When the announcement is better than the technology
There’s something curious about how people talk about artificial intelligence. Every week there’s a new model that “changes everything,” a startup that “revolutionizes” some industry, a CEO explaining in a three-hour interview why we’re about to witness the greatest shift in human history. The noise is constant. And so is the feeling that something doesn’t quite add up.
I’m not saying AI is a fraud. That would be absurd. But I do think there’s a massive gap between what gets announced and what actually happens when you sit down and work with it for real. And that gap has a name: marketing.
Silicon Valley has spent decades perfecting one particular art: selling the narrative before the product. AI is no exception. It might be the most sophisticated example we’ve seen.
What happens when you actually use it
I’ve been using these tools seriously for nearly two years. The last six months especially, combining real automation work with my master’s degree, I’ve reached a point where I can say something with a reasonable basis: AI is genuinely good at concrete, simple, well-defined things. Bounded tasks. Clear questions. Clean contexts.
The problem starts when you ask it for something more complex. Real useful work, real problems. When the context stretches, when you need to connect many different pieces of information, when the work requires actual judgment and not just execution. That’s where it fails. And it fails in a particular way worth pointing out: not all at once, but by accumulating patches.
Instead of undoing a mistake and rebuilding properly, it tends to correct on top. Layer over layer. And over time that stack of corrections becomes fragile, hard to control, full of inconsistencies that aren’t always visible at first glance. This deserves its own deeper analysis, and I’ll do that in another article. But I mention it here because it’s part of the honest conversation about what we’re actually being sold.
The marketing of the ethical narrative
One of the examples that has struck me most is Anthropic, the company behind Claude, the model I use most. A few months ago, the news went viral that they had refused to sell their AI to the Pentagon for autonomous military use. Many users uninstalled ChatGPT and installed Claude. The story worked perfectly.
The detail is that Claude had already been working with the Pentagon since 2024, through Palantir. And Anthropic had applied to a military drone control contract. The condition they set, “no autonomous weapons,” was something their technology can’t do today anyway.
Does that make them liars? Not necessarily. But it illustrates something worth keeping in mind: frontier tech companies have learned that an ethical narrative is a financial asset as real as the technology itself. Every press release, every public gesture, every study they publish also serves a purpose: attract investors, retain talent, prepare for an IPO.
Understanding that doesn’t make you cynical. It makes you a better reader.
What’s left when you turn down the volume
I don’t want to close this with a firm verdict, because I don’t have one. AI is real, it’s useful in many contexts, and it’s probably one of the most important technologies we’ll see in our lifetime. But there’s an enormous difference between that and what gets announced every week.
What I find myself asking more and more is: how much of what we consume about AI is information, and how much is an extraordinarily well-executed marketing campaign?
I don’t have the answer. But I think asking the question is already a good place to start.