Join Bellwether’s Artificial Intelligence (AI) Cohort!

Bellwether is convening a year-long, virtual learning and action network for leaders of education and youth-serving organizations across the country. Participants will build practical skills to evaluate and responsibly use AI, lead change with confidence, learn from peers, and leave with an actionable AI roadmap aligned to their mission. The cohort experience will include: 10 monthly 90-minute strategy and skill-building sessions; monthly office hours with Bellwether advisers; guest experts; and curated tools. Organization leaders who are committed to move from experimentation to strategy and organizationwide adoption are encouraged to apply.

Learn more here.

Over the last year, the major AI companies (OpenAI, Anthropic, Google, etc.) have released several reports analyzing their user data to map trends about how people are accessing generative AI. We’re taking their conclusions with a grain of salt and so should you; each of the reports so far have methodological flaws that should give discerning readers pause.

• In February 2025, OpenAI analyzed college students’ adoption of ChatGPT across the country to find that “more than one-third of college-aged young adults in the U.S. use ChatGPT.”  
  • However: The methodology section of the report is very light, and draws on usage in January 2025 combined with a third-party survey (Conducted by what organization? With what questions? Analyzed how? We don’t know!). Moreover, OpenAI identified “college users” as just individuals between ages 18 and 24 — which doesn’t necessarily mean that they are in college.
• In April 2025, Anthropic released its Education Report, which analyzed how college students were using Claude and found 1) that STEM students were early adopters, compared to business- or humanities-focused students, and 2) the top three AI use cases for students included generating educational content, asking for technical solutions (including cheating), and conducting data analysis.
  • However: Anthropic identified college students as any account with a linked email address that ended in “.edu,” but in the authors’ own words, they may have included staff or faculty members. Additionally, the chats studied spanned only an 18-day window due to Anthropic’s privacy policies, which creates a much smaller pool of data that may or may not be generalizable.
• Last month, OpenAI released another report looking at the usage patterns across all users (not just students), and claimed that 10% of chats were learning-related — specifically “tutoring or teaching”-related.
  • However: OpenAI defined “tutoring or teaching” as “explaining concepts, teaching subjects, or helping the user understand educational material.” For anyone who has ever tutored or taught, that definition seems fairly reductive and definitely doesn’t encompass all that goes into learning. (A natural follow-up to that is, “What does go into learning?” For some more nuanced thinking on that piece, check out Bellwether’s Productive Struggle report if you haven’t already.) Additionally, some of the other functions OpenAI identified but did not include as part of “learning” are also, well … part of learning. For example, two completely separate categories were “argument or summary generation” and “creative ideation” — both of which many will recognize as key to digesting new concepts.

To be clear: Our issue isn’t with the research per se, or who is conducting it, but with how these findings are being presented. Many of these claims are being repeated in other forums as Truths or Absolutes, when in fact there are assumptions made behind the scenes that may or may not be generalizable. The sector should be careful when consuming these reports and read them with a critical eye that asks more not just of the researchers, but also of our own understanding.

And it’s not just tech companies falling into this trap: Earlier this year, an MIT paper made headlines with its finding that “[large language model or LLM] users consistently underperformed at neural, linguistic, and behavioral levels.” However, as this critique outlines, there were a lot of methodological choices in the experiment design that may not be generalizable (and are definitely not headline-worthy).

Usage research is crucial to a deeper understanding of how AI is being integrated in our society, and major tech companies should absolutely play a role in conducting and disseminating this research given their overall influence and positions as product developers. But there’s still a gap between how AI companies view learning and how education researchers measure outcomes. Bridging this gap with thoughtful partnerships could be key to unlocking better insights and a clearer picture of how AI is showing up across education and learning.

Education Evolution: AI and Learning

In our last issue, we covered tragic instances where AI may have played a role in children being harmed, as well as the responses (and lawsuits) that followed. Since then, and given the heightened scrutiny of AI chatbots’ interactions with children, OpenAI announced new protections for kids. One guardrail is “age prediction” to conversations to detect if the user is a minor; if so, that user will be directed to a “GPT for Teens” experience. While this feature was clearly intended for child safety, an AI experience purpose-built for teenagers could naturally lend itself to wider adoption in the classroom and beyond. Just saying.

