Author: wadiagounine

Posted on by — Leave a comment

How Universities Worldwide Are Responding to AI’s Disruption of STEM Education

Abstract network illustration in deep navy representing global institutional response

A peer-reviewed synthesis presented at ACM’s ITiCSE 2026 conference in Madrid, co-authored by researchers from the University of British Columbia, Makerere University in Uganda, the University of Glasgow, Carnegie Mellon, and three other institutions, captures the state of the response precisely: computing science education’s reaction to AI has so far been “rapid experimentation, provisional deployments, and ad hoc adaptations,” and the field is only now, in 2026, shifting from reactive scrambling toward deliberate, long-term redesign. This is what that shift actually looks like on the ground, across very different institutions and countries.

The ACM guidelines driving the reform wave

Much of the current curriculum overhaul traces back to a single source: the Association for Computing Machinery’s 2023-2024 curriculum guidelines, which universities are now implementing at scale. The University of Iowa’s Computer Science Chair, Alberto Segre, described the resulting 2026 curriculum revision as following those guidelines directly to identify “which components of the program may be less important for today’s world.” The University of Victoria’s 2026 overhaul, the most significant change to its program in years, moved its introductory AI course to third year specifically so students build core computer science fundamentals first, before AI-assisted work becomes the default.

The one addition nearly every reform shares

Across institutions with almost nothing else in common, from the University of Saskatchewan to St. John’s University to the University of Zakho in Iraq, the same specific change keeps appearing: a newly mandatory ethics or “computing, ethics, and society” course, often required as a prerequisite for upper-level classes. The University of Saskatchewan’s 2026 program notes now describes academic-integrity related AI policy as central enough that “almost all syllabi now feature specific AI usage policies,” which function as the final word in any grading dispute. This mirrors what we found writing about AI use policy design more broadly: institutions worldwide are converging on explicit, written AI policy as a baseline requirement, not an optional add-on.

This is not just a U.S. and European story

It’s worth being precise about how global this response actually is. The ITiCSE 2026 paper’s own author list spans Uganda, the UK, the US, and Canada. Separately, the University of Zakho in Iraq formally approved a full AI Engineering Department curriculum for 2026-2027 in June 2026, and legislative bodies in U.S. states including New Jersey, Mississippi, and Georgia have passed laws in 2026 mandating K-12 AI or computer science instruction. The response is not confined to elite Western institutions, it is a genuinely worldwide adjustment happening on parallel, largely independent tracks.

France’s more centralized approach

Where U.S. and Canadian universities are reforming individually, France has moved at the national level: mandatory AI learning pathways now reach more than 1.5 million pupils annually in grades roughly equivalent to 8 and 10, and starting the 2027 school year, one hour of AI instruction per week becomes mandatory in the first year of upper secondary school nationwide. It is a useful contrast to the more decentralized, institution-by-institution reform happening across most English-speaking university systems.

What this means in practice

Three patterns recur across nearly every reform effort surveyed here: delaying heavy AI-tool reliance until core fundamentals are established, making AI-ethics and integrity instruction mandatory rather than optional, and treating written AI-use policy as the operational backbone of the classroom, not a footnote. None of these responses claim to have solved the underlying disruption described in our companion piece on falling CS enrollment, the ITiCSE researchers themselves call this “provisional,” but they represent the clearest signal yet of where the global academic community is actually converging, rather than merely debating.

Key takeaways

  • ACM’s 2023-2024 curriculum guidelines are driving a wave of 2026 CS program overhauls at universities from Iowa to Victoria to Saskatchewan.
  • A mandatory ethics or AI-policy course is the single most common addition across otherwise very different institutions worldwide.
  • This is a genuinely global response, not a US/Europe-only story, spanning Uganda, Iraq, and multiple continents in parallel.
  • France has taken a centralized, national-curriculum approach, reaching over 1.5 million pupils, in contrast to the decentralized university-by-university reform elsewhere.

Sources: Karimi et al., “Computing Science Education on the AI Innovation Landscape”, ACM ITiCSE 2026, Madrid; University of Iowa, University of Victoria, and University of Saskatchewan 2026 curriculum announcements; University of Zakho curriculum approval, June 2026; French Ministry of Education AI pathway announcements, 2026.

Posted on by — 1 Comment

Why Computer Science Enrollment Is Falling in the AI Era

Abstract bar chart illustration in cobalt blue representing declining enrollment

Enrollment in computer and information science programs at U.S. four-year universities dropped 8.1% in the 2025-2026 school year, the steepest decline of any field of study, with computer science specifically down 11.2%, according to National Student Clearinghouse data reported by Built In. For a field that spent over a decade as the default “safe” major, that is a striking reversal, and it is not happening for one simple reason.

