What do you think?
Rate this book
Atlas da IA
Em Atlas da IA, Kate Crawford revela os impactos ocultos desse sistema tecnológico, que se apresenta como algo positivo –divertido, até –, mas se enraíza em exploração ambiental e social, que começa na extração de minérios essenciais ao seu funcionamento, inclui subemprego maciço de trabalhadores desprotegidos e a coleta maciça de dados. Kate Crawford observa que é preciso questionar o que é a IA, a que interesses serve, quem explora e quais ameaças apresenta. Esta é uma leitura essencial para quem quer entender o que está por trás desse sistema que redefine nosso presente e molda o nosso futuro.
479 pages, Kindle Edition
First published January 1, 2020
929 people are currently reading
10232 people want to read
Ratings & Reviews
Friends & Following
Create a free account to discover what your friends think of this book!
Community Reviews
Displaying 1 - 30 of 381 reviews
January 10, 2022
I really had high hopes for this book considering the reviews, but even the way the book is framed is a red flag. Let's just walk through this book linearly:
Introduction: Let's define the entirety of AI as stupid, dumb, and mean calculations.
Not only does Crawford never even attempt to give a sufficient definition of what an Artificial Intelligence is, she mostly just calls it names. Granted, she loosely ties it to common usages and popular interpretations of how AI has been used, but she never gives it a clear definition, or explains broadly how it is different from other kinds of computation. But she only uses these straw person positions of what AI are in order to say we don't need one. This is intentional. It enables a particular perspective of AI that does not have to be very nuanced, and it allows her to make all sorts of arguments without justification. It essentially makes it really easy to read this book and think you learned something about AI when really what you are about to learn is the history of basically every technology invented since the 1800s. This book was praised as being the book of the moment and an instant classic. This book might as well be about the printing press. The arguments of its devastation to society would hold just the same at the time of its industrialization.
Chapter 1 Earth: Mining is bad. Computers use parts that are mined.
Yes, Crawford initiates with an ethnographic style that she had a paid field trip to visit a mining site, yet she never seems to use any ethical ethnographic practices. "I'm here to learn what AI is made from." Okay, but she doesn't use any of the information she got there that she couldn't have gotten from a Google map search and literal hundreds of years of news reports about the hazards and labor issues in mining. As best as I can tell, she doesn't quote a single person from the site, she doesn't help anybody there, and most of her arguments aren't about AI specifically but are about computers we all use everyday. It's almost like this chapter intends to say AI did this when literally the entire world of computing plays a part. Your smartphone is a part of this. The entire chapter (and even the book as a whole) is riddled with narrative holes such as connecting all of statistics with eugenics without further explanation or motivation for reason in a chapter about mineral mines. Am I to believe that anyone who computes a t-statistic is connected to a goal of eugenics oriented computation? The author is not clear on this but she certainly wants the reader to uncritically follow that narrative.
Chapter 2 Labor: There's a history of labor exploitation in computation.
Uh, yes. There's a history of labor exploitation. How is this unique to AI? It isn't. Why is this a book about AI again? Why are we ignoring more than 200 years of labor literature on this exact subject?
Chapter 3 Data: Models often use data abstracted from their origins which carry information we should address.
This is likely the best chapter in the book. For the first time, the author addresses something interesting about machine learning and AI. The authors argument that we use data that are deeply personal to some people (specifically in the use of mug shots used for crime prediction or facial recognition), because it strips those people of their humanity through abstraction and sets a norm of being able to use people's image data without asking them. However while this argument is obvious, the way in which she makes it doesn't address data itself as the chapter proclaims. It addresses models. Otherwise, why not critique the Census while we're at it? Is that not basically the same? No, because according to the author it's not about the data itself (except that the image data make her feel more strongly about this particular case) but it's about dehumanizing modeling techniques. But later, it's not about modeling techniques because it's about the fact that these data contain histories of things these models don't model. This story is not well unpacked. She cannot argue both models and data are steps after each other in dehumanizing people simultaneously. This is where Crawford's refusal to directly explain how AI is defined and works mechanistically hurts her the most. WIthout this nuance, it's not clear where the problem is. And so for the remainder of the book, we talk about government programs without clarity of what was built at a datafied level or at a modelled level, and so a casual reader would have to already know how AI works to actually learn the ethics here. The book does not give anyone the chance to reflect on their own relationship to data and programming. What practical ethical consideration should I take away from this? Despite critiquing induction/deduction in computation, the implication is to induct that all things about social data trace back to unnecessary and dehumanizing abstractions. But all a data scientist can do is say, "Yes, this case is sad," and be educated enough to know that this isn't true in every case. But Crawford implies again and again that this basically is the history of AI and datafication.
