In this article
Today we get into an experiment I keep recommending cautiously because it sounds like a joke and isn't: stop reading headlines about AI for three months. Replace them with a book a month, a paper a week and an analysis newsletter a day. Why? Because AI evolves on a scale of months, not hours, and the headline format fits that scale terribly. My opinion: I've done this experiment three times in two years, lasted all three, and come out understanding more at the end than at the start each time. Form your own: try a month, not three, and judge.
I've been subscribed for years to more X, LinkedIn and breaking-news accounts than I'll admit in company. The first time I switched it all off for a month, I felt stupid. By the second week I missed something. By the third I stopped missing it. At the end of the month, when I reopened the backlog of accumulated headlines, I found three serious things and a pile of noise. The experiment turned into an intermittent habit, and that's where the article comes from.
Why the headline is the worst format for understanding AI
There's a structural problem with the headline format when applied to generative artificial intelligence. A headline's informational density is deliberately low, to grab attention. That low density works well for national politics, where the reader has prior context, or for football, where the result is understood instantly. It works terribly for a sector where every story includes technical jargon, a figure that needs qualifying and an actor whose name isn't known.
The serious part isn't that the headline informs little. It's that it produces the illusion of having informed. When a reader has read eighty headlines about AI in a month, they have the sense of knowing the sector. The sense is real; the knowledge isn't. Ask them what exactly a transformer is, why Anthropic was founded in 2021, why Stochastic Parrots mattered, or what the difference is between OpenAI's deal with Microsoft and Anthropic's with Amazon, and the answer is usually silence or vagueness.
There's also a rhythm problem. AI progresses in spaced-out milestones. The Attention Is All You Need paper came out in June 2017; its sector effects began to show in 2019; mass consumerisation with ChatGPT arrived in November 2022; the DeepSeek moment came in January 2025. Between each milestone there are long periods of internal evolution inside the labs, with no visible news. The headline tends to manufacture news in those empty periods, inventing urgencies — "what the CEO said," "the new model X that's 3% better on MMLU," "the fight between two executives" — that within a year nobody remembers. The consumer of headlines stays hooked to that conveyor belt. Whoever skips it loses nothing.
The concrete three-month diet
What follows is the operational recipe. It's not the only one possible, but it's the one I've tested.
For the ninety days, replace all consumption of headlines about AI — X, LinkedIn, technology supplement, breaking-news newsletters, news podcasts — with three streams.
The first stream is a book a month, read calmly in the afternoon, not on the metro. For the three months I suggest, in this order: Atlas of AI by Kate Crawford (Yale University Press, 2021), Human Compatible by Stuart Russell (Viking, 2019) and, if you have time, AI Superpowers by Kai-Fu Lee (Houghton Mifflin, 2018) or Cloud Empires by Vili Lehdonvirta (MIT Press, 2022). Each covers a dimension — the materiality of the infrastructure, the alignment problem, geopolitics, the platform economy. Three months, three books, three different layers.
The second stream is a paper a week. Twelve papers in all. The minimum list: Attention Is All You Need (Vaswani et al., 2017, arXiv:1706.03762); On the Dangers of Stochastic Parrots (Bender, Gebru, McMillan-Major, Mitchell, FAccT 2021); On the Opportunities and Risks of Foundation Models (Bommasani et al., Stanford CRFM, 2021, arXiv:2108.07258); Sparks of Artificial General Intelligence (Bubeck et al., Microsoft Research, 2023, arXiv:2303.12712); Constitutional AI (Bai et al., Anthropic, 2022, arXiv:2212.08073); Emergent Abilities of Large Language Models (Wei et al., 2022, arXiv:2206.07682); Will We Run Out of Data? (Villalobos et al., Epoch AI, ICML 2024); AI Models Collapse When Trained on Recursively Generated Data (Shumailov et al., Nature 2024); Chatbot Arena (Chiang et al., 2024, arXiv:2403.04132); The Leaderboard Illusion (Singh et al., April 2025, arXiv:2504.20879); Training Compute-Optimal Large Language Models (Hoffmann et al., DeepMind, 2022); and one free choice each month, depending on the reader's interest.