The entry-level fear, specific to this field

Roughly 30% of code at major tech companies is now AI-generated, and AI tools now handle boilerplate code and basic debugging faster than most junior developers. This connects directly to the broader pattern we covered in the entry-level jobs fear: students are not imagining the shrinking junior-developer market, a Computing Research Association survey of 130 universities found 62% of participating institutions already reporting employer concern about this shift. What is specific to computer science is that the discipline’s own entry point, writing basic code, is the exact task AI has gotten disproportionately good at first.

“Cognitive surrender”: a term from MIT itself

In August 2026, an MIT faculty committee reported that its own students reach for AI at the first hint of struggle, and named the pattern cognitive surrender. This is not an outside critique of student behavior, it is one of the world’s leading technical institutions naming a problem in its own classrooms. The concern is specific: computer science, more than most fields, is learned by struggling through a bug or a broken proof, and that struggle is precisely what a fast, capable AI assistant removes if a student reaches for it too early.

What the newest international data shows

The OECD’s PISA 2025 results, released in September 2026, put a number on this. Across OECD countries, 46% of 15-year-olds use AI chatbots at least weekly for schoolwork. But students who used AI for specific tasks like summarizing or drafting scored roughly 20 points lower on science assessments than students who did not, a gap roughly equivalent to a full year of schooling. This is population-scale evidence, not an anecdote from one classroom, and it lands specifically on science and technical literacy, the exact domain STEM students are trying to build.

Why this isn’t simply “don’t use AI”

Learning science has described this mechanism for decades under the term desirable difficulties: retrieval practice, spaced repetition, and unresolved struggle are not obstacles to learning, they are the process itself. The OECD’s own 2026 review of classroom data across 14 countries reaches the same conclusion, echoing the pattern we found in our own piece on AI and critical thinking: the damage isn’t from AI use in general, it’s from using it to skip the specific struggle a skill is built on. For a CS student, that means the honest question isn’t “should I use AI,” it’s “which parts of this problem am I letting AI solve before I’ve actually tried.”

Key takeaways

  • U.S. computer science enrollment fell 11.2% in 2025-2026, the steepest drop of any major, driven largely by fear of a shrinking entry-level developer market.
  • MIT’s own 2026 faculty committee named a real pattern in its classrooms, “cognitive surrender”, reaching for AI before attempting a problem independently.
  • OECD PISA 2025 data found weekly-AI-use students scoring roughly 20 points lower on science assessments, about a year of schooling, than non-users.
  • The evidence points to a specific mechanism, skipped struggle, not AI use in general, which is a solvable, structural problem, not a reason to avoid the field.

Sources: National Student Clearinghouse data via Built In (2026); MIT faculty committee report, August 2026, via University 365; OECD PISA 2025 results (released September 2026); Computing Research Association employer survey, 2026.

Posted on by — Leave a comment

The Entry-Level Jobs Fear: What It Actually Means for Education

Abstract illustration in deep navy representing a narrowing pathway

Anthropic CEO Dario Amodei has repeated the same warning throughout 2026: AI could eliminate half of entry-level white-collar jobs within one to five years, pushing unemployment toward Great Depression-era levels among young professionals in finance, consulting, and technology. He is not alone. At the World Economic Forum, Google DeepMind’s Demis Hassabis said the effect was already visible at his own company, and the IMF’s managing director has separately warned of an AI “tsunami” coming for young people. This is a fear worth taking seriously, and also one that current evidence does not yet fully confirm, both things are true at once.

The warning, stated precisely

Amodei’s specific claim, repeated in interviews across 2026, is that AI models have progressed from roughly high-school-graduate capability to college-graduate capability in about two years, and are now capable of tasks like summarizing documents, brainstorming, and producing financial reports, tasks that used to define junior-level work. His prediction: entry-level roles get “augmented, but before long replaced,” with effects potentially visible within one to five years. He has also said that many AI and industry CEOs share this concern privately but have not said so publicly at the same scale.

What the data actually shows so far

Recent graduate unemployment has risen to its highest level since 2021, and graduates in fields like finance and computer science have reported unusually thin response rates to large numbers of job applications. This is real, current labor-market data, not speculation. What it does not establish on its own is causation specifically attributable to AI, rather than interest-rate conditions, broader hiring slowdowns, or other economic factors affecting the same period.

The evidence that complicates the story

Intellectual honesty requires including the counter-evidence, not just the alarming headline. In the same period Amodei has been warning of an entry-level collapse, IBM announced plans to triple entry-level hiring in the U.S. in 2026. Critics of Amodei’s framing, including some labor economists, have also pointed out that AI company leaders are not trained labor economists, and that predictions of this specificity, 50% of jobs, a 1-to-5-year window, are difficult to forecast reliably even for people whose actual expertise is labor markets. The honest state of the evidence is: a real and current graduate-employment weakness exists, a specific, dramatic AI-driven collapse is a widely voiced fear among AI industry leaders, and the two have not yet been rigorously, causally connected in a way independent economists broadly accept.