Chapter 4 Classification: Ontologies have implications.
Now she turns on the laborer she earlier claimed was exploited, and says that when we use cheap labor to classify stuff, we end up with racist outcomes and the degradation of linguistic principles. The author remains conservative to following an epistemology that she doesn't explain about language and classification. This is nothing but an elitist trope that classification experts should do classification since data scientists aren't willing to ethically clean their own datasets. While there are good points here about the harms of classification, just read Bowker and Star instead. It seems more reasonable for me to say this is a mix of (1) inductions without critique are bad, (2) some people are overtly problematic in their speech and classification (and in fact classification itself is problematic ontologically), and (3) there simply isn't enough labor and financial ability for scientists on minimal budgets to do this work. In summary what Crawford's policy here is, don't let pedestrian people datafy things because they will muddy up your data. This feels elitist about interpreting how MTurkers work, and unfair to researchers who cannot meet publication requirements on their own. It also doesn't account for the argument that researchers muddy their own data all the time, computational data or not. "Expert" checking doesn't solve this on it's own. Crawford is turning her back on her own labor chapter. Researchers can't work in isolation for most of academia in order to answer the questions posed to them, and the use of crowdsourced data (while problematic and should be addressed more by the standards of the Labor chapter) is not necessarily lesser quality data just because they are unaware of the academic training. If anything, an oppositional reading of this section teaches us that researchers can get better at cleaning crowdsourced data with experience of what to look for or can use that same data for a different purpose related to the biases it contains. Unfortunately, instead of actively looking for ways to continue valuing the labor of our peers, she suggests we induct that all datafication processes using classification are problematic necessities of AI because nobody is looking at datasets they use, which is laughable. While its been described this way, clearly somebody eventually looked at all the datasets Crawford claims weren't looked at because it's those researchers trying to correct these problems that gave her the historical data to write the chapter.
Chapter 5 Affect: Reading people's emotions from their faces is not a science.
While the simple argument here is true: we can't look at a person's face and know their feelings, this chapter badly butchers its theory. Firstly, while her main antagonist, Elkman, actually did attempt to do this and its still used in computational research, the way she frames Tomkins is horrendous. He argues that every body has an affective interpretation, not necessarily that all bodies essentially are the same in how we interpret their affects. Secondly, Tomkins does not fixate on singular human affects as the only individuals of interest. Tomkins is a post-human psychologist. Crawford reduces Affect theory to simple emotions of individuals (as do most communication scholars actually; it's kind of a meme at this point.) In total, this is probably one of the worst readings of Tomkins I've ever seen. There's no "affect theory" here in this interpretation. This is just badly interpreted psychology theory. For anyone who wants to check this out, look up Eve Sedgewick's Tomkins reader _Shame and its Sisters_. For affect theory more broadly, Spinoza's _Ethics_ and Deleuze make great starting points also. Even a brief overview from experts in affect should convince you of how badly Crawford misunderstands affect theory.
Chapter 6 State: Lots of computational tools used to police and track people online are AI tools that create logics to hurt "others" as a part of datafied "precision."
Again, we knew this. She briefly discusses Benjamin Bratton's book (which this chapter is merely a summary of), but also _We Are Data_ describes this very carefully. Crawford supposedly, in hacker-like style, had access to all of Snowden's data release and still the best she could do with it was repeat the same arguments others did. Yes, AI is used to track people and occasionally decide to kill them. Yes, this is dehumanizing to reduce humans to numbers. Yes, this is what new nationalist racism looks like. But more could have been said that summarizing the work of others as a part of having pulled a couple of slides from Snowden's files and putting them in a book.