Some will think twelve papers are too many for someone with no technical background. They aren't, if you read them the way you read an essay: introduction and conclusions first, the main figure next, the technical sections if it interests you. The paper is accessible if you accept that you won't understand all of it, and that understanding a third already puts the reader ahead of 95% of the audience.
The third stream is a daily slow-analysis newsletter. I recommend Stratechery's free Monday column, the free Platformer pieces, Import AI by Jack Clark — Anthropic's former head of policy — and the weekly AI links from Marginal Revolution. Total time invested: between fifteen and thirty minutes a day.
That's all. Fifteen to thirty minutes of reading a day, a weekend paper, a book a month. Total: two and a half to three hours a day over ninety days. It's exactly the time many of us used to lose scrolling the news.
What happens in the first week
The first week is the worst and it helps to anticipate it. The brain misses the microdose of novelty. The feeling is exactly the same as with added sugar when you stop drinking soda: absence, irritation, a distraction hard to place. The temptation to open X "just to look for a second" appears. The fear of being left out of something important appears. It's calibration anxiety, not a sign of error.
What helps in that first week is knowing two things. The first is that almost nothing appearing in the headline stream over those seven days will be remembered in a month. The vast majority is noise self-induced by the format itself. The second is that the truly important stories — OpenAI's board coup, the DeepSeek-R1 launch, ChatGPT passing a billion users — will reach the newsletter consumer and the book reader with a day or two's delay, already contextualised. The advantage of being first to know is very small, unless you're trading on a horizon of minutes.
What happens in the second and third week
Here something changes. Past the initial withdrawal, the reader starts to notice the effect of the new stream. The first book finishes. The first complete paper gets read. The first Monday Stratechery newsletter arrives and some part is hard to follow because there's no prior context. They open the next one the following Monday with more context and follow it better.
What the reader starts to notice is something the headline stream had been hiding: the sector's pieces connect. Why Anthropic was founded in 2021 by people who left OpenAI isn't an anecdote; it's a deliberate decision about alignment that Russell's book explains in detail. Why OpenAI needed Microsoft isn't a tactical alliance; it's an inevitable consequence of the training costs that Lehdonvirta's book frames in the cloud economy. Why the DeepSeek coverage was so loud wasn't because of the model itself, but because of what the model meant within the geopolitical frame Kai-Fu Lee's book had been describing for years.
Every connection the reader discovers for themselves — not one they were told — is a piece of the sector they can no longer forget. That solidity is what the headline never produced, because the headline treated each story as isolated.
What happens by the end of the month
Past the first month, there's a quantifiable change. The reader starts to read the weekend paper — if they decide to open it again — as someone with a prior frame. The figures appearing in headlines have context. The names cited have a biography. The arguments reproduced have a tradition.
The interesting side effect is that the reader starts to notice errors in media coverage. Not minor errors; structural ones. When a supplement mistranslates a figure from the AI Index. When a columnist confuses two benchmarks. When an interview omits the interviewee's conflict of interest. Those errors went unnoticed before because the reader had no context. Now they're visible. And, once visible, they can't be unseen.
This has an uncomfortable consequence. The reader who has done the diet comes back less comfortable with the standard media ecosystem. The things that used to look like information now look like infodecoration. The loss of trust is real and isn't unpleasant: it's operational.
The month-three test
At the end of the experiment, it's worth asking an honest question. Do I need to go back to the old stream?
The most common answer, among those who've tried it, is not entirely. The vast majority of readers who've done the diet don't return to their prior headline consumption at a hundred per cent. Some open X once a day for ten minutes instead of four times a day for an hour. Others check the headlines only on Sundays. Others keep the new stream and reduce the old to zero. Each finds their own mix.
What does change, systematically, is the internal hierarchy of consumption. The analytical piece comes to occupy more mental space than the news piece. The book takes priority over the supplement. The paper gets read, not ignored. The general feeling drops: less urgency, less anger, less hooking; but more density of information retained per unit of time.