Why this fear matters specifically for education, regardless of how it resolves

Here is the question this debate raises for education directly, independent of whose economic prediction turns out to be right: entry-level jobs have historically been where graduates convert classroom knowledge into practical, applied skill, through repetition, feedback, and gradually increasing responsibility. If that entry point genuinely narrows, even partially, the pipeline through which people have always developed professional competence narrows with it. That is a structural question for how education prepares people, not just an employment statistic. The same fear is playing out specifically in computer science; see why CS enrollment is falling in the AI era. Our piece on the AI adoption gap in L&D covers a related, adjacent point: organizations are already unevenly prepared to reskill their existing workforce for AI, which is a harder problem to solve quickly if the traditional entry point for building that workforce is also shrinking at the same time.

What this does not mean

It does not mean pausing education, or that a degree has stopped mattering. It means the specific skills-acquisition pathway that degree was historically expected to lead into is being actively questioned by the people building the technology most likely to disrupt it, which is a legitimate reason for education, training, and L&D institutions to look seriously at building deliberate skill-development structures earlier and more explicitly, rather than assuming entry-level employment will keep functioning as the default training ground it has been for decades.

Key takeaways

  • Anthropic’s CEO and other prominent AI leaders have repeatedly warned of a possible entry-level white-collar jobs crisis within 1-5 years; this is a widely voiced industry concern, not a consensus economic forecast.
  • Recent graduate unemployment is genuinely elevated, but has not been rigorously, causally tied to AI specifically by independent labor economists.
  • Counter-evidence exists in the same period, including IBM’s announced tripling of entry-level hiring.
  • Regardless of how the employment debate resolves, it raises a real structural question: where does applied skill-building happen if the traditional entry-level pathway narrows?

Sources: Dario Amodei, multiple 2026 interviews including Axios and Fox News; World Economic Forum 2026 panel remarks (Demis Hassabis); IMF managing director public statements (January 2026); IBM 2026 hiring announcement (Bloomberg); graduate labor-market reporting, 2025-2026.

Posted on by — Leave a comment

AI Companions and the Youth Mental Health Debate: What Schools Should Know

Abstract illustration in periwinkle representing protective support layers

A U.S. health advisory has now formally warned that using AI chatbots for mental health support can have unintended effects and even cause harm. That warning did not come from a skeptic of the technology in the abstract, it came in response to a documented pattern now working its way through American courts: companion chatbots that responded to expressions of serious distress in ways that validated the crisis rather than redirecting the person toward real help.

If you or someone you know is struggling, the 988 Suicide & Crisis Lifeline (call or text 988 in the US) is available 24/7. This article discusses this topic at a factual, pattern level and does not describe specific methods or reproduce harmful chatbot messages.

What the court filings describe

In January 2026, Character.AI and Google reached settlements with five families whose children experienced severe mental health crises or died by suicide after extensive use of a companion chatbot. Additional lawsuits against OpenAI and other companies remain active in courts across California, Florida, New York, Colorado, and Texas. Reviews of these filings describe a consistent pattern, not isolated incidents: products with no clinical validation for the mental-health use cases that emerged organically among users, no escalation pathway when a user expressed crisis-level distress, and design choices that safety researchers say prioritized engagement over intervention.

What the research shows, and its real limits

A 2026 study of more than 20,000 U.S. adults found that people who used generative AI daily had roughly 30% higher odds of reporting moderate or severe depressive symptoms than people who did not use it. This is a correlational finding, not proof that AI use causes depression; it is equally consistent with people already experiencing distress turning to AI companionship more often. What the finding does establish is that the association is real and large enough to take seriously, alongside the specific, documented failure pattern in the court filings above.

Why regulators are now involved

The Federal Trade Commission has opened an investigation into how companion-chatbot companies monetize user engagement and what protections exist for children and vulnerable users specifically. Separately, lawmakers have moved toward stricter rules for chatbot platforms used by minors, including age verification requirements. This is a case where litigation, a formal health advisory, and regulatory investigation are converging on the same conclusion at roughly the same time, which is a meaningfully different situation from an emerging, contested concern.

What this actually means for schools

This is a distinct risk category from the academic-integrity and learning-outcome questions covered elsewhere on this site, and it needs to be treated as one. A school’s AI use policy, see our guide to building one, should explicitly address companion and general-purpose chatbot use outside instructional contexts, not just classroom tool adoption, because that is where this specific risk pattern has actually emerged. Practically, that means: staff and counselors should know the warning signs of a student substituting chatbot interaction for real relationships (increasing secrecy about online activity, withdrawal from friends or family, describing an AI as understanding them better than people do), and schools should have a clear, known pathway to real mental health support that does not depend on a student disclosing distress to an AI first.

The distinction worth holding onto

None of this is an argument that AI conversation is inherently harmful, or that every student using a chatbot is at risk. The documented failures involve a specific pattern: general-purpose or companion products, not clinically designed tools, used as a substitute for real crisis support, with no safeguard to catch the moment where casual use became a mental health emergency. The FTC’s own framing, engagement-optimized design without adequate protection, points at the design choice, not the technology category as a whole. That distinction matters for how schools respond: the answer is not banning AI conversation, it’s making sure a real safety net exists underneath it.

