A Longitudinal Analysis of Public Discourse on AI Ethics in Education Using Twitter Data

Synthesis: Bagale and colleagues (2026) analyze five years of Twitter/X discourse (2019–2024) on AI ethics in education and find a public conversation that is overwhelmingly positive (81.65% of 14,201 tweets) but punctuated by four event-driven negative peaks, most decisively around the release of ChatGPT. The study argues that the public is not rejecting AI in education so much as asking for frameworks to use it responsibly, while warning that platform algorithms amplify technically sophisticated and aspirational voices at the expense of teachers and students.

Overview

As AI and generative AI moved into classrooms, the authors frame a problem for institutions: policy documents articulate principles without operational guidance, practitioner accounts are constrained by compliance logistics, and media coverage stays episodic and shallow. Where, then, does one look to understand what the public actually expects of AI in education? Their answer is social media, where debates unfold at scale and in real time, and where the ethical vocabulary is negotiated as events happen rather than after the fact. Twitter (referred to throughout by its former name) is treated as both a barometer of public sentiment and a site where those sentiments are contested and amplified. The paper's central concern is that if institutions read "public opinion" off algorithmically curated timelines, they may mistake a well-engaged technical minority for the broader stakeholder base of teachers and students.

Study Design & Method

• Corpus. 14,201 tweets with metadata (retweet, reply and like counts, hashtags, author id) collected via the twarc2 tool and Twitter's Academic Research API. The query combined AI terms ("artificial intelligence", "AI", "generative ai", "chatgpt", "llm", "large language model") with education and ethics terms ("education", "educator", "learning", "teach", "teaching", "ethics", "AIED", "ethical", "ethic"). Collection was restricted to English-language tweets from U.S.-based users, dated 2019-01-01 to 2024-11-10.
• Sentiment analysis. A supervised binary classifier built on SetFit, a sentence-transformer fine-tuning framework chosen for its robustness to short, informal social-media text and its label efficiency. Two annotators independently labeled 90 tweets; agreement was 91.1% with Cohen's κ = 0.797. The annotated set was split 80/20 into training and validation, and the trained classifier then labeled the full corpus. Confidence was high: 95.87% of predictions fell in the ≥0.80 confidence band. Month-level proportions of positive and negative sentiment were used to build a temporal timeline so that sentiment intensity could be read independently of tweet volume, with peaks detected against a local baseline.
• Topic modeling. BERTopic with the all-MiniLM-L6-v2 Sentence-BERT embedding model, combining UMAP, HDBSCAN (min cluster size = 30, min samples = 10) and c-TF-IDF keyword extraction. Search terms were treated as stop words to prevent over-representation. The initial run produced 57 topics with 6,520 outliers (46% of the dataset); c-TF-IDF reassignment cut outliers to 12 documents (0.08%). Because inter-topic distances showed heavy semantic redundancy (cosine similarity ≈ 0.8), the authors applied topic reduction to nrtopics = 16, chosen for semantic distinctiveness (merge distances 0.97–1.64).
• Comparison. The paper positions Twitter discourse against three other text types — policy documents, practitioner accounts, and media coverage — arguing that each has a different temporal rhythm and risk vocabulary: policy balances rights against innovation, media highlights crises, practitioners stress mitigations, and the public alternates between integrity concerns and classroom experimentation.