That shift of axis is what the experiment is after. It isn't a digital retreat, it isn't a moral purge, it isn't a protest. It's a pragmatic readjustment of the informational consumption of a sector the current format covers badly.
The political question
This is personal opinion, but I hold it with Reuters Institute data. News fatigue has grown systematically over the last five years. The Digital News Report 2025 documents that 43% of surveyed Spanish readers actively report avoiding the news at least part of the time, a figure that in 2017 was around 27%. The main reason cited is the negative emotional impact and the sense of not understanding what's going on despite reading a lot.
This matters. The headline news diet not only doesn't work; it's producing exhausted readers who disconnect entirely. The alternative to the headline isn't silence, it's slow reading. Whoever only chooses between "maximum noise" or "total silence" is left with no middle options and, almost always, ends up in silence.
What I hold is that the reader who spends time reading books and papers about the sector is, besides better informed, less burned out. The chronic anger of the X reader on AI isn't due only to what gets published. It's due to the diet. When you change the diet, the anger drops. Not entirely, not forever, but it drops.
The concrete figure that closes it: in an unpublished survey that circulated among English-speaking Stratechery readers at the end of 2024, picked up on Ben Thompson's own blog, 71% of subscribers reported having drastically reduced their X consumption in the previous twelve months. The figure is anecdotal, not statistically robust. But it points to a pattern many sector professionals recognise: the more you know the sector, the less you follow it via the fast lane. Slow reading is, too, a form of craft.
Definitions
Informational density: the amount of useful information a format transmits per unit of time or words. The headline has very low density; the technical paper has very high density for someone with context, and low for someone without it.
News diet: the deliberate combination of sources a reader regularly consumes. As with food, it's not just about how much is consumed, but about the proportion of each type of source.
News fatigue or news avoidance: a phenomenon documented by the Reuters Institute, whereby a growing share of the audience actively avoids the news out of emotional exhaustion. In Spain the share is 43% according to the 2025 report.
Slow reading: the practice of reading long texts without interruption, with sustained attention, without jumping between tabs. Documented as a habit by Nicholas Carr in The Shallows and by Cal Newport in Deep Work, among others.
References
Reuters Institute for the Study of Journalism, Digital News Report 2025 (Oxford University, June 2025). Figures on news fatigue and consumption by country.
Kate Crawford, Atlas of AI (Yale University Press, 2021). Recommended book for the first month; a material analysis of the infrastructure.
Stuart Russell, Human Compatible: Artificial Intelligence and the Problem of Control (Viking, 2019). Recommended book for the second month; the alignment problem.
Vili Lehdonvirta, Cloud Empires (MIT Press, 2022). Optional recommended book for the third month; the cloud economy.
Kai-Fu Lee, AI Superpowers: China, Silicon Valley, and the New World Order (Houghton Mifflin, 2018). Alternative recommended book for the third month; the geopolitical frame.
Jack Clark, Import AI (jack-clark.net). A weekly analysis newsletter on AI by Anthropic's former head of policy.
Ben Thompson, Stratechery (stratechery.com). The free-access Monday column; weekly strategic analysis.
Tyler Cowen & Alex Tabarrok, Marginal Revolution (marginalrevolution.com). Weekly links on AI, economics and politics.
Going deeper
Nicholas Carr, The Shallows: What the Internet Is Doing to Our Brains (W. W. Norton, 2010). A classic text on the cognitive effect of fragmented reading.
Cal Newport, Digital Minimalism: Choosing a Focused Life in a Noisy World (Portfolio, 2019). An operational manual for reducing digital consumption without losing functionality.
Neil Postman, Amusing Ourselves to Death (Viking, 1985). A conceptual precedent for why the medium determines the content.
Karen Hao, Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI (Penguin Press, 2025). For anyone wanting to go deeper after the diet with the best journalistic coverage of OpenAI in recent years.
You might also like
- Stratechery, Platformer and 404 Media, AI journalism done seriously
- Why Spanish journalism about AI is very bad
- Can you learn AI seriously by watching YouTubers?
- Ten reasons never to read another listicle about AI in your life
Comments0
No comments yet.
Leave a comment