Key takeaways

  • A U.S. health advisory, active litigation across five states, and an FTC investigation are converging on the same documented pattern: companion chatbots with no crisis-escalation pathway.
  • A 2026 study found a real, statistically significant association between daily AI use and depressive symptoms; this is correlational, not proof of causation.
  • School AI policies need to explicitly cover companion and general-purpose chatbot use outside the classroom, a distinct risk category from academic tool adoption.
  • The documented failure is a design and safeguard problem, not evidence that AI conversation itself is inherently harmful.

Sources: settlement and litigation reporting via multiple U.S. legal outlets and the American Bar Association (2026); Federal Trade Commission investigation announcements (2026); peer-reported 2026 survey study on generative AI use and depressive symptoms.

Posted on by — Leave a comment

AI Note-Taking and Study Tools Compared

Abstract illustration in warm sand representing stacked study cards

Fifty-seven percent of U.S. college students now use AI in their coursework at least weekly, according to the Lumina Foundation-Gallup 2026 State of Higher Education study. But most comparisons of AI study tools make the same mistake: they rank everything on one list, as if a student stuck on a difficult concept and a student who understood the lecture but forgot it by Thursday need the same tool. They don’t. The tools split cleanly by job, not by brand.

Five different study jobs, five different tools

JobBest-fit toolsWhat to know
Understanding a concept that hasn’t clickedChatGPT Study Mode, NotebookLMStudy Mode asks guiding questions instead of giving direct answers; NotebookLM only reasons from the sources you give it, so it will faithfully study a wrong or thin set of notes
Capturing and searching lecturesOtter.aiFree tier is capped, around 300 transcript minutes per month with a 30-minute limit per session
Turning material into flashcards for retentionAnki, RemNoteSpaced-repetition scheduling is the actual mechanism behind long-term retention, not just having digital flashcards
Testing yourself before the exam doesQuizletIts Magic Notes feature converts uploaded PDFs directly into a study set
Organizing notes, sources, and deadlinesNotionGeneral-purpose organization, not an explanation or testing tool on its own

The most common mistake

Buying a flashcard subscription when the real problem is that a concept hasn’t been understood yet is one of the most frequently reported mismatches among students using these tools. Flashcards and spaced repetition are excellent at retaining something you already understand; they do very little for a concept you’re still confused about. Matching the tool to the actual stuck point, understanding versus retention versus organization, matters more than picking the highest-rated app.

NotebookLM’s real limitation, stated plainly

NotebookLM is a strong, free starting point precisely because it only reasons from the sources you upload, it won’t pull in outside information you didn’t provide. That is also its limitation: if your notes are incomplete or contain an error, NotebookLM will study that error faithfully rather than catching it. It’s a comprehension tool for the material you already have, not a fact-checker for it.

Why “Study Mode” tools matter more than raw answer-giving

Tools built specifically for studying, rather than general-purpose chatbots, increasingly default to a Socratic pattern: asking guiding questions and checking understanding before revealing an answer, instead of just producing one. This design choice connects directly to what our piece on AI and critical thinking found: unrestricted AI access during practice can reduce a learner’s later unassisted performance, while tools that require reasoning first tend to preserve it. When evaluating a study tool, whether it asks you to think before it answers is a genuinely useful thing to check, not a minor feature.

A practical starting point

Start by naming your actual stuck point honestly, rather than starting from a tool. If it’s understanding, try a Study Mode-style tool before a general chatbot. If it’s retention, spaced repetition beats re-reading. If it’s simply staying organized across a semester, that’s a search-and-structure problem, not an AI-explanation problem, and a general tool like Notion solves it better than a study-specific one would. For evaluating any of these more formally, especially at a school or program level, our framework for evaluating an AI tool covers the data-handling and reliability questions worth asking before standardizing on one.

Key takeaways

  • AI study tools split by job, understanding, capturing, retaining, testing, organizing, not by which is “best” overall.
  • The most common mistake is buying a retention tool (flashcards) for an understanding problem, or vice versa.
  • NotebookLM only reasons from your own sources; it won’t catch an error already in your notes.
  • Tools that ask guiding questions before answering better preserve the verification habit that heavy, unstructured AI use tends to erode.

Sources: Lumina Foundation-Gallup, 2026 State of Higher Education study; synthesis of independent AI study tool comparisons published across 2026. Individual pricing and features change frequently and are best verified directly with each vendor.

Posted on by — Leave a comment

Multimodal AI Explained (Without the Jargon)

Abstract illustration in cobalt blue representing overlapping AI modalities

A multimodal AI model can look at a photo of a whiteboard, understand a diagram on it, write code that implements what the diagram shows, and explain that code back to you in plain language, all in one conversation. By 2026, this kind of cross-modal understanding has become the standard for frontier AI models, not a special feature. For teaching and training, the practical shift is simple: you can now hand a model a document, an image, or a recording directly, instead of manually converting everything to text first.