This book is simple to interpret because it's written to be devastating and have the sense of absolute completion. While I would like to say that I agree with most of it, half of the stuff that I do agree with isn't new here even though it's written stylistically to look like she's uncovering conspiracies like some Jason Borne-level academic. The other half of stuff that I agree with ties itself to framings that might be more problematic to academic standards and labor ethics. This does not even include that I think that most of this felt really disingenuous. The narrative turns from the ethnographic writing style to carefully selected historical tangents you can find on Wikipedia that have nothing to do with the immediate ethnographic framing give me the feeling that I'm not being told the whole truth... and I know I'm not. It's too simple. It's too easy to consume. It makes too many hidden and contradictory assumptions that can be teased out quickly if someone is looking. While I wanted to like this book, its a case of how stuff with AI in the name sells, and it cheapens a lot of the work that others have worked towards on this subject. While someone who isn't trained in this might walk away with a sense of gained knowledge, it doesn't well introduce how critical AI research is being done well. This book seems to shut down interest more than raise it and leave open questions. This is a book that only someone trained in critical communication studies could agree with in summary, but it is condescending to everyone else that it's written for.
This is a book that required a lot more nuance and should have trusted its audience to understand more than the consumable drama and oversimplifications this book narrates. While good things are here, they can only be interpreted with clarity by someone who already knew those things. This is a common problem I've seen in "accessible academia": don't trust the public with the clear and fair investigated truth-claims; trust them instead with patchworked, token truths sewn together with drama.
Introduction: Let's define the entirety of AI as stupid, dumb, and mean calculations.
Not only does Crawford never even attempt to give a sufficient definition of what an Artificial Intelligence is, she mostly just calls it names. Granted, she loosely ties it to common usages and popular interpretations of how AI has been used, but she never gives it a clear definition, or explains broadly how it is different from other kinds of computation. But she only uses these straw person positions of what AI are in order to say we don't need one. This is intentional. It enables a particular perspective of AI that does not have to be very nuanced, and it allows her to make all sorts of arguments without justification. It essentially makes it really easy to read this book and think you learned something about AI when really what you are about to learn is the history of basically every technology invented since the 1800s. This book was praised as being the book of the moment and an instant classic. This book might as well be about the printing press. The arguments of its devastation to society would hold just the same at the time of its industrialization.
Chapter 1 Earth: Mining is bad. Computers use parts that are mined.
Yes, Crawford initiates with an ethnographic style that she had a paid field trip to visit a mining site, yet she never seems to use any ethical ethnographic practices. "I'm here to learn what AI is made from." Okay, but she doesn't use any of the information she got there that she couldn't have gotten from a Google map search and literal hundreds of years of news reports about the hazards and labor issues in mining. As best as I can tell, she doesn't quote a single person from the site, she doesn't help anybody there, and most of her arguments aren't about AI specifically but are about computers we all use everyday. It's almost like this chapter intends to say AI did this when literally the entire world of computing plays a part. Your smartphone is a part of this. The entire chapter (and even the book as a whole) is riddled with narrative holes such as connecting all of statistics with eugenics without further explanation or motivation for reason in a chapter about mineral mines. Am I to believe that anyone who computes a t-statistic is connected to a goal of eugenics oriented computation? The author is not clear on this but she certainly wants the reader to uncritically follow that narrative.
Chapter 2 Labor: There's a history of labor exploitation in computation.
Uh, yes. There's a history of labor exploitation. How is this unique to AI? It isn't. Why is this a book about AI again? Why are we ignoring more than 200 years of labor literature on this exact subject?
Chapter 3 Data: Models often use data abstracted from their origins which carry information we should address.
