The 10 dangers of AI and pigs fly

Today we talk about a minor literary genre that fills half the Spanish SERP on artificial intelligence: the listicle of "the X dangers of AI." I'll warn you up front that this article keeps a number in the title —ten— but it's a joke at its expense, not a listicle. The thesis is harsh. The real dangers aren't ten, they don't come out round, they don't fit a pretty infographic and, above all, they aren't the ones the listicle dares to number. My take is biased because I've written listicles in a former life and I know the logic from the inside. Form your own: open any "dangers of AI" listicle in Spanish, count how many are noise and how many are serious risks, and judge.

I've kept a file since 2023 where I hang every listicle of AI dangers that passes through my feed. I have twenty-three. Nineteen of the twenty-three have exactly ten points. Three have seven. One —gym press, don't ask how it got there— has five. None has four, eleven, or twenty-two. The distribution isn't random; it's the result of the format.

The economy of the listicle, explained calmly

The listicle —a numbered article with a headline of the type "the X best Y" or "the X dangers of Z"— was born as a web genre around 2008, at the moment outlets like BuzzFeed began aggressively optimizing for click-through rate on social media. The formal structure of the listicle has three advantages for the editor and almost none for the reader.

The first advantage is the click-through rate. A headline with a specific number promises digestible, bounded content completable in a few minutes. The traffic studies of Anglophone outlets published throughout the 2010s consistently documented that headlines with a number in a prominent position generated between 30% and 70% more clicks than equivalent headlines without a number. The editor who knows that figure applies it.

The second advantage is retention. When a reader starts a ten-point listicle, they know that reaching point ten will take them between four and seven minutes. The promise of a near close reduces the abandonment rate compared to an open article with no visible structure. The reader goes bullet by bullet feeling progress.

The third advantage is sharing. A listicle is shared better than a discursive article because it lets you cite specific points —"look at point 4, exactly what I was saying." Social-media algorithms reward that kind of content.

All three advantages are the editor's. The reader, meanwhile, gets in exchange a false hierarchy where important points and incidental points appear as equivalent items in a numbered list, with no distinction of weight.

Why always ten

The choice of number isn't the author's, it's the editor's or the SEO manager's. After nearly fifteen years of experimentation, most digital newsrooms have concluded that ten is the optimal figure. It's enough to seem exhaustive but not so much as to seem overwhelming. It's digested in a single reading session. It fits in a vertical infographic for Instagram. And, above all, it sounds like a complete inventory: "the ten dangers," not "some dangers," not "the most important," not "a sample."

That apparent solidity is the problem. Ten dangers suggest a closed taxonomy. But a closed taxonomy doesn't exist in a technology that changes every six months. Any listicle from three years ago had as its central danger "job loss through automation." Today that danger still figures but has mutated: it's no longer mass replacement but the redesign of tasks within the same job, with asymmetric effects by sector. A listicle from today should say different things than a listicle from three years ago. Most don't; they reproduce the same scheme with slightly updated names.

To reach ten dangers, moreover, you have to pad. When a newsroom inventories the real dangers that appear in serious academic literature —papers from FAccT, NeurIPS, AI Now Institute, Stanford HAI, AI Index— it finds two things. The structural dangers are fewer than ten. The flashy dangers are more than ten. To reach the magic number, the editor mixes both without distinction.

The real dangers, the ones that don't get padded

Here's the uncomfortable part and I go to it, precisely, without numbering it. The structural dangers documented with evidence in the recent academic and empirical literature are roughly the following, and the hierarchy within the list matters more than the total number.

The first structural danger is the corporate concentration of compute. Five or six companies control the infrastructure on which generative AI is trained and served: Microsoft, Amazon, Google, Meta, Nvidia, OpenAI. The costs we saw in 0055 make it practically impossible for new competitors to enter. That concentration has economic and political consequences that almost every listicle keeps quiet, probably because their own outlets receive advertising from the concentrated players.

The second structural danger is cognitive atrophy through delegation. When a tool does cognitive tasks for you that you used to do yourself —summarizing, drafting, calculating, remembering—, the corresponding skill weakens. The literature on GPS dependence and the decline of spatial orientation skills is just a preview. The studies published in Nature and in Science during 2024 and 2025 on intensive ChatGPT use and performance on independent reasoning tasks point, with nuances, to comparable effects.

