Free AI training from the vendor who grades it: what OpenAI Academy's role-based paths teach, and what they leave out

Two announcements about who learns AI, and how

On September 21, 2026, OpenAI announced an expansion of OpenAI Academy, its free online training platform, with role-based learning paths: Apply AI at Work for knowledge workers, Build with AI for developers, Lead AI Adoption for leaders, AI for Educators, and AI for College Students. The company also added course assessments that let learners earn an OpenAI Academy course badge and published a Deployment Guide for organizations that want to roll the courses out across teams.

Two days later, on September 23, OpenAI and Grab announced GO Forward with AI, a two-year program that aims to help 30,000 of Grab's driver-, delivery-, and merchant-partners across Southeast Asia build practical AI skills. It starts in Singapore and is planned to expand to Thailand, Indonesia, and the Philippines later this year, then Malaysia and Vietnam in 2027. The workshops are based on the OpenAI Academy curriculum and adapted to local needs, with GrabAcademy trainers trained to extend the program over time.

Taken together, these announcements describe a serious investment in free, vendor-provided AI training. On the surface, that looks like a straightforward win for access. Anyone with an internet connection can study the same curriculum that large organizations deploy to their staff, at no cost. But vendor-run training also sets the terms of what counts as competent and safe use of AI, and what does not. This article examines what the Academy actually teaches, who it is for, and what it leaves out.

What the curriculum actually teaches

The Apply AI at Work pathway is the most broadly relevant of the new paths. According to OpenAI's announcement, it teaches learners to give clear instructions, add relevant context, and review AI responses against the task at hand. It then moves to building reusable workflows and, at the advanced end, directing larger pieces of work with agents: deciding what to delegate to AI and where to add checkpoints and human review.

That last element is worth pausing on. The curriculum explicitly frames the human as the party who retains responsibility for the final result, even as more work is delegated to AI systems. Whether learners internalize that principle in practice is an open question, but the framing itself is a responsible one. It treats agent delegation not as a hands-off handover but as a supervision skill that must be learned.

The other paths follow the same structure of role-specific practice. Build with AI covers planning and implementing changes across the software development lifecycle, plus solution design, evaluations, retrieval, and operating AI systems in production for teams building on the API. AI Leadership asks leaders to assess where AI could create business value, connect initiatives to business priorities, define ownership and governance, and produce an adoption roadmap. AI for Educators has teachers plan classes and create assessments with permissioned materials, then review ChatGPT's responses against their own learning objectives, with the announcement stating that the educator decides what belongs in their teaching. AI for College Students focuses on academic and career tasks such as organizing study plans and reviewing drafts against assignment requirements.

A common thread runs through all of them: learners practice on real tasks, review results, and make the final decision themselves. That is a defensible pedagogy, and it is more substantive than promotional tutorials. It teaches verification, not just prompting.

The credential question: who recognizes a badge?

Passing a course assessment earns a badge. Badges are useful signals inside an organization that recognizes them, and the Deployment Guide is designed to help companies integrate the courses into onboarding and transformation programs. But OpenAI Academy badges are vendor-issued credentials for vendor-specific tools. There is no indication in the announcement that they are recognized by employers outside OpenAI's ecosystem, by educational institutions as credit, or by professional certification bodies.

This matters for the access angle. If free vendor training becomes the main way workers build AI skills, the credential gate is defined by the vendor itself. An employer or a school cannot independently verify what a badge holder actually knows, because the curriculum, the assessments, and the passing standards are all controlled by one company. Compare that with vendor-neutral programs such as national digital-skills frameworks or university continuing education, where curriculum decisions are at least subject to external governance.

None of this makes the badges worthless. As evidence of having worked through a structured curriculum, they have real value. But they should be understood as company certifications, not portable qualifications, and readers making hiring or admissions decisions should treat them accordingly.

Who defines competent and safe use?

Here is the deeper issue. Every element of the Academy encodes judgments about what good AI use looks like: where to place human checkpoints, how to evaluate AI output, what to delegate and what to retain. OpenAI is not wrong to hold positions on these questions. But a curriculum is never neutral, and when a single vendor supplies both the dominant tool and the dominant training on how to use it, the vendor's own product assumptions become the definition of competence.

Some examples are visible in the course descriptions themselves. The developer path teaches teams to build on OpenAI's API and use Codex, not how to evaluate whether an open-weights model or a competitor's service might serve better. The educator path teaches teachers to review ChatGPT's responses, not how to evaluate AI tools across providers or how to weigh the privacy implications of bringing student data into a commercial system. The leaders' path covers governance and ownership in the abstract, which is genuinely useful, but within a deployment mindset that presumes adoption rather than asking whether a given task should use AI at all.

This is not an accusation of bad faith. It is a structural observation about who sets the standard. A course that taught identical skills around a competitor's products would encode a different set of assumptions. The problem is not that the assumptions exist; it is that free, high-quality, vendor-provided training crowds out the space where those assumptions could be debated, because it is the easiest and cheapest option available to organizations and individuals.

What is missing: outcomes evidence

The GO Forward with AI program makes the access promise concrete, and it also shows the limits of what we can currently claim. OpenAI's announcement cites a Grab survey of Singapore driver-partners in which half said they use AI tools, and 87 percent of those who do not said they were open to trying them. Those figures are Grab's own survey of its own partners, and the announcement does not publish the methodology. They are best read as the program's motivation, not as independent evidence of need or impact.

What the program will deliver is clear: half-day in-person workshops, adapted curricula, and trained local trainers. What no one can yet claim is that these workshops improve earnings, business outcomes, or job quality for participants. No outcomes data exists. The quotes in the announcement, including from a driver-partner describing how workshops helped her explore ideas for her family's hawker business, are individual testimonials gathered by the program itself. They are encouraging and they are not evidence.

This distinction matters for policy. Programs like GO Forward with AI are often cited in access and digital-inclusion debates as proof that industry can close the AI skills gap without public intervention. That may turn out to be true. But two-year skilling commitments should be evaluated the way we evaluate workforce programs generally: with baseline data, follow-up measurement, and independent assessment. Until then, the honest description is that a large vendor is funding broad, free training on its own terms, and the results are unknown.

A checklist for readers who care about access

Independent evaluation of skill claims. Every assessment of what these courses teach comes from OpenAI. An external review of learning outcomes, even a small one, would materially strengthen the access case.

Recognition pathways. Whether governments, universities, or industry bodies will treat Academy badges as evidence of competence is unresolved. Access advocates should push for answers.

Coverage of model choice, privacy, and refusal. A curriculum that teaches how to use one vendor's tools but not when to decline, how data flows, or how to compare providers trains users for that vendor's world. Safe use includes knowing the limits.

Support for the excluded. Free online training still assumes internet access, device access, and time. The in-person Grab workshops address some of this. Whether the model reaches people outside platform partnerships is an open question.

The verdict: real access, unresolved questions

Free vendor-provided AI training is, on the evidence available, a genuine improvement in practical access. Millions of people can learn workflow building, agent supervision, and verification habits at no cost, and the curriculum's emphasis on human review and retained responsibility is sound. The Grab partnership extends that access to gig workers and small merchants who are often left out of corporate training programs.

But access and empowerment are not the same thing. When one company defines the curriculum, the assessments, the badges, and the deployment playbook, it also defines what competent and safe AI use means. That definition deserves scrutiny, independent evaluation, and eventually external recognition mechanisms, precisely because the training is good enough to become dominant. The right response is not rejection. It is engagement with a demand for evidence and openness: publish outcomes, invite independent assessment, and let the credential be portable. Progress needs a voice, and that voice should include people who did not write the course.