Two years ago, if anyone had told me the most AI-hostile demographic in 2026 would turn out to be Gen Z, I would have laughed. The generation that grew up with screens in its hands seemed like the obvious early adopter, ready to use the tools to look more skilled, more productive, and more employable than everyone else.

Instead, May commencement season turned into an open revolt. At the University of Arizona, students booed Google chairman Eric Schmidt for pitching AI’s world-changing potential. Gloria Caulfield, VP of strategic alliances for the investment firm and real estate developer Tavistock, drew the same reaction at the University of Central Florida when she compared the rise of AI with the Industrial Revolution. At Middle Tennessee State University, students shouted down Big Machine Records CEO Scott Borchetta for the offense of saying the word out loud.

The numbers back up the booing. A recent Gallup poll measuring AI adoption and attitudes among Gen Zers found the share who say they’re excited about AI fell from 36% to 22% in a year. The share who say they feel anger toward AI climbed from 22% to 31%. Older cohorts are skeptical too, but a sentiment swing this sharp inside the youngest part of the workforce is something I’ve never seen for a new technology this early in its lifecycle.

That swing matters because the AI industry has a serious PR problem at the worst possible moment. Anti-AI sentiment is hardening as the midterms approach, and politicians are picking up data centers as a wedge issue, supporting efforts to halt or slow the build-out of the facilities that fuel AI with the computing power it needs to function. If capacity can’t keep up with demand, the cost of compute will keep rising, and that will put hard ceilings on what newsrooms, marketing teams, and comms shops can actually do with AI in the year ahead.

The infrastructure squeeze is already here

It’s already happening. Anyone who runs Claude as part of their daily workflow knows the rhythm of the outages, and Anthropic’s own status dashboard tells the story in red over the past 90 days. The Claude Code boom has driven demand through the roof in 2026, and the company is scrambling to keep up. Anthropic signed a deal to buy computing power from Elon Musk’s SpaceX, and at the same time it closed the loophole that let builders run third-party software on top of their Claude subscriptions. Some of those setups were burning thousands of dollars of compute against a $200-a-month Claude Max plan. Now those teams have to use Anthropic’s platform directly or move to pay-as-you-go.

The angry posts in response were predictable, but the more useful read on the change is that it forced builders to reckon with the actual cost of what they were running. The choices are probably familiar to anyone trying to budget AI spend: switch to a cheaper model, possibly an open-source one, rebuild on Anthropic’s own platform, or shut the project down.

This was always going to happen at some point. As demand grows, free-compute workarounds will keep closing. The industry’s argument is that the squeeze would hurt less if compute were cheap and plentiful, which is the case being made for trillion-dollar infrastructure projects like OpenAI’s Stargate. For AI to deliver on its promises, compute has to flow like water. That means more data centers, and more power plants behind them.

Which loops back to why Gen Z is angry in the first place. Environmental concerns are near the top of their list, and AI’s energy footprint has only gotten harder to ignore since I wrote about it months ago. This isn’t a fight confined to politicians and podcasters anymore. At the companies I advise on AI adoption, employee surveys keep surfacing the same worry, and in some cases it’s starting to shape whether teams use AI at all.

Governance is the new AI strategy

You can argue about whether the pollution and water concerns are overblown. The cost question is harder to wave off. The leaders running the most ambitious AI programs are past the era of handing every employee a ChatGPT seat. They want agentic workflows, automated processes, and rapid prototyping through vibe coding, and they may be telling their engineers to get obsessed with “tokenmaxxing.”

It’s unclear so far how data center politics will play out. What leaders can actually control right now is governance. That’s more than running training sessions on which model does what, although that matters. Real governance is the balance between experimentation and direction. People need room to invent their own workflows, and the organization needs a way to make sure the compute it’s paying for is being put to good use. That doesn’t just mean “keeping costs down” it means accepting that the bill will sometimes be high and being confident the outcome will be worth it.

Through my consulting work with media companies and PR agencies, I’ve watched this play out in practice. One agency I worked with piloted a vibe-coding tool. Usage spiked early as employees tested its limits, and several different teams ended up building near-duplicate prototypes. The thing that saved them was high-bandwidth communication. They ran regular workshops and project reviews, learned from their people, and steered the work as they went. They eventually homed in on the use cases that actually delivered, in their case automating media intelligence, and the experiment surfaced something unexpected. The original platform wasn’t the right one. The agency ended up adopting a different tool and sunsetting the one it started with.

