the centre of my PhD research at USIU-Africa, and has already produced two published papers: a systematic literature review on the key characteristics of political misinformation on X in Kenya, and a study exploring the link between content gratification and misinformation-sharing behaviour among Kenyan social media users. This post draws on both to lay out what we know, what we can now predict, and what the limits of a machine learning approach actually are.

I want to be direct about something from the outset: this is not a triumphalist AI story. Machine learning can help detect and predict misinformation at a scale no human moderator team could match. But deployed carelessly — without understanding the social and linguistic context of the platform it is operating in — it can also suppress legitimate political speech, encode the biases of its training data, and create a false sense of certainty about a problem that is fundamentally a social one, not a computational one.

Kenya's information environment: why it matters

Kenya is one of Africa's most connected countries, with one of the continent's highest rates of smartphone penetration and a deeply engaged social media culture. X (formerly Twitter) is disproportionately influential in shaping political discourse relative to its raw user numbers — it is where journalists, politicians, civil society actors, and opinion leaders spend their time, and where narratives originate before cascading to WhatsApp, Facebook, and TikTok.

This concentration of political influence in a single platform creates a specific vulnerability. A coordinated inauthentic operation — a cluster of accounts posting the same false narrative in a short window — can saturate the platform's algorithmic feed fast enough to reach mainstream media before any verification process can catch it. Kenya has experienced this pattern in each of the last three election cycles, with documented campaigns around vote tallying, candidate eligibility, ethnic violence rumours, and fabricated statements attributed to real public figures.

My SLR on political misinformation characteristics on X in Kenya, published in the International Journal of Technology and Human Interaction, identified three structural features that consistently distinguish misinformation content from accurate content in the Kenyan context: emotional amplification language, strategic vagueness about source attribution, and timing relative to election milestones. These features are detectable — with the right features engineered into a model.

A brief history of the problem

False reports of ethnic violence circulate on early Twitter and SMS networks, later linked to real-world incidents in Mombasa and Kisumu. IEBC results portal crash triggers a wave of competing tallies on social media.

Cambridge Analytica's involvement in the Jubilee campaign becomes a reference point. Fabricated opinion polls and edited video clips spread widely. The nullified election produces two distinct misinformation waves separated by three months.

Misinformation moves in coordinated clusters across X, Facebook, and WhatsApp simultaneously. Swahili and Sheng content is under-moderated by platforms. A false result declaration spreads to 500K+ accounts in under four hours before correction.

Synthetic audio clips attributed to political figures, AI-generated images of candidates, and LLM-assisted copy production are documented for the first time at scale in Kenyan political discourse. Detection methodology must evolve accordingly.

What makes Kenyan political misinformation distinctive

One of the most important findings from my research is that political misinformation in Kenya is not simply a local version of a global phenomenon. It has structural characteristics that differ meaningfully from the misinformation patterns studied in the US and European literature — and those differences matter enormously for model design.

First, language is not monolithic. Kenyan political discourse on X moves fluidly between English, Swahili, and Sheng — a Nairobi-origin creole that blends Swahili, English, and ethnic languages in patterns that are highly contextual and constantly evolving. A model trained on English-language misinformation datasets from the Global North will be essentially blind to Sheng-language content, which is precisely where the most inflammatory material tends to circulate, because it is also where platform moderation is weakest.

Second, ethnicity is the primary political organising principle, which means that the most consequential misinformation is ethnically coded. Claims are not always false on their face — they are distorted, decontextualised, or selectively framed to activate ethnic threat perceptions. Detecting this requires models that can identify not just factual inaccuracy but narrative framing — a harder problem by an order of magnitude.

Third, the gratification-sharing link is direct and measurable. My published paper in Online Social Networks and Media found that users who consume political content primarily for entertainment and status signalling — rather than information — are significantly more likely to share content without verification. This is not a moral failing; it is a predictable behavioural pattern that can itself be used as a model feature.

Content gratification type is a stronger predictor of misinformation sharing behaviour than education level or media literacy self-assessment. This has a direct implication for intervention design: literacy campaigns alone will not move the needle. Platform design that changes the default gratification loop is more likely to.

"A model trained on American election misinformation will fail in Nairobi. Not because the mathematics is different — but because the language, the context, and the social logic are."