In other news:

• A new Bellwether report highlights how educators, developers, and funders can create outcomes-focused measures to understand AI’s true impact in classrooms.
• Data from RAND indicates that middle and high school students using AI for their schoolwork grew from 37% in 2024 to 54% in early 2025. Educators’ use of AI for instructional planning or teaching more than doubled from 24% last year to 53% as of spring 2025. 
• Stanford University Human-Centered Artificial Intelligence released a policy brief that illuminates another danger of AI to students — deepfakes depicting minors nude — and suggests steps schools can take to better prepare for these situations.
• Robin Lake has a great round-up of the summer’s push for educator-focused AI literacy, as well as thoughtful questions about how effective these initiatives can and will be.
• A curriculum design specialist worked with a teacher to create an AI-powered chemistry lesson that exemplifies how AI could incorporate productive struggle (or what he calls “cognitive investment”) for real learning.
• A recent EdTech Insiders podcast episode discussed what differentiates smaller, more education–focused shops from the big players like OpenAI or Google, and whether that will be enough to survive as ed tech-focused venture capital declines.
• The State Educational Technology Directors Association and Whiteboard Advisors 2025 State EdTech Trends report finds that AI is (perhaps unsurprisingly) states’ top ed tech priority.

The Latest: AI Sector Updates

OpenAI announced the release of Sora 2, their latest video and audio generation model, which will power a new standalone social app for iPhone called “Sora.” Two days after launch, the app hit the No. 1 spot on the App Store despite invite-only access (for now). The app is already pushing up against (and maybe violating) copyright laws: Users can create videos with their favorite characters, such as this video depicting Pikachu taking part in the D-Day landing. 

You may already be tired of AI-generated images flooding your social media feeds — now there’s a new social media app that contains only AI-generated content. X (née Twitter) was a perfect engine of negative attention generation, leading millions to the habit of “doomscrolling.” Sora offers an alternative: less outrage, more oblivion. It’s a digital version of Aldous Huxley’s “Brave New World” soma, offering a never-ending stream of meaningless video memes coaxing users to “hollow-scroll” their day away. At least the builders of the Tower of Babel were trying to reach Heaven; Sora’s creators seem content to have us strive for the end of the feed. This is a good time to ask: What are we trying to build with AI tools? The answer matters now more than ever.

In other news:

• Anticipating the rise of an era where AI agents will do all the shopping, Google is proactively working to ensure secure and trustworthy “agent-led” payments.

• Top platforms are converging on a multi-provider approach — intensifying competition among the major LLM providers:
  • Despite its ties to OpenAI, Microsoft recently added the option for Copilot users to switch between ChatGPT and Anthropic’s Claude models; the move to diversify could signal greater competition among AI providers in the future.
  • Apple has reportedly tapped Google’s Gemini to improve Apple Intelligence for a more multimodal Siri, despite already having integrated ChatGPT.

Pioneering Policy: AI Governance, Regulation, Guidance, and More

Voters nationwide are paying attention: A recent Gallup-Special Competitive Stories Project poll shows that 98% of Americans say they’ve heard about AI in the past year, yet trust and use in the technology remain low. At the same time, 87% of Americans expect foreign powers to deploy AI as a threat, and 72% support policies expanding AI training and education. These numbers suggest both anxiety and opportunity. For state policymakers, the tension between fear and curiosity will shape the next wave of AI regulation.

Across the country, statehouses are learning that grand, universal “AI bills” tend to stall while narrower, risk-specific efforts — like disclosure rules for chatbots, algorithmic hiring, or health care decisions — move. Disclosure requirements have become the de facto entry point: a low-cost way to build public visibility before layering on stricter obligations. Procurement is quietly emerging as another lever. Even when legislatures deadlock, states can use contracts to require vendors to explain, audit, and test high-risk AI systems used in public services or education.

However, the fine print matters. Definitions of “AI,” “generative,” and “automated decision-making” are diverging across jurisdictions, creating friction for companies and agencies alike. States that harmonize terminology and pair rules with support — particularly around training, educator capacity, and workforce development — may find themselves better positioned to build both trust and capability. The throughline from the Gallup data is clear: The public wants AI safety, transparency, and skills, in roughly that order. States that balance all three will be setting the tone for national AI policy conversations around the corner.

In other policy developments:

• California’s Leading Ethical AI Development (LEAD) for Kids Act was significantly narrowed from its original scope to primarily ban companion chatbots for minors unless they meet safety criteria, reflecting state-level hesitation to implement comprehensive AI regulations without federal guidance.
• After a 2024 ban, the U.S. House of Representatives is now allowing Microsoft Copilot for select staff use and is testing additional AI tools for potential approval.
• Anthropic is offering to sell Claude to the government for $1, while simultaneously increasing its Washington, D.C. presence to influence AI policy discussions.
• Later this month, the federal Office of Science and Technology Policy will seek public and business input on regulations potentially hindering AI development and deployment.