This is likely the best chapter in the book. For the first time, the author addresses something interesting about machine learning and AI. The authors argument that we use data that are deeply personal to some people (specifically in the use of mug shots used for crime prediction or facial recognition), because it strips those people of their humanity through abstraction and sets a norm of being able to use people's image data without asking them. However while this argument is obvious, the way in which she makes it doesn't address data itself as the chapter proclaims. It addresses models. Otherwise, why not critique the Census while we're at it? Is that not basically the same? No, because according to the author it's not about the data itself (except that the image data make her feel more strongly about this particular case) but it's about dehumanizing modeling techniques. But later, it's not about modeling techniques because it's about the fact that these data contain histories of things these models don't model. This story is not well unpacked. She cannot argue both models and data are steps after each other in dehumanizing people simultaneously. This is where Crawford's refusal to directly explain how AI is defined and works mechanistically hurts her the most. WIthout this nuance, it's not clear where the problem is. And so for the remainder of the book, we talk about government programs without clarity of what was built at a datafied level or at a modelled level, and so a casual reader would have to already know how AI works to actually learn the ethics here. The book does not give anyone the chance to reflect on their own relationship to data and programming. What practical ethical consideration should I take away from this? Despite critiquing induction/deduction in computation, the implication is to induct that all things about social data trace back to unnecessary and dehumanizing abstractions. But all a data scientist can do is say, "Yes, this case is sad," and be educated enough to know that this isn't true in every case. But Crawford implies again and again that this basically is the history of AI and datafication.
Chapter 4 Classification: Ontologies have implications.
Now she turns on the laborer she earlier claimed was exploited, and says that when we use cheap labor to classify stuff, we end up with racist outcomes and the degradation of linguistic principles. The author remains conservative to following an epistemology that she doesn't explain about language and classification. This is nothing but an elitist trope that classification experts should do classification since data scientists aren't willing to ethically clean their own datasets. While there are good points here about the harms of classification, just read Bowker and Star instead. It seems more reasonable for me to say this is a mix of (1) inductions without critique are bad, (2) some people are overtly problematic in their speech and classification (and in fact classification itself is problematic ontologically), and (3) there simply isn't enough labor and financial ability for scientists on minimal budgets to do this work. In summary what Crawford's policy here is, don't let pedestrian people datafy things because they will muddy up your data. This feels elitist about interpreting how MTurkers work, and unfair to researchers who cannot meet publication requirements on their own. It also doesn't account for the argument that researchers muddy their own data all the time, computational data or not. "Expert" checking doesn't solve this on it's own. Crawford is turning her back on her own labor chapter. Researchers can't work in isolation for most of academia in order to answer the questions posed to them, and the use of crowdsourced data (while problematic and should be addressed more by the standards of the Labor chapter) is not necessarily lesser quality data just because they are unaware of the academic training. If anything, an oppositional reading of this section teaches us that researchers can get better at cleaning crowdsourced data with experience of what to look for or can use that same data for a different purpose related to the biases it contains. Unfortunately, instead of actively looking for ways to continue valuing the labor of our peers, she suggests we induct that all datafication processes using classification are problematic necessities of AI because nobody is looking at datasets they use, which is laughable. While its been described this way, clearly somebody eventually looked at all the datasets Crawford claims weren't looked at because it's those researchers trying to correct these problems that gave her the historical data to write the chapter.
Chapter 5 Affect: Reading people's emotions from their faces is not a science.
While the simple argument here is true: we can't look at a person's face and know their feelings, this chapter badly butchers its theory. Firstly, while her main antagonist, Elkman, actually did attempt to do this and its still used in computational research, the way she frames Tomkins is horrendous. He argues that every body has an affective interpretation, not necessarily that all bodies essentially are the same in how we interpret their affects. Secondly, Tomkins does not fixate on singular human affects as the only individuals of interest. Tomkins is a post-human psychologist. Crawford reduces Affect theory to simple emotions of individuals (as do most communication scholars actually; it's kind of a meme at this point.) In total, this is probably one of the worst readings of Tomkins I've ever seen. There's no "affect theory" here in this interpretation. This is just badly interpreted psychology theory. For anyone who wants to check this out, look up Eve Sedgewick's Tomkins reader _Shame and its Sisters_. For affect theory more broadly, Spinoza's _Ethics_ and Deleuze make great starting points also. Even a brief overview from experts in affect should convince you of how badly Crawford misunderstands affect theory.
Chapter 6 State: Lots of computational tools used to police and track people online are AI tools that create logics to hurt "others" as a part of datafied "precision."