Building a machine learning prediction model: the feature engineering challenge

The core research question of my PhD is whether a machine learning model can predict — not just classify after the fact — political misinformation on X in a Kenyan context. Classification is a solved-enough problem: given a post that has already been labelled as misinformation or accurate, a well-trained classifier can achieve respectable accuracy. Prediction is harder. It requires identifying signals that are present before the content goes viral, or even before a piece of content has enough engagement to be evaluated at the account level.

The feature space I have developed through iterative research breaks into four categories. Each encodes a different dimension of the misinformation signal.

• Emotional intensity scores (valence, arousal)
• Hedging and vagueness markers
• Unverified source attribution patterns
• Cross-language code-switching frequency
• Named entity density (people, places, institutions)

• Account age relative to posting surge
• Follower/following ratio anomalies
• Retweet velocity in first 30 minutes
• Coordinated posting time clustering
• Quote-tweet to retweet ratio

• Distance to nearest election milestone
• Time-of-day posting distribution
• Burst pattern detection (posting frequency spikes)
• Content lifecycle relative to news cycle

• Reply sentiment toward content
• Engagement-to-impression ratio
• Community cluster membership
• Historical sharing behaviour of early amplifiers

Model performance: what actually works

Across the model architectures I have evaluated on the Kenyan-context dataset, the results are illuminating — not because any single approach is clearly dominant, but because the performance gaps reveal something about the nature of the problem itself.

Traditional NLP classifiers (Naive Bayes, SVM) perform reasonably well on purely textual features but degrade sharply when the post is in Sheng or contains a high proportion of code-switched content. Transformer-based models fine-tuned on English political text show strong English performance but similar degradation on multilingual content. The approach that has shown the most promise in my research is a multi-modal ensemble that combines a multilingual transformer (fine-tuned on a Swahili and Sheng corpus) with network propagation features — capturing both the content signal and the behavioural signal simultaneously.

The performance gap between English-only models and multilingual ensemble approaches narrows considerably on English-dominant content but widens dramatically — sometimes by 20+ percentage points — on Sheng-dominant threads. This is not a minor technical footnote. In the Kenyan context, Sheng is the language of the street, of youth political mobilisation, and of the most emotionally charged content. A model that cannot read Sheng cannot read Kenyan political misinformation. It is that simple.

All model evaluation in my research uses a dataset labelled by a panel of Kenyan journalists, fact-checkers, and academic researchers — not crowdsourced annotation. The labelling protocol explicitly accounts for satire, hyperbole, and culturally specific exaggeration that does not constitute misinformation despite surface-level similarity. This distinction is critical to avoiding a model that flags vigorous political opinion as false information.

The patterns that emerge: what misinformation looks like before it spreads

One of the more striking outputs of the prediction modelling work is the identification of consistent pre-viral patterns — signals that appear in the data before a piece of misinformation achieves the engagement threshold that makes it visible to platform moderators or fact-checkers.

The highest-risk misinformation posts show a distinctive engagement curve: slow initial uptake (under 10 interactions in the first 20 minutes), followed by a sudden acceleration between minutes 30 and 90, typically triggered by amplification from an account with above-average follower counts. Intervening during this window is significantly more effective than after it.

Misinformation campaigns in Kenya consistently involve a small core of accounts — typically 5 to 15 — whose posting and retweeting behaviour is temporally synchronised to a degree that is statistically improbable for organic activity. Identifying these clusters early is more reliable than classifying individual posts in isolation.

The baseline false-positive rate for political content rises sharply in the 14 days before and 7 days after an election — but so does the true positive rate for misinformation. Models must be recalibrated for election-period operation; a threshold tuned for normal conditions will produce unacceptably high false-positive rates during the surge window.

A growing proportion of misinformation originates on closed platforms (WhatsApp groups, private Telegram channels) and is seeded onto X by a small number of bridge accounts. The X post is not the origin — it is the amplification step. Detecting this pattern requires looking at the network structure of early amplifiers, not just the content of the post itself.

What the model cannot do — and why that matters

I want to be precise about the limits of the approach, because they are not incidental — they are structural, and anyone deploying these tools in a real political information environment needs to understand them clearly.

First, the model predicts virality risk, not truth. A high-risk score means the post exhibits the linguistic, network, and temporal characteristics associated with past misinformation campaigns. It does not mean the post is false. Satire, hyperbole, contested factual claims, and vigorous political rhetoric can all trigger high-risk scores. Every high-risk flag requires human review before any action is taken.