That’s one of many examples, and the lesson behind all of them is the same. If AI agents are going to do real work for your team, they need to run compute-heavy jobs. Compute is going to stay expensive for a while. The way to avoid the kind of top-down restrictions that suffocate innovation is for leaders to define what success looks like up front, get their teams trained on the tools and models, and build systems that surface collaboration and catch waste before it compounds.

That is what good governance actually looks like. The political fight over AI and data centers isn’t going anywhere, but the companies leaning into AI can still find a way through. The goal is to work inside the real cost constraints while shielding the people doing the actual work from feeling them.

A version of this column appears in Fast Company.

The post The AI industry has a Gen Z problem appeared first on The Media Copilot.

My usual focus is the cutting edge of AI in media, examining how journalists and media companies are using the technology to change the way they work, reach new audiences, and transform their organizations. But the reality is that a persistent stigma still hangs over artificial intelligence in the journalism world. In conversations I have with working reporters and editors, there’s clearly still a lot of reluctance, if not outright disdain, for using AI in almost any part of their work.

Recent media coverage, though, paints a different picture. The Wall Street Journal recently profiled how Fortune business editor Nick Lichtenberg uses AI to turbocharge his output, sometimes writing as many as seven stories in a single day. The same day, Wired highlighted how several prominent reporters—including independents like Alex Heath and Taylor Lorenz as well as The New York Times’ Kevin Roose—use AI in various editorial tasks, sometimes in the writing itself.

Taken together, it feels like a dam has finally burst. And I don’t think the timing is accidental—this shift is happening alongside the arrival of Claude Code and Cowork, which has put remarkably powerful agentic AI within reach of everyone and reshaped what people expect from these tools. (An interesting aside buried in all this coverage of journalists’ use of AI is that it appears Claude is rapidly becoming what the Mac became among media pros: the platform of choice for creatives who “know better.”)

A plagiarism scandal puts AI trust on ice

But just as the relationship between journalists and AI seemed to be thawing, a high-profile incident threw it back into doubt. Last week, The New York Times severed its relationship with a freelance writer who had submitted a book review that was at least partially AI-written. The review by Alex Preston, published in early January, included passages that were nearly identical to Christobel Kent’s review of the same book that was published in The Guardian months earlier.

Preston admitted he used AI to assist in writing his book review, saying that he had “made a serious mistake.”

The episode is a clear wake-up call for the Times—and not its first—about communicating AI policy to freelancers. But it also sends a warning signal to every newsroom that has been inching toward greater AI adoption. Here, suddenly, was an error that appeared to validate all the restrictive rules.

Confronting what happened directly matters. The incident steers us back into the dark cave of AI scandals in media—from CNET’s bot-authored service journalism to the made-up book titles in the Chicago Sun-Times’ “summer reading list” last year. It risks erasing the productivity and content optimization gains that many journalists and newsrooms have been making, and could push those just beginning to experiment with AI back toward the simplest possible rule: don’t use it at all.

That makes it essential to examine specifically how AI was deployed here, so we can draw a clearer line between responsible and irresponsible use. It’s easy to say there wasn’t enough “human in the loop” (an increasingly unhelpful term)—but where in the loop? With prompting, fact-checking, something else? The whole point of AI is to outsource some human decision-making to sophisticated machines, so rather than pointing out the obvious—that humans need to shape and monitor the process—it’s better to zero in on the specific decisions that AI was asked to make, and whether the human gave the right parameters and restrictions.

When you look at the details, the answer is clearly no. According to The Guardian story, the two reviews have eerily similar language—so close that it’s difficult to argue against outright plagiarism. Consider these side-by-side passages:

• Original review, published August 21, 2025: “most significantly a song of love to a country of contradictions, battered, war-torn, divided, misguided and miraculous: an Italy where life is costume and the performance of art, and where circuses spring up on wasteland.”
• Times review, published January 6, 2026: “populate what is ultimately a love song to a country of contradictions: battered, divided, misguided and miraculous. This is an Italy where life is performance, where circuses rise on wasteland.”

Given the dates and the undeniable overlap, a few things become clear. Preston evidently asked the AI—directly or indirectly—to generate text he planned to use in the piece, and not just from his own notes. The four-month gap between the two reviews (and likely an even longer lead time given the Times’ editing process) almost certainly means the AI’s training data didn’t include Kent’s review. That points to the AI tool pulling from web search (also known as RAG) to produce the copy.