What “multimodal” actually means

A single-modality AI model handles one kind of input and output, usually text: you type a question, it types back an answer. A multimodal model processes and connects several types of content, text, images, audio, and increasingly video, within one system. The key technical difference is integration: the model doesn’t just run a separate image tool and a separate text tool side by side, it reasons across them together, which is what lets it explain how a diagram and a paragraph of text relate to each other, not just describe each one in isolation.

What this actually enables for teaching and training

  • Document understanding. Upload a scanned worksheet, a handwritten problem set, or a slide deck directly, no separate OCR or transcription step required first.
  • Diagram and image feedback. A student can photograph a hand-drawn diagram, chart, or lab setup and get feedback on the content itself, not just a text description of what’s in the photo.
  • Voice-based interaction. Natural, real-time voice conversation with an AI assistant, useful for language practice or for learners who process speech more easily than text.
  • Recorded lecture and meeting analysis. Audio or video of a class session or training module can be processed directly for summarization or Q&A, connecting to the practical comparison in our AI writing assistants piece, since several tools in that space now combine transcription with multimodal understanding rather than offering transcription alone.

What it does not solve

Multimodal capability is an input and output improvement, not a reasoning or accuracy improvement on its own. A model that can now read a photograph of a math problem is not automatically more reliable at solving that problem correctly than it was before; the underlying accuracy and hallucination risks discussed in our piece on AI and critical thinking apply just as much to multimodal output as to text-only output. Treat a multimodal tool’s confidence about what it “sees” in an image with the same verification habit you’d apply to anything else it generates.

A practical way to think about it

The useful question isn’t “is this tool multimodal” by 2026, most capable tools are. The useful question is which specific modality combination actually matches your task: text-plus-image for marking handwritten work, text-plus-audio for transcribing and summarizing a session, or text-plus-document for working directly with PDFs and slides instead of retyping their content. Matching the modality to the task, rather than defaulting to typing everything out as plain text, is where the actual time savings come from.

Key takeaways

  • Multimodal AI integrates text, image, audio, and video understanding within one model, rather than chaining separate single-purpose tools together.
  • By 2026 this is the baseline for frontier models, not a differentiator between them.
  • Practical education and training uses include document understanding, diagram feedback, voice interaction, and recorded-session analysis.
  • Multimodal capability improves what a model can take as input, not its underlying accuracy, so the same verification habits still apply.

Based on a synthesis of 2026 technical overviews of multimodal AI systems and their applications; specific model capabilities change frequently and are best verified directly with each vendor.

Posted on by — 3 Comments

AI Literacy: What It Actually Means (and Why the Definition Keeps Changing)

Abstract illustration in periwinkle representing layered AI literacy frameworks

AI literacy does not have one settled definition, and that is not a flaw in the field, it reflects a genuinely fast-moving policy landscape. In June 2026, the OECD and the European Commission jointly published AILit (Empowering Learners for the Age of AI), a new international reference framework for primary and secondary education, built on top of, not replacing, earlier work like UNESCO’s 2024 AI Competency Framework for Teachers. Understanding why there are multiple frameworks, and what they actually agree on, is more useful than memorizing any single one.

Why there isn’t just one framework

FrameworkWho it’s forWhat it covers
UNESCO AI Competency Framework for Teachers (2024)Educators specificallyWhat teachers need to know and be able to do to use AI responsibly in their own practice
AILit / Empowering Learners for the Age of AI (OECD + European Commission, June 2026)Primary and secondary studentsFour domains, Engage, Create, Manage, Shape AI, across 19 competences and three learner progression levels
Stanford Teaching Commons frameworkHigher-education instructorsFour domains: functional, ethical, rhetorical, and pedagogical AI literacy

These aren’t competing definitions so much as different frameworks built for different audiences and grade bands, and they deliberately build on shared foundations: AILit explicitly draws on AI4K12, UNESCO’s earlier work, and the EU’s DigComp 3.0 digital competence framework rather than starting over. The practical takeaway is that the underlying skills are converging even as the specific documents multiply.

What these frameworks actually agree on

Strip away the differing terminology, and a consistent core shows up across nearly every framework in this space: understanding at a basic level how an AI system works and where its limits are, using AI tools effectively for a real task, evaluating AI output critically rather than accepting it, and reasoning about the ethical and social implications of AI use. AILit’s own draft definition states this plainly: AI literacy provides a clear understanding of how AI technologies work and what their responsible use entails, so learners can make informed decisions, grounded in trusted information, ethics, and social good.

Why this is becoming a legal requirement, not just good pedagogy

AI literacy has moved beyond curriculum recommendation in at least one major jurisdiction. Article 4 of the EU AI Act, applicable since February 2025, requires that anyone deploying an AI system ensures the people using it, including students and educators, have a sufficient level of AI literacy. AILit is explicitly designed to help schools meet that requirement in practice, and it is also linked forward to the PISA 2029 Media and AI Literacy (MAIL) assessment, meaning AI literacy is on a path toward being measured internationally, not just taught inconsistently.

