Is the generative AI revolution stalling out?

Aug 28, 2025 - 17:40

0 0

Is the generative AI revolution stalling out?

Welcome to AI Decoded, Fast Company’s weekly newsletter that breaks down the most important news in the world of AI. You can sign up to receive this newsletter every week via email here.

Could generative AI be just a minor revolution?

On a recent episode of the TBPN podcast, Jordi Hays asked his cohost John Coogan whether his life would really be that much worse if he couldn’t access generative AI tools like ChatGPT and Claude. Would AI’s absence be as disruptive, he asks, as the sudden disappearance of smartphones, or TVs, or electricity? Coogan conceded it wouldn’t. The real-life impact of AI varies—someone going through a hard time might find relief in the counsel of a chatbot—but it’s reasonable to judge the technology in societal and historical terms, because that’s how its biggest cheerleaders and hype-spreaders frame it.

So the question becomes, “What is the real progress of the AI revolution?” Based on the past two years and nine months since ChatGPT’s debut, should we believe generative AI will change the world on the scale of the industrial revolution, the internet, or the mobile revolution?

As Financial Times politics and culture columnist Janan Ganesh notes in a recent op-ed, those predicting that generative AI will usher in abundance and well-being are the ones working closest to the technology and presumably understand it best. But they are also the ones with the most to gain from overselling it, and the most reluctant to admit they’ve dedicated their careers to something with only modest impact.

To be sure, generative AI is an amazing technology. Anyone who has used ChatGPT’s Deep Research, Anthropic’s Computer Use feature, or Google’s Veo 3 video generator can see that. It’s also the fastest-adopted technology in modern history: Gen AI apps reached 39.4% adoption among U.S. adults in just two years, compared to the four years it took for smartphones to hit 35% adoption following the iPhone’s 2007 launch.

Yet adoption hasn’t translated to willingness to pay. Only about 3% of users subscribe to premium tiers, according to Menlo Ventures’ State of Consumer AI report. (Mobile computing, by contrast, always required buying a handset and a cellular plan.) Globally, AI apps are bringing in only about $12 billion in annual revenue from 1.8 billion users. Two of generative AI’s biggest players, OpenAI and Anthropic, remain far from profitability. Meanwhile, Nvidia, which sells the chips that power AI, made $130 billion last fiscal year. OpenAI and Anthropic remain far from profitability.

In the late summer of 2025 generative AI’s honeymoon period may be coming to an end. Consumers and businesses are less interested in being dazzled and more focused on actually being helped by it. The attention has shifted to whether AI can actually overhaul aging business practices. And AI is indeed making some tasks, like coding, more efficient, saving time and sometimes payroll. But few CTOs are claiming generative AI is transforming their business, at least not yet.

Large enterprises are pouring money into AI projects, but many stall. An MIT report last week shook investors by finding that 95% of enterprise AI projects fail to substantially improve efficiency or profits. The research shows that the models aren’t the problem; the challenge is integrating them into company data, workflows, and infrastructure. In other words, it’s an application problem—one that’s lingered since 2023.

AI companies seem to recognize this. Covering them, I hear less talk of artificial general intelligence (AGI) and “superintelligence,” and less emphasis on monolithic models that can do everything. In reality, most AI workloads are handled by teams of specialized models. (OpenAI even described GPT-5 not as a “model” but as a “system” of models.)

Generative AI is technically complex and hard to grasp in detail. But that shouldn’t mean its impact can only be judged by those inside AI labs or big tech companies. What really matters is whether it measurably boosts productivity in business, and whether it leaves societies healthier, better educated, freer, more prosperous, more creative, and less bored at work.

Anthropic’s settlement with authors could set a precedent in future copyright cases

Anthropic is poised to settle a lawsuit brought by a group of authors who alleged the company trained its Claude models on their copyrighted books. A court filing Tuesday shows the parties have agreed on preliminary terms. Judge William Alsup gave them until September 5 to finalize the details and submit the proposed settlement.

In June, Alsup ruled that Anthropic’s use of digitized books qualified as “fair use” under the Copyright Act, but that the company had obtained the works unlawfully from a “shadow library” (including the notorious LibGen site). In late July, he certified that the class could include any author whose copyrighted book Anthropic downloaded from such libraries, meaning the company could have faced damages of $150,000 per book across potentially thousands of titles. Instead, Anthropic opted to settle.

Like its peers, Anthropic relies on vast amounts of online text to pretrain its large language models (LLMs). These models process data for weeks or months to build an understanding of language and context, forming a basic knowledge of how the world works. Content owners continue to sue AI companies over this practice, with several major cases still ongoing.

In his June ruling, Alsup addressed the broader issue at the heart of these lawsuits, writing that AI companies use copyrighted content in a “transformative” way—even when an LLM is just memorizing text—and therefore in a manner protected by “fair use.” That reasoning could become the defining legacy of Bartz v. Anthropic.