The same week's research documents both AI-driven grade inflation and AI-driven learning gains — the difference, in both cases, is instructional design.
Today's Signals at a Glance
01Friday Classroom Signal—Foreign Language: AI tools reached language classrooms before AI training reached language teachers. Frontiers research pinpoints the gap.Lang
02UC Berkeley traced a 30% spike in A grades directly to AI-assisted take-home writing and coding. The inflation is concentrated exactly where you'd expect.AI / EdTech
03Google ran an 8-week randomized controlled trial on AI math tutoring in Sierra Leone. The learning gains were real — and they required more than deploying the tool.Pedagogy
04Columbia's Melissa DuPont-Reyes let 41 Latino teens co-author research on their own social media use. The data complicates both the harm narrative and the resilience narrative.Youth Culture
Classroom Signal — Friday · Foreign Language
Foreign Language
AI Tools Reached Language Classrooms Before AI Training Reached Language Teachers.
Ren and Li published a study in Frontiers in Education this year assessing AI literacy in foreign language teachers using a TPACK-based framework — combining technological, pedagogical, and content knowledge. The core finding: generic AI literacy assessments miss the specific demands of language instruction. The cognitive and linguistic work of teaching conversation, pronunciation, authentic register, and cultural context is different from teaching other subjects, and AI tools engage all of those dimensions differently than a writing-assistance tool in an ELA class.
Teachers are at an early integration stage. AI is reducing administrative workload, but it has not yet changed how language is actually taught. The separate finding from Lu in Foreign Language Annals this year names AI literacy as a core world language practice — not a technology add-on. The teachers who build that literacy now, before curriculum requirements formalize it, will define what it looks like in their departments.
Try This — Ready to Use
Have students use an AI tool to generate a short spoken response in the target language — two or three sentences on a prompt you assign. Then, as a class, critique the AI's output: Is the register appropriate for the context? Would a native speaker in a specific region say it this way? What's missing culturally? Students who can identify what the AI got wrong about authentic language use have demonstrated deeper knowledge than a student who conjugated a verb correctly on a worksheet. The AI becomes the object of analysis, not the replacement for thinking.
Try This in Any Class — Today
Before students begin any task — writing, problem-solving, translation, lab work — ask them to state aloud or write in one sentence what they already know about this topic and what they're genuinely unsure about. Two minutes. It forces retrieval and surfaces genuine prior knowledge rather than AI-assisted performance. What they name as uncertain is your actual learning target for the day. No device required, no setup, no grading.
Signal Analysis
SIGNAL 01—AI / EdTech
A 30% Spike in A Grades Since ChatGPT. Berkeley Traced Where the Inflation Landed — and What It Cost.
The Development
UC Berkeley researcher Igor Chirikov published findings in Science on May 21, 2026, analyzing more than 500,000 student-course enrollments across 84 departments at a large Texas university from 2018 to 2025. The finding: courses exposed to AI tools saw a 30% increase in "A" grades since ChatGPT's release, concentrated in courses relying heavily on take-home writing and coding assignments. A companion paper, the largest study of undergraduate AI use yet conducted, covered 95,000 students across 20 research-intensive universities and documented broad patterns of AI-assisted cheating. Both studies were co-authored with researchers from the University of Technology Sydney and Cornell University and published simultaneously in Science.
Why It Matters to You
The grade inflation is not distributed evenly. It is highest in courses where take-home essays and coding assignments carry the most weight and where the work happens out of the teacher's view. In-class work, oral assessments, and process-visible assignments did not show the same inflation pattern. That is not a coincidence — it is a precise description of which assignment structures allow substitution and which do not. The Chirikov data gives teachers a specific, evidence-based answer to the question of which assignments need redesign: the ones producing grades that no longer require the student to have done the thinking.
Why This Matters
Grade inflation data is often abstract. This study is not. It identifies the specific assignment type producing the inflation and the specific mechanism causing it. That makes it a design guide, not just a warning.
Around the Corner
The companion Science paper on 95,000 students will be cited in every academic integrity debate for the next two years. Districts that haven't addressed AI use in take-home writing and coding specifically — only in general academic integrity policies — are not addressing the problem the research identifies. That distinction will matter when the data reaches school board conversations this fall.
Google Ran an 8-Week AI Math Trial. The Gains Were Real — and They Required More Than Deploying the Tool.