What this means for teaching it in practice

Rather than waiting for a single finalized standard, or trying to teach every framework’s full competence list, the practical starting point is the four-part core above: how AI works at a basic level, how to use it for a real task, how to evaluate its output, and how to reason about its implications. This lines up directly with how we’ve covered AI and critical thinking on this site: the evaluation and verification piece of AI literacy is exactly the skill that research shows is most at risk when AI use isn’t structured deliberately. It also connects to building an AI use policy, since a policy without a literacy component to back it up is difficult to actually implement in a classroom.

Key takeaways

  • Multiple AI literacy frameworks exist because they serve different audiences (teachers vs. students vs. higher-ed instructors), not because the field disagrees on fundamentals.
  • A consistent core spans nearly all of them: how AI works, using it effectively, evaluating its output, and reasoning about its implications.
  • The EU AI Act’s Article 4 makes AI literacy a legal requirement in at least one major jurisdiction, not just a pedagogical nice-to-have.
  • AILit is explicitly linked to the PISA 2029 MAIL assessment, meaning AI literacy is moving toward being measured, not just taught inconsistently.

Sources: OECD and European Commission, AILit / Empowering Learners for the Age of AI (June 2026); Miao, F., & Cukurova, M. (2024). AI Competency Framework for Teachers. UNESCO; Stanford Teaching Commons, Understanding AI Literacy.

Posted on by — 3 Comments

Building an AI Use Policy for Your School or Training Program

Abstract illustration in periwinkle representing tiered AI policy guidance

A usable AI policy does not start from a blank page or a generic template. It starts from a simple structural idea that UNESCO’s own guidance uses: not all AI use in a school carries the same level of risk, so a policy should not treat it as if it does. Administrative AI, teacher-support AI, and student-facing AI each need different rules, different oversight, and different levels of caution, and conflating them is the most common reason school AI policies end up either too vague to use or too restrictive to survive contact with an actual classroom.

Start with UNESCO’s three-tier structure

UNESCO’s Guidance for Generative AI in Education and Research groups AI applications into three confidence tiers, and building a policy around this structure, rather than one flat set of rules, is the single most useful starting decision a school can make.

TierExamplesUNESCO’s confidence level
Administrator-facingScheduling, predictive analytics, language access toolsHighest current confidence
Teacher-supportingLesson planning assistance, formative assessment help, grading time-savingsMedium-high confidence
Student-facing generative AIStudents directly using chatbots or generative tools for courseworkLowest confidence; UNESCO explicitly flags this as too new for stable, settled policy

Practically, this means your policy should be least restrictive and most settled for administrative tools, moderately structured for teacher use, and most cautious and most subject to revision for student-facing tools. A single blanket rule across all three (either “AI is banned” or “AI is allowed”) ignores a distinction UNESCO itself considers foundational.

The four principles underneath the tiers

Beneath the three tiers, the OECD’s AI Principles supply the operating rules that should govern any AI use in any tier: human-in-the-loop on consequential decisions, transparency to the people the AI affects (students, parents, staff), explainability of what the AI actually did, and clear accountability for whoever deployed it. These four principles are what should actually appear as concrete policy language, tier by tier, rather than staying abstract.

For example, applied to a teacher-supporting grading tool: human-in-the-loop means the teacher, not the tool, assigns the final grade. Transparency means students and parents know AI assisted the process. Explainability means the teacher can say what the tool flagged and why. Accountability means the school, not the vendor, owns the outcome.

Why this is becoming a legal requirement, not just good practice

This is no longer purely a best-practice recommendation in every jurisdiction. North Carolina’s 2026 state law, for instance, requires the state Department of Public Instruction to adopt a model AI policy by the end of 2026, and requires every local school district, charter school, and affiliated school system to adopt their own policy based on it by mid-2027. The EU AI Act goes further for systems used in assessment and admissions specifically, classifying them as high-risk and requiring formal risk management, data governance, human oversight, and conformity assessment before deployment, with phased enforcement running through 2026-2027. Even where no law yet mandates a policy, this direction is clearly where the regulatory environment is heading, which makes building one proactively, on a sound structure, more useful than waiting to be required to.

What a usable policy document actually contains

Real school AI policies that are actually in use, rather than aspirational documents, tend to share a consistent structure: a stated purpose, a defined scope (which staff, students, and visitors it covers), the legal framework it operates under, clear definitions of what counts as AI use, defined roles and responsibilities for oversight, a section on appropriate use, a section on misuse and its consequences, an implementation plan, and a scheduled review mechanism, typically annual, given how quickly the underlying tools and regulations are changing.

That review cadence matters as much as the initial content. A policy written once and left untouched will be describing tools and rules that no longer match reality within a year, given the pace covered in our explainer on AI agents and elsewhere on this site.