Again, we knew this. She briefly discusses Benjamin Bratton's book (which this chapter is merely a summary of), but also _We Are Data_ describes this very carefully. Crawford supposedly, in hacker-like style, had access to all of Snowden's data release and still the best she could do with it was repeat the same arguments others did. Yes, AI is used to track people and occasionally decide to kill them. Yes, this is dehumanizing to reduce humans to numbers. Yes, this is what new nationalist racism looks like. But more could have been said that summarizing the work of others as a part of having pulled a couple of slides from Snowden's files and putting them in a book.
This book is simple to interpret because it's written to be devastating and have the sense of absolute completion. While I would like to say that I agree with most of it, half of the stuff that I do agree with isn't new here even though it's written stylistically to look like she's uncovering conspiracies like some Jason Borne-level academic. The other half of stuff that I agree with ties itself to framings that might be more problematic to academic standards and labor ethics. This does not even include that I think that most of this felt really disingenuous. The narrative turns from the ethnographic writing style to carefully selected historical tangents you can find on Wikipedia that have nothing to do with the immediate ethnographic framing give me the feeling that I'm not being told the whole truth... and I know I'm not. It's too simple. It's too easy to consume. It makes too many hidden and contradictory assumptions that can be teased out quickly if someone is looking. While I wanted to like this book, its a case of how stuff with AI in the name sells, and it cheapens a lot of the work that others have worked towards on this subject. While someone who isn't trained in this might walk away with a sense of gained knowledge, it doesn't well introduce how critical AI research is being done well. This book seems to shut down interest more than raise it and leave open questions. This is a book that only someone trained in critical communication studies could agree with in summary, but it is condescending to everyone else that it's written for.
This is a book that required a lot more nuance and should have trusted its audience to understand more than the consumable drama and oversimplifications this book narrates. While good things are here, they can only be interpreted with clarity by someone who already knew those things. This is a common problem I've seen in "accessible academia": don't trust the public with the clear and fair investigated truth-claims; trust them instead with patchworked, token truths sewn together with drama.
October 2, 2024
A wide ranging book focussing on the dark sides of techno utopianism, and how AI requires vast amounts of unwillingly provided input, cheap labor and energy to function
What remains is a persistent asymmetry of power, where technical systems maintain and extend structural inequality, regardless of the intention of the designers
Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence is in some ways a chilling book, at times quite winding but definitely clear eyed on many of the negative aspects lurking under the surface of Silicon Valley’s belief in the malleability of the world.
What is being optimized and for whom is a recurring question Kate Crawford asks, while touching upon the many parts of the economy that enable the rise of AI.
The scope is more broadly a critique of capitalism than really tackling AI as such, and takes a more sociological angle than I initially expected. In contrast to similar books like The Mushroom at the End of the World: On the Possibility of Life in Capitalist Ruins, the narrative is supported with examples and the goal of critique does not obscure analysis.
In the long run AI is the only science some aficionados say, but Crawford zooms in much more on the thesis AI as an extractive industry. Extraction, power and politics form the core of the book.
Material extraction is first examined, with the extracting of one tonne of rare earth minerals requiring not just displacing 499 tons of earth (since the purity is just 0.2% in nature) but also 17.000 liters of toxic acids and 1.000 kilo of radioactive waste. The extraction via deep mines of gold in California being mirrored in skyscrapers in the sky, is an imagine the author uses. Global aviation and server emissions of AWS, Azure and Google cloud being roughly the same, with Chinese datacenters being 73% powered by coal plants. Some examples don't really hit the mark, I for instance don’t see how the container example has much of a link with AI, but this overall chapter was definitely interesting in reassessing the tech sector as inherently clean and low in emissions.
Labor extraction is next, and starts with a meditation on the concept of a panopticon stemming from a wish to better monitor workers. The problematic implication that control is equivalent to intelligence would continue to shape the AI field for decades.
Crawfurd introduces a concept of Potemkin AI, being an AI application being heavily reliant on low paid gig economy workers to function appropriately (take the example of content moderators).