Second, the model will reflect the biases of its training data. If the labelled dataset under-represents a particular political tendency or ethnic community — either because the labellers were not sufficiently diverse or because the sampling methodology over-weighted certain account types — those biases will be encoded in the model's predictions. Ongoing audits of false positive rates across demographic groups are not optional; they are a fundamental operating requirement.

Third, adversarial adaptation is real and fast. As soon as a detection approach is known, operators of misinformation campaigns adjust. The evasion techniques documented in the 2022 election cycle included deliberate spelling variation to defeat keyword matching, the use of memes and image text to evade NLP analysis, and the fragmentation of claims across multiple posts to avoid triggering single-post classifiers. A detection system that is not continuously updated will be obsolete within months of an election cycle.

Coordinated inauthentic behaviour from state-adjacent actors is qualitatively different from organic misinformation. It has more resources, more sophisticated evasion, and a clearer strategic objective. A detection system designed for opportunistic misinformation may underperform significantly against a coordinated state-level campaign. Being honest about this distinction is essential for calibrating expectations — and for policy conversations about where technical tools can help and where legal and diplomatic instruments are needed instead.

Recommendations for platforms, researchers, and policymakers

Based on three years of research in this space, these are the interventions I believe will move the needle — in order of the evidence supporting them.

• 01 Invest in African-language NLP infrastructure

  There is no shortcut here. Effective detection of Kenyan political misinformation requires training data and model capacity in Swahili and Sheng. Platform investments in African-language NLP have been a fraction of those in European languages. This gap is not technically justified — it is a resource allocation decision, and it can be changed.

• 02 Fund pre-election detection infrastructure, not post-election inquiries

  The pattern of response in Kenya has been consistent: a misinformation surge occurs, an inquiry is commissioned, findings are published after the political moment has passed, recommendations sit unimplemented. The intervention needs to happen before the election, not after. Pre-electoral monitoring programmes with standing technical infrastructure are more effective and far less expensive than post-hoc remediation.

• 03 Make fact-checkers part of the model pipeline, not separate from it

  Kenya has several excellent fact-checking organisations — Africa Check, PesaCheck, and others. Their output is underutilised as a training signal. A formal integration between ML detection systems and fact-checker databases — where verified claims feed directly into model retraining cycles — would compress the time between detection and validated response significantly.

• 04 Design for friction, not removal

  The evidence on content removal as a misinformation intervention is, at best, mixed. Removed content often migrates to less-moderated platforms and gains credibility as "suppressed truth." Friction interventions — interstitial prompts that ask users to read before sharing, labels that indicate content is disputed, reduced algorithmic amplification — have shown more consistent positive effects with lower risks of unintended consequences.

• 05 Localise the research base

  The majority of published research on social media misinformation uses US or European datasets. Findings do not transfer cleanly across political and linguistic contexts. African research institutions need the funding, infrastructure, and journal access to conduct and publish context-specific work. Policy decisions about African information environments should not rest primarily on findings from a 2016 US election study.

What I have learned — and what I am still learning

Three years into this research, I am more convinced than ever that machine learning has a genuine role to play in making Kenya's information environment healthier around elections. I am also more aware of where the limits are, where the risks of overreach are real, and where the intervention that will actually make a difference is social, institutional, or political — not algorithmic.

The most important lesson is the one that applies to every domain where AI is being applied to complex human problems: the model is a tool for amplifying human judgement, not replacing it. A system that flags misinformation at scale is valuable only if the humans receiving those flags have the time, the authority, and the context to act on them wisely. Building that capacity — in platforms, in fact-checking organisations, in regulators — is where the real work lies. The model is the easy part.

Doreen is a data scientist, AI researcher, and Adjunct Lecturer in Cyber Forensics at Riara University, Nairobi. Her PhD research at USIU-Africa focuses on building machine learning models for predicting political misinformation on X in the Kenyan context. She is CEO of Sd Pro Technology Ltd and holds a Master's in Data Analytics from KCA University.

Dr Doreen Bundi

IT Professional, Researcher & University Lecturer. Data-driven IT and research professional with 7+ years of experience spanning data analytics, AI-driven solutions, and academia. Proven expe