This was the critical error. Giving Preston the benefit of the doubt, he may not have deliberately told the AI he was using to synthesize other reviews of the book, and perhaps it grabbed The Guardian review on its own. But he certainly didn’t tell the AI not to do that, which would seem to be an essential part of your prompt if you want to avoid the very plagiarized text he ended up including.

Moving from stigma to smart adoption

It bears repeating: in most cases, how you use AI matters far more than whether you use it. Getting there requires deep familiarity with these tools’ strengths and weaknesses, careful attention to prompt design, and a commitment to continuous adaptation. It’s an ongoing process, and it needs guardrails—such as “always” and “never” commands to avoid specific problems and (human) fact-checking. Without those safeguards, you’re handling a loaded weapon that can easily misfire.

Broader structural protections help, too. Whether you’re an independent writer or a full newsroom, it pays to have an AI policy. As a media AI trainer, I of course would encourage investing in training, but I think it’s still objectively a good idea. But most importantly, the trial-and-error that comes with figuring out the boundaries of “good AI” should be kept out of public view if you can avoid it.

When it comes to AI-assisted writing specifically, developing your prompts and safeguards in a private sandbox is critical. That might seem obvious, but one of AI’s most deceptive qualities is that it produces outputs that look indistinguishable from work that went through a rigorous human process. To someone without experience, that surface-level competence feels sufficient.

Truly making AI work as a writing and journalism partner means going beyond trusting the process—it means accepting responsibility for building, testing, and refining that process yourself. The more journalists do that, the more the stigma will fade.

A version of this column appears in Fast Company. It has been lightly “remixed” (alternate words and phrasings used) with AI assistance and human review.

The post Journalists are opening up about AI, but one mistake shows how fragile that progress is appeared first on The Media Copilot.

The Trump administration released its National Policy Framework for Artificial Intelligence on March 20 — a four-page document that tells Congress what the White House wants federal AI law to look like, and signals clearly what it doesn't want: state-level regulation, a new federal AI agency, or courts deciding that AI training violates copyright law.

The framework isn't binding. It's a legislative wish list that still needs Congress to act. But it maps the administration's priorities across seven areas, and for publishers and media companies, two of them matter most.

On intellectual property, the White House punts to the courts while tipping its hand. The document states the administration "believes that training of AI models on copyrighted material does not violate copyright laws" — then says it supports judicial resolution. That's a tell. For publishers currently suing OpenAI, Google, and others over training data — including Encyclopedia Britannica's recent suit against OpenAI and News Corp's ongoing case against Perplexity — the administration has effectively signaled it's rooting against them. The framework does contemplate collective licensing frameworks and protections against unauthorized replicas of people's voices and likenesses, but the core fair use question is left to judges who now know where the White House stands. That's also bad news for publishers pushing for statutory licensing models as a structural solution.

On federal preemption, the framework pushes hard to override state AI laws that "impose undue burdens" on a national strategy for "global AI dominance." The immediate target is Colorado's AI Act — the first state law requiring impact assessments and transparency for high-risk AI deployments — which was already delayed from February to June 2026 under industry pressure. The framework would put federal law above a patchwork of state rules, effectively neutering the most aggressive state-level efforts to regulate AI behavior. It's the opposite direction from the EU AI Act, which the administration's framework implicitly positions itself against.

The rest covers child safety — age verification and deepfake protections via the Take It Down Act, which targets non-consensual intimate imagery — infrastructure buildout, workforce development, and a preference for regulatory sandboxes and industry-led standards over a new AI regulator. The deepfake protections are notable given Grok's ongoing global regulatory scrutiny over sexualized AI imagery, though the framework addresses individual harm rather than platform accountability.

The overall posture is: light-touch federal rules, no new agency, and existing sector-specific regulators handle the rest. The contrast with the EU is deliberate. The administration's framework is a bet that the US approach — let companies build, let courts sort out the edges — will outcompete Europe's more prescriptive compliance regime. For publishers, that bet means the most important AI policy battles are now happening in courtrooms, not legislatures — and the traffic consequences of losing those battles are already here.

The post The White House AI blueprint tells publishers where the administration stands on copyright. Spoiler: It’s not with them appeared first on The Media Copilot.