The third structural danger is input capture. Every time you type into a chatbot you generate data that, depending on the service's policy, may be used to train the next generation of the model. The company that gathers the highest-quality private human corpus between 2024 and 2028 will have an insurmountable advantage when the public corpus saturates, as we saw in 0057. The user is a supplier of raw material without knowing it and, almost always, without compensation.

The fourth structural danger is the offshoring of critical labor. The big generative-AI platforms rest on chains of precarious workers in developing countries —annotators in Kenya, the Philippines, Venezuela and Colombia working for two dollars an hour— who label data, moderate violent content and train filters. Artificial intelligence looks automatic and dematerialized; in reality it's subcontracted, invisible human labor. The journalism of 404 Media and other outlets has documented this in detail.

The fifth structural danger is the erosion of intermediate institutions of knowledge. Education, journalism, book publishing, university teaching, vocational training. Each of these institutions performs a function of cognitive mediation between abstract knowledge and practical application. When a conversational model offers instant and seemingly competent synthesis, the mediation becomes economically unviable. The affected institutions may take decades to rebuild, if they rebuild at all.

To this are added verifiable secondary risks: algorithmic bias in high-risk applications —hiring, credit, court rulings— that Cathy O'Neil has documented for over a decade in Weapons of Math Destruction (Crown, 2016); military use of autonomous systems without adequate human supervision, exemplified by the Lavender case in Gaza documented by +972 Magazine in April 2024; pressure on energy and water resources as we saw in 0079.

Counting, adding structural and secondary, I get around eight. Neither ten nor five. The exact figure doesn't matter; the hierarchy does.

The padded dangers that always appear

To reach ten in the average listicle, you have to pad. The standard padders are always the same.

"The singularity" appears, the hypothetical moment when artificial intelligence surpasses human intelligence in all domains. It's a respectable philosophical hypothesis, defended by Ray Kurzweil among others, but it isn't an immediate measurable danger. Mixing it with the current structural dangers is mixing five hundred years of future with fifteen years of present.

"Skynet" appears, the cinematic reference to the autonomous system in Terminator that decides to exterminate humanity. It's a useful metaphor for science-fiction discussions but, as an operational danger, it figures in no serious taxonomy of short-term risks. It appears in the listicle because it sounds good. Not because it's likely.

"Killer robots" appear. Autonomous weapons are a serious debate in international humanitarian law, yes. But the framing "killer robots" trivializes the debate and shifts it toward a Hollywood imaginary instead of the documentable reality —which is, precisely, semi-autonomous systems like Lavender, with inadequate human supervision, already in operational use.

"The AI that becomes conscious" appears. The question of artificial consciousness is philosophically interesting. As a short-term operational danger, it figures in no serious technical literature. The human brain remains the best-documented example of consciousness and we don't know how to replicate it. Listing artificial consciousness as danger number 9 next to the corporate concentration of compute is an act of rhetorical balancing, not analysis.

"Deepfakes in general" appear, without distinguishing between non-consensual pornographic deepfakes —real, documented, frequent harm, with specific victims— and political deepfakes of heads of state —technically easy to produce but of limited electoral efficacy in markets where information is fact-checked. The failure to discriminate between the two adds points to the list but degrades the analysis.

The practical rule for the reader

There's a simple test to tell, faced with a listicle of AI dangers, whether it's worth reading or not.

Look at the title. If it says "the X dangers" with X being a round number —five, seven, ten—, you already know what you're dealing with. If it says "the main dangers" or "the structural risks" without a number, it hints at honesty.

Look at the authors referenced. If the listicle doesn't cite a single verifiable academic or journalistic source, it isn't popular science; it's text generated for SEO. If it cites Crawford, Russell, Gebru, O'Neil, Mitchell, Bender, some specific paper, there's a basis behind it.

Look at the internal hierarchy. A serious analysis distinguishes which risk is more relevant and why. A mediocre listicle presents them all as equivalent with identical bullets. If the ten dangers have the same paragraph length and the same visual weight, the false hierarchy is giving itself away.

Look at whether it mentions the conflict of interest. Does the listicle talk about corporate concentration? Does it mention who controls the compute? Does it discuss why its own outlet receives advertising from Microsoft or Amazon? The outlet's self-criticism is a marker of seriousness. Its absence is a marker of soft editorial capture.

If the listicle fails those four tests, it isn't serious popular science about the dangers of AI. It's content to fill the SERP, written by a copywriter on a four-hour deadline and an editor who's paid per click. That's neither bad nor good in a moral sense; it's exactly what it looks like. But the reader should know what they're consuming before taking it as a basis for forming an opinion.