Key Findings

• The discourse is predominantly positive. Of 14,201 tweets, 11,595 (81.65%) were classified positive and 2,606 (18.35%) negative — an asymmetry the authors read as resilient baseline optimism. The positive curve did not collapse even as negative events emerged.
• Four positive peaks, each tied to adoption milestones. June 2019 (edtech conferences and early curriculum integration across K–12 and higher education), May 2021 (cloud AI platform launches, most notably Google Cloud Vertex AI), April 2022 (enterprise AI and professional upskilling), and May 2024 (Google Gemini for Education, Microsoft and Khan Academy's Khanmigo free tutoring for U.S. teachers, and personalized learning platforms). The 2024 peak was the most pronounced and framed AI as complementing rather than replacing human teaching.
• Four negative peaks, each anchored to a specific controversy. April 2019 — the collapse of Google's AI Ethics Board, the largest early spike, read by users as performative ethics and governance failure; February 2021 — the firing of AI ethics researcher Timnit Gebru, framed as evidence that corporate labs cannot self-regulate; February 2023 — the largest negative peak after ChatGPT, driven by academic integrity worries, students using ChatGPT to pass MBA exams, AI-generated art and copyright, and sensitive-content handling, with Noam Chomsky's "high-tech plagiarism" characterization circulating widely; and the study's final negative event. Unlike the earlier institutional triggers, the 2023 surge was diffuse — systemic anxiety about AI's role in knowledge, creativity and authenticity rather than reaction to one event.
• Four themes dominate the topic model. Topic 1, AI in Education and Pedagogy, was 5,600 tweets (39.43%), framing AI as augmenting instruction through feedback generation, differentiation and personalized pathways with a "human-in-the-loop" ethos. Topic 2, Applied Machine Learning and Data Science, was 3,023 tweets (21.33%), a technically oriented stream dominated by tools, methods and health/medical case studies that had carried over into the education ethics stream. Topic 3, AI Ethics, Risk and Governance, was 2,958 tweets (20.83%), centering on algorithmic bias, equitable deployment and responsible data use. Topic 4, ChatGPT for Teaching, Writing & Coding, was 7.12%, and was disproportionately actionable — a dual focus on risk mitigation (cheating, plagiarism-detection limits, policy guidance) and constructive integration (prompt design, feedback generation, coding support, lesson planning) with repeated calls for educator training. All remaining topics fell below 3% of tweets.
• Temporal divergence across themes. AI-in-Education discussion peaked sharply in 2023 with the rise of ChatGPT before declining in 2024, while Applied ML, Ethics/Governance and the ChatGPT-in-practice theme followed steadier trajectories with noticeable increases after 2022.
• Algorithmic amplification shapes who is heard. The authors argue that engagement-optimizing algorithms reward technical sophistication and moral complexity, creating a feedback loop that privileges expert-driven contributions and may exclude educators and students without AI/ML backgrounds. The comparatively small share of practical implementation concerns (Topic 4) is read as algorithmic de-prioritization of mundane institutional challenges relative to aspirational or controversial content.

What this means for practice

• Administrators. Shape engagement rather than prohibit it: 11,595 of 14,201 tweets (81.65%) were positive, and the study reads the persistent optimism alongside deepening concern as a public request for frameworks for responsible use, not for bans.
• Administrators. Do not read policy priorities directly off trending timelines, because engagement-optimizing algorithms amplify technically sophisticated and aspirationally framed voices and may exclude teachers and students without AI/ML backgrounds.
• Instructors. Update academic misconduct policy and build critical AI literacy instead of blanket prohibition — the message the authors take from the discourse.
• Designers. Treat social-media sentiment as a signal about the conversation, not as a proxy for the stakeholder base: the sophisticated bias-and-fairness reasoning in the discourse presupposes a technical literacy that narrows who participates, and the prominence of healthcare AI cases shows both useful cross-domain transfer and the risk of importing high-stakes medical analogies into classroom tools.
• Researchers. Use public discourse as a complement and corrective to top-down ethical framings — it is more ethically nuanced than media coverage and more operationally specific than policy documents — while remembering the analysis cannot show whether online debate translates into classroom change.

Limitations

• The corpus is Twitter/X only, restricted to English-language tweets from U.S.-based users collected 2019-01-01 to 2024-11-10 (14,201 tweets); Reddit, YouTube and academic forums are excluded.
• The sentiment classifier was trained on just 90 hand-labeled tweets (91.1% agreement, Cohen's κ = 0.797) and evaluated on an 18-instance validation set, while topic annotation reliability was checked on 65 double-coded tweets (κ = 0.715).
• The authors state that peak detection is sensitive to parameter choices and that annotating peaks from hashtags and selected tweets introduces interpretive subjectivity.
• The platform's user base skews more technically literate and professionally engaged than the general population, and its algorithm may overrepresent emotionally charged posts, which limits generalization.

Citation

Bagale, A., Mehjabin, N., Unlu, A., Johri, A., et al. (2026). A Longitudinal Analysis of Public Discourse on AI Ethics in Education Using Twitter Data.