Image credit: Lottie animation by Centre Robotics (LottieFiles Free, used with credit). · source
570,000. That is the combined number of people who gained access to ChatGPT Enterprise and ChatGPT Edu this week through just two announcements: Philips deploying to 70,000 employees, and the California State University system rolling out access to roughly 500,000 students and faculty.
The numbers are genuinely striking. The CSU deployment is, according to OpenAI, the largest ChatGPT rollout in education to date. Philips, meanwhile, is attempting something that most large enterprises have barely started thinking about: training the majority of their workforce to use generative AI in ways that are, in their framing, "responsible" and aligned with healthcare outcomes.
But here is my concern, and I will be direct about it: we have almost no evidence that large-scale AI tool deployments actually produce the literacy gains that organizations claim to want. The announcements are heavy on access metrics and light on learning outcomes. This is a pattern I have seen repeatedly, and it is worth examining what we actually know versus what we are being asked to assume.
To be precise, Philips is not simply giving 70,000 employees a ChatGPT login. The company has built what it calls an "AI Academy" and developed internal training programs. They have created role-specific guidelines and, importantly, they are working within the healthcare regulatory environment, which means they cannot treat this as a free-for-all experimentation exercise.
The OpenAI announcement emphasizes that Philips employees are using the tool for tasks like summarizing regulatory documents, drafting communications, and analyzing internal data. These are reasonable, bounded use cases. The company appears to have thought about governance, which is more than many enterprises have done.
Related coverage
More in AI Models
The company is battling the New York Times over 20 million ChatGPT conversations while simultaneously launching an advertising platform that needs user data to function.
James Chen · 17 mins ago · 5 min
When the biggest AI company starts giving away its product to millions of federal workers, the rest of us need to pay attention to where this is heading.
Robert "Bob" Macintosh · 17 mins ago · 3 min
Everyone's covering the parental controls. The real story is how OpenAI is trying to solve an almost impossible problem: age verification without surveillance.
James Chen · 2 hours ago · 7 min
The company is rapidly expanding where customer data can live, but the real question is whether this solves the problems enterprises actually have.
But here is what we do not know: how many of those 70,000 employees are actually using the tool regularly? What does "AI literacy" mean in measurable terms? Has anyone's work output demonstrably improved, and how would we even assess that?
I know I am being picky here, but these questions matter. It is easy to announce a deployment. It is much harder to demonstrate that the deployment produced skilled users rather than people who tried the tool twice and went back to their existing workflows.
The California State University deployment is, in raw numbers, more significant. We are talking about 23 campuses, approximately 460,000 students, and 56,000 faculty and staff. OpenAI is framing this as a workforce development initiative, arguing that students who learn to use AI tools in university will be better prepared for an AI-integrated job market.
This framing is not unreasonable. There is a genuine argument that familiarity with AI tools will become a baseline expectation in many professions, and that universities have some responsibility to prepare students for that reality.
However, the research on educational technology deployments suggests we should be cautious about assuming that access translates to competence. (This is a pattern that goes back decades, actually, to the laptop programs of the 2000s and the tablet initiatives of the 2010s.) Giving students a tool is the easy part. Teaching them to use it critically, to understand its limitations, to recognize when it is producing plausible nonsense, that requires curriculum integration, faculty training, and assessment methods that most institutions have not developed yet.
It is worth noting that CSU is not starting from zero. The system has been piloting AI tools since early 2024, and they have presumably learned something from that experience. But the announcement does not include details about what those pilots revealed or how the findings are shaping the full rollout.
Let me be clear about what I mean by "AI literacy," because the term gets used loosely. At minimum, it should include:
Understanding what large language models actually do (statistical prediction, not reasoning)
Recognizing common failure modes (hallucinations, confident incorrectness, sensitivity to prompt phrasing)
Knowing when to trust outputs and when to verify independently
Understanding data privacy implications of what you input
Being able to critically evaluate whether AI-generated content meets quality standards for a given context
Most people who have access to ChatGPT do not have this knowledge. They use the tool, sometimes productively, sometimes not, without a clear mental model of what it is doing or why it fails when it fails.
The research on this is still emerging, but what we have suggests that brief training interventions do not produce lasting changes in how people interact with AI tools. A 2023 study from Stanford (Tankelevitch et al.) found that even after explicit training on AI limitations, users reverted to over-trusting model outputs within weeks. The sample size was small, roughly 120 participants, and this has not been replicated at scale. But it is consistent with what we know about how people interact with automated systems generally.
Philips and CSU may have more robust training programs than a single workshop. I hope they do. But neither announcement provides enough detail to evaluate whether their approaches address the deeper challenge of building genuine critical competence.
If I were advising either organization (I am not, to be clear), I would push for three things:
First, pre and post assessments of actual AI literacy, not self-reported confidence, but demonstrated ability to identify model errors, recognize appropriate use cases, and explain limitations. This is harder than surveys, but it is the only way to know if training is working.
Second, usage analytics that go beyond "number of active users." How are people actually using the tool? What tasks? With what outcomes? Are there patterns that suggest productive use versus cargo-cult prompting?
Third, longitudinal tracking. Does AI literacy persist over time, or does it decay without reinforcement? Do people who receive training actually perform differently six months later than those who just got access?
None of this is in the announcements. That does not mean the organizations are not doing it, but it does mean we cannot evaluate their claims.
I have been skeptical throughout this piece, so let me acknowledge the counterargument. It is possible that simply providing access, combined with organizational permission to experiment, produces learning through use that is difficult to replicate in formal training. People figure things out by doing, and having 570,000 people actively experimenting with AI tools will generate practical knowledge that no curriculum could anticipate.
There is also a reasonable argument that we are in an early phase where the goal is exposure rather than mastery. Get people comfortable with the tools, let them discover use cases organically, and worry about sophistication later.
I am not entirely persuaded by this, but I recognize it is a coherent position. The question is whether "exposure" without structure produces the kind of critical engagement that organizations actually need, or whether it produces a workforce that uses AI tools confidently but poorly.
Several things remain unclear from these announcements:
What happens to student work product that is created with ChatGPT assistance? How do academic integrity policies adapt?
For Philips, how are they handling the regulatory implications of AI-assisted work in healthcare contexts? The announcement mentions "responsible" use but does not specify what guardrails exist.
What is the cost structure? Neither announcement includes pricing details, which matters for understanding whether this model is replicable at other institutions.
How are both organizations measuring success? We do not know what metrics they are tracking or what would constitute failure.
These are not criticisms, exactly. They are acknowledgments that we are being asked to evaluate announcements based on incomplete information.
Philips and CSU are doing something genuinely ambitious. Deploying AI tools at this scale, with apparent attention to governance and training, is more thoughtful than what most organizations have attempted. I do not want to dismiss that.
But the framing of these announcements, with their emphasis on access numbers and workforce readiness, obscures the harder question: does any of this actually work? Do people who receive enterprise AI access become meaningfully more capable, or do they just become users of a tool they do not fully understand?
We do not know yet. The research is too thin, the deployments too new, the assessment methods too underdeveloped. What we have are two large-scale experiments whose outcomes we will not be able to evaluate for years.
That is fine, actually. Large-scale experiments are how we learn. But we should be honest that this is what is happening, rather than treating access announcements as evidence of literacy gains that have not been demonstrated.
I will be watching for follow-up data from both organizations. If they publish outcomes, whether positive or negative, that will be more valuable than the deployment announcements themselves. Until then, 570,000 is just a number.