The political question

This is personal opinion, but I hold it on two years of files. The proliferation of listicles about the dangers of AI isn't an individual failing of copywriters. It's a symptom of a media ecosystem that rewards form over substance when it comes to a sector most of the audience doesn't master. The social consequence is serious: the public conversation about AI polarizes between apologists who ignore the risks and catastrophists who exaggerate them, with no intermediate space for serious analysis. That polarization isn't accidental; it's produced by the format.

What I maintain is that the real dangers —corporate concentration, cognitive atrophy, input capture, the offshoring of critical labor, the erosion of intermediate institutions— are more serious than the padded ones, but less visceral. A killer robot moves you emotionally; a concentration of compute doesn't. That explains why the listicle always picks the visceral and why the public conversation has shifted accordingly.

What I would ask of the Spanish press, and again with no realistic hope of getting it, is that the listicles about AI dangers be replaced by structured analyses that differentiate verifiable systemic risks, documentable secondary risks, and speculations about a distant future. Those three categories don't fit in ten bullets. They fit in an article of several thousand words that the current format of the Spanish web doesn't dare to produce.

In the meantime, the individual reader can at least read the listicles knowing what they are. Not information. Numerical decoration with technical vocabulary.

The concrete figure that closes it. According to the Reuters Institute Digital News Report 2025, numbered formats —listicles, rankings, closed taxonomies— represent around 23% of the most-read content in the technology sections of the main European digital outlets, while in-depth analytical content —pieces over 1,500 words with verifiable citations— represents less than 6% of measured consumption. That proportion describes the ecosystem better than any abstract critique. And it describes, too, why this article, deliberately long and deliberately without ten numbered points, is doomed to be read by fewer people than the listicle it criticizes. That's exactly what the problem produces.

Definitions

Listicle: an article structured as a numbered list with a headline of the type "the X reasons why Y" or "the X dangers of Z." A web genre optimized for click-through rate and sharing on social media.

Analytical hierarchy: the explicit distinction between elements of a taxonomy according to their weight, available evidence and foreseeable consequences. The typical listicle suppresses it; serious analysis requires it.

Corporate concentration of compute: a situation in which a small number of companies control the computing infrastructure needed to train and serve AI models. It's probably the best-documented structural risk in the contemporary sector.

Soft editorial capture: a situation in which an outlet modulates its coverage out of economic or access dependence on the companies it covers. It doesn't require an explicit agreement; it's produced by sustained incentives. Examined in detail in 0067.

References

Cathy O'Neil, Weapons of Math Destruction (Crown, 2016). The canonical text on algorithmic bias in high-risk applications; one of the few books that sustains serious analysis without falling into the listicle.

Kate Crawford, Atlas of AI (Yale University Press, 2021). A material framework on the physical, labor and political costs of AI infrastructure.

Stanford HAI, Artificial Intelligence Index Report 2026 (April 2025). Section on media coverage and public perception.

Reuters Institute for the Study of Journalism, Digital News Report 2025 (Oxford University, June 2025). Figures on the consumption of numbered formats versus in-depth analysis.

Yuval Abraham, 'Lavender': The AI machine directing Israel's bombing spree in Gaza (+972 Magazine, April 3, 2024). A documented case of military use with inadequate human supervision.

Karen Hao, Empire of AI: Dreams and Nightmares in Sam Altman's OpenAI (Penguin Press, 2025). In-depth reporting that exemplifies serious analysis in long form.

International Energy Agency, Electricity 2024 (IEA, January 2024). Chapter on the energy impact of data centers, the basis for the structural danger over resources.

Going deeper

Nicholas Carr, The Shallows: What the Internet Is Doing to Our Brains (W. W. Norton, 2010). A framework for understanding cognitive atrophy through delegation.

Daron Acemoglu & Simon Johnson, Power and Progress (PublicAffairs, 2023). A historical analysis of how the benefit of a technology is distributed when the key resource concentrates.

Stuart Russell, Human Compatible (Viking, 2019). A rigorous framework on the control risks the listicle oversimplifies as "Skynet."

Ryan Holiday, Trust Me, I'm Lying (Portfolio, 2012; revised edition 2017). An insider's manual on how viral content, including the listicle, is designed.

You may also like

• Ten reasons never to read another listicle about AI in your life
• Five advantages of AI, depending on who's trying to sell it to you
• OpenAI's press release isn't journalism
• The real dangers of AI

Elsewhere

• AI Now Institute
• Center for Humane Technology