The Development
Google published results from an eight-week pre-registered randomized controlled trial conducted with Fab AI and local teachers in Sierra Leone. The trial assigned 48 mathematics classrooms — approximately 1,800 Grade 7 and 8 students — to either use Guided Learning via Gemini or continue with regular instruction. Students using the AI tool improved their scores on externally validated assessments by +0.26 standard deviations, equivalent to 1.2 to 1.7 years of typical learning progress in comparable contexts. Students who engaged with the tool for at least 12 hours over the 8-week period saw larger gains, moving from the 50th to the 64th percentile. The gains did not appear for low-engagement students.
Why It Matters to You
This is the counterweight to the Stanford SCALE finding from two weeks ago that only 20 out of 800+ AI-in-education studies meet causal standards. This study meets those standards. The gains were real, measured by external assessment, and attributable to the tool rather than correlated with it. The variable that determined whether the gains materialized was not which tool was used — it was how much time students spent actively using it. Passive AI exposure produced nothing. Twelve or more hours of active engagement produced results equivalent to nearly two additional years of learning. The instructional implication is specific: AI tools produce learning when they are embedded in structured, time-intensive practice, not when they are available and occasionally used.
Why This Matters
Most US schools deploying AI tools are measuring adoption, not time-on-task. The Sierra Leone data argues that time-on-task is the number that predicts learning gains. That is a different accountability question than "did the teacher use the tool?"
Around the Corner
The Sierra Leone context matters for reading this study carefully. Students there had access to fewer instructional alternatives, which may have increased both engagement and the relative benefit of the tool. Direct transfer to well-resourced US classrooms requires caution. But the core finding — that active, sustained engagement with AI produces measurable learning and passive access does not — holds across contexts and is consistent with how effective instruction has always worked, with or without technology.
Columbia Let 41 Latino Teens Co-Author Research on Social Media and Mental Health. The Data Complicates Both Narratives.
The Development
Melissa DuPont-Reyes, assistant professor of epidemiology and sociomedical sciences at Columbia University Mailman School of Public Health, published SocialsVoice this month — a book co-authored with 41 Latino youth ages 13 to 24 and 28 of their parents. The methodology was deliberate: youth participants first defined what they considered positive and negative mental health content, were then randomly assigned to groups exposed to their respective category, and spent seven video-chat sessions discussing their assigned content. They concluded by co-creating videos of their own research findings for peers and parents. The book draws on the resulting data and participant-generated content. The June 2026 release was covered by EurekAlert and Medical Xpress.
Why It Matters to You
The standard social media and mental health narrative runs in one direction: harm. The SocialsVoice data is more precise. Youth identified rampant stigmatizing content — posts claiming mental illness isn't real, content minimizing depression, posts reinforcing toxic masculinity — alongside a youth-led anti-stigma movement that is substantive, not marginal. Mental health education, symptom management, suicide awareness, and self-care content is present and actively consumed by teens on the same platforms causing harm. Most teachers working with majority-Latino student populations have had very little research specific to how their students experience social media. This is among the first systematic studies with this demographic, and the result is not a simple harm story.
Why This Matters
The anti-stigma movement online is youth-led. Some of your students are likely already participating in it. A classroom conversation that treats social media as a one-dimensional harm misses what the data actually shows — and misses an entry point for genuine discussion.
Around the Corner
DuPont-Reyes's methodology — youth participants defining the research categories, then generating the findings themselves — is a model for classroom-based digital citizenship inquiry. Students can replicate this process at the school level. Assigning students to document and categorize the mental health content they encounter in a week, then analyzing it as a class, produces a very different kind of digital literacy than a lecture about screen time.
The Bottom Line—Three Things for a High-Agency Professional
1If any take-home writing or coding assignments in your class can be completed by AI without your knowledge, the Chirikov data applies to your classroom. Not as a problem to detect — as a design prompt. The assignments where substitution is easiest are the ones most worth redesigning before fall.
2The Sierra Leone RCT shows AI produces learning gains when students engage actively for sustained time. If your district deploys AI tools and sees weak results, the question to ask is time-on-task, not which tool was chosen. Passive access and active use are not the same intervention.
3Before the next time you address social media with your students, read the SocialsVoice summary at Medical Xpress. Forty-one teenagers already analyzed the problem. Their finding — that stigmatizing content and a youth-led counter-movement coexist on the same platforms — is a more accurate starting point than a lecture about screen time.