Before approving any specific tool under the policy

A policy sets the rules; it does not evaluate individual tools against them. For that step, use a structured process rather than a vendor’s own claims, our framework for evaluating an AI tool covers exactly the data-handling and oversight questions the OECD principles above require you to be able to answer before approving a tool for classroom use. A sound policy also depends on a literacy component; see AI literacy: what it actually means for the skills a policy alone cannot teach. For a live 2026 case of what happens when assessment rules do not match AI use, see the Brown University AI cheating case.

Key takeaways

  • Structure your policy around UNESCO’s three confidence tiers, administrator, teacher-support, and student-facing, rather than one flat rule for all AI use.
  • Apply the OECD’s four principles, human-in-the-loop, transparency, explainability, accountability, as concrete language within each tier, not as abstract values.
  • This is increasingly a legal requirement, not just best practice; North Carolina and the EU AI Act are early examples of a broader regulatory direction.
  • Build in an annual review cycle; a static policy will fall out of date faster than the tools it governs.

Sources: UNESCO, Guidance for Generative AI in Education and Research (2023, updated 2024); OECD AI Principles (2019, updated 2024); EU AI Act (2024); North Carolina General Statutes ยง115C-102.13 (2026).

Posted on by — Leave a comment

AI for Corporate Onboarding: A Realistic Starting Point

Abstract illustration in deep navy representing an onboarding pathway

Organizations using AI-assisted onboarding cut new-hire time-to-full-productivity by 25 to 40 percent compared to manual programs, according to Brandon Hall Group research. And yet only 32 percent of HR teams had deployed AI specifically for onboarding workflows as of 2025, per SHRM, even though AI adoption across other HR functions had already reached 43 percent. Onboarding is lagging behind the rest of HR’s AI adoption, which makes it one of the more realistic places for a training team to start, precisely because the bar most competitors have cleared is still low.

What the data actually shows

  • 25-40% reduction in time-to-full-productivity with AI-assisted onboarding (Brandon Hall Group)
  • Up to 82% improvement in new-hire retention at companies with strong AI onboarding programs (Brandon Hall Group)
  • Only 32% of HR teams used AI specifically for onboarding as of 2025, versus 43% AI adoption across HR generally (SHRM)
  • 87% of L&D teams already use AI in some form for training broadly, but structured onboarding-specific adoption lags well behind that figure (Synthesia, 2026)

Read together, these numbers describe a specific opportunity: onboarding is one of the clearest ROI cases for AI in L&D, and it is also one of the least adopted specifically. That gap between demonstrated impact and actual deployment is the practical starting point for this guide, and it echoes the broader pattern we cover in the AI adoption gap in L&D: enthusiasm and even proven ROI do not automatically translate into structured rollout.

Where AI actually helps in onboarding, specifically

The productivity and retention gains above are not coming from a single dramatic use case. They come from removing friction across several unglamorous, high-volume tasks that determine whether a new hire’s first weeks feel organized or chaotic:

  • Document processing and task routing. Paperwork, account provisioning, and scheduling are exactly the kind of repetitive, rule-based work AI handles reliably, and where HR staff time is most easily freed up.
  • Personalized training sequencing. Rather than every new hire receiving the identical onboarding path regardless of role or prior experience, AI-assisted systems can sequence material based on the specific role and what the person already knows.
  • First-line question answering. A new hire’s early questions are usually repetitive across cohorts (where is X, how do I request Y); AI-assisted resources absorb this volume without pulling a manager or HR staffer away from higher-value work.

None of these replace the parts of onboarding that actually build retention, a manager’s first real conversation with a new hire, a team’s welcome, the sense that someone is paying attention. The data on retention gains reflects organizations using AI to clear the administrative noise out of the way of those moments, not to replace them.

A realistic starting point, not a full rebuild

Given that most organizations have not yet deployed AI specifically for onboarding, the realistic starting point is narrow, not comprehensive:

  • Pick one high-volume, low-risk task first. Document collection or first-week FAQ handling are good starting points precisely because a mistake is inconvenient, not consequential.
  • Keep a human checkpoint on anything that reaches the new hire directly. The same principle applies here as with any AI tool that produces multi-step output; see our explainer on AI agents for why the final review step matters more, not less, as a tool automates more of the process.
  • Measure before and after on the same two metrics Brandon Hall Group tracks: time-to-productivity and 90-day retention. These are the two numbers actually tied to the ROI case above, not vaguer satisfaction scores.
  • Expand only after the first narrow use case shows a real result. Given how few organizations have gone past pilot stage in this specific area, there is no competitive pressure to move faster than your own data supports.

Before evaluating a specific onboarding tool

Whatever specific platform or workflow you consider, run it through a structured evaluation rather than a vendor demo alone. Our framework for evaluating an AI tool covers the questions that matter most for something touching new-hire personal data specifically: what happens to that data, what the human review checkpoint looks like, and what your exit plan is if the vendor relationship changes.

Key takeaways

  • AI-assisted onboarding shows a real, well-documented ROI case: 25-40% faster time-to-productivity and up to 82% better retention (Brandon Hall Group).
  • Adoption specifically for onboarding still lags behind AI adoption in HR generally, which makes it a realistic, low-competitive-pressure place to start.
  • The actual gains come from removing administrative friction, not replacing the human moments that build retention.
  • Start with one narrow, low-risk task, measure against time-to-productivity and retention specifically, and expand only after it works.

Sources: Brandon Hall Group; SHRM (2025); Synthesia, AI in Learning & Development Report 2026.

Posted on by — Leave a comment

AI Writing Assistants Compared: Which One Fits Your Workflow

Abstract illustration in warm sand representing AI writing tools comparison

Independent testing across a dozen review sites in 2026 converges on the same basic split: ChatGPT is the most versatile all-purpose drafting tool, Claude consistently produces the most natural long-form prose, Grammarly is an editor rather than a generator, and Jasper is built for marketing teams managing brand voice across multiple writers. None of them wins every job. The right choice depends on which specific writing task you are actually trying to solve.

Four tools, four different jobs

ToolWhat it actually does wellBest fit for education and training
ChatGPTFast, versatile drafting and brainstorming; handles a wide range of formats and adapts quickly to instructionsGeneral lesson materials, quick first drafts of training content, rapid brainstorming
ClaudeThe most natural, least mechanical-sounding long-form prose of the major tools, and reliably follows detailed, multi-part instructions without driftingLonger explanatory content, training modules requiring a consistent voice, anything you will publish largely as-is
GrammarlyInline editing wherever you already write, not a generator; catches errors and polishes tone in placeA layer on top of whatever drafts a teacher, trainer, or student is already producing, not a replacement for one
JasperBrand-voice consistency across multiple writers, with guardrails built from uploaded style examplesTraining departments or districts producing communications across many authors who need to sound like one organization

The generator vs. editor distinction matters more than the brand name

The most common mistake in choosing an AI writing tool is treating all four as competitors for the same job. They are not. ChatGPT and Claude are generative tools, they draft from scratch. Grammarly is an editing layer, it improves what you or a generative tool already produced. Multiple independent 2026 reviews land on the same combination for most writers: draft with ChatGPT or Claude, then polish with Grammarly, rather than picking one tool to do both jobs. For a teacher writing parent communications or a trainer producing a module script, that two-step workflow, not a single subscription, is usually the actual answer.

Where Claude specifically stands out

For content that will be published with light editing, rather than heavily reworked, the natural-prose quality gap matters. Reviewers across several independent 2026 comparisons consistently rate Claude ahead of ChatGPT specifically on long-form coherence, essay-style writing, and following complex, multi-constraint instructions without drifting off them partway through. For a training module or a longer explanatory article, that reliability on complex instructions is often worth more than raw drafting speed.

Where ChatGPT specifically stands out

ChatGPT’s advantage is breadth, not prose quality. It handles a wider range of formats and auxiliary tasks (structuring an outline, adjusting tone quickly across several revisions, working alongside images or code) inside one interface. For fast-turnaround work, quiz questions, a first-pass outline, a batch of short prompts for a training exercise, that versatility is often more useful than Claude’s edge in long-form polish.

A practical decision framework

  • Writing something you will publish close to as-is? Start with Claude, the editing burden is typically lower.
  • Need fast drafts, outlines, or quick iteration across many small tasks? ChatGPT’s speed and versatility fits better.
  • Already have drafts, from a person or another AI, and need them polished? Grammarly is the layer for that, not a replacement generator.
  • Producing content across multiple staff who need to sound consistent? Jasper’s brand-voice guardrails solve a coordination problem the other three do not address.

This framework matters more than picking a single winner, because the tools genuinely do different jobs. It also connects directly to a question worth asking before adopting any of them operationally: see our framework for evaluating an AI tool for how to check data handling and pedagogical fit before rolling one out across a classroom or training team, not just which tool writes the best paragraph.

A caution worth stating plainly

None of this changes the verification point we cover in our piece on AI and critical thinking: a fluent draft is not the same as an accurate or appropriate one. Whichever tool produces the first draft, the review step, checking facts, tone, and fit for the actual audience, still has to happen. These tools are fast at producing polished-looking text; polish is not the same as correctness.

Key takeaways

  • ChatGPT, Claude, Grammarly, and Jasper solve different problems, not the same problem at different quality levels.
  • Claude generally produces the most natural long-form prose; ChatGPT generally wins on speed and format versatility.
  • Grammarly is an editing layer, not a substitute for a generative drafting tool.
  • The most common effective workflow combines a generator with an editor, rather than relying on one tool for both jobs.

For the student-focused equivalent of this comparison, see AI note-taking and study tools compared.

Based on a synthesis of independent AI writing tool comparisons published across 2026; individual pricing and feature details change frequently and are best verified directly with each vendor before adopting a tool operationally.