Reducing the link between labor and value creation, making the bargaining power of tech employers larger and larger is another topic, including we as users training Google’s AI image recognition via ReCatchpa, which ironically forces us to prove we are human.
Campaigns like We are all techworkers -to view the tech supply chain and its reach more broadly, seem sympathetic, but I wonder on the viability of these initiatives versus the massive scale of big tech.
Data is the next topic, with many companies espousing a regime of There is no data like more data
IBM is an example, with the company using an anti-trust case against it of 13 years and 100 million words as a basis of its speech recognition software.
Further down the line the Enron corpus, a major fraud, is used for natural language processing.
This all feels rustic versus the 315 million photos on Meta’s platforms and 500 million tweets are published every day in 2019.
Extraction is justified if it comes from a primitive and unrefined source - on the narrative of data as the new oil.
These sources of data are worrying, as a 2014 model of Amazon shows. The model lead it to reinforce hiring men and masculine language in resumes.
In this sense artificial intelligence is a register of power, with over 95% of AI models treating gender as binary and inalterable, reinforced by classifying people as objects in datasets which are used to train AI.
Affect, or emotion recognition, is a new frontier.
Job applicants being scored on their facial expressions for suitability for a job is disturbing, or Palantir using its software, with a point system, makes police move into intelligence roles, reminiscent to Minority Report in terms of self reinforcement.
Vigilant system selling mass surveillance of license plates to local governments, with these kind of companies even being contracted to collect fines against a 25% fee for police.
Amazon using its Ring cameras to monitor its delivery staff and signing agreements with 600 local police departments to provide them with access to recordings and providing credits for subscribers to its Neighbors app.
We kill people based on metadata e.g. we track them, you whack them, as a vendor of "killer AI" says on a military conference, is not that far away.
Even seemingly benign initiatives, like IBM creating a hypothetical “terrorist credit score” for Syrian refugees to prevent backlash against immigrants, lead to discrimination and prosecution of those already in a very weak position. Then we have 60.000 Michigan citizens being penalized by AI, taking away their social benefits based on model outcomes, costing much more than the "savings" realised, and eroding trust in society.
A grim and thought provoking book, on how (as Margaret Atwood already said so poignantly in the The Handmaid's Tale) Better is never better for everyone.
What remains is a persistent asymmetry of power, where technical systems maintain and extend structural inequality, regardless of the intention of the designers
Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence is in some ways a chilling book, at times quite winding but definitely clear eyed on many of the negative aspects lurking under the surface of Silicon Valley’s belief in the malleability of the world.
What is being optimized and for whom is a recurring question Kate Crawford asks, while touching upon the many parts of the economy that enable the rise of AI.
The scope is more broadly a critique of capitalism than really tackling AI as such, and takes a more sociological angle than I initially expected. In contrast to similar books like The Mushroom at the End of the World: On the Possibility of Life in Capitalist Ruins, the narrative is supported with examples and the goal of critique does not obscure analysis.
In the long run AI is the only science some aficionados say, but Crawford zooms in much more on the thesis AI as an extractive industry. Extraction, power and politics form the core of the book.
Material extraction is first examined, with the extracting of one tonne of rare earth minerals requiring not just displacing 499 tons of earth (since the purity is just 0.2% in nature) but also 17.000 liters of toxic acids and 1.000 kilo of radioactive waste. The extraction via deep mines of gold in California being mirrored in skyscrapers in the sky, is an imagine the author uses. Global aviation and server emissions of AWS, Azure and Google cloud being roughly the same, with Chinese datacenters being 73% powered by coal plants. Some examples don't really hit the mark, I for instance don’t see how the container example has much of a link with AI, but this overall chapter was definitely interesting in reassessing the tech sector as inherently clean and low in emissions.
Labor extraction is next, and starts with a meditation on the concept of a panopticon stemming from a wish to better monitor workers. The problematic implication that control is equivalent to intelligence would continue to shape the AI field for decades.
Crawfurd introduces a concept of Potemkin AI, being an AI application being heavily reliant on low paid gig economy workers to function appropriately (take the example of content moderators).
Reducing the link between labor and value creation, making the bargaining power of tech employers larger and larger is another topic, including we as users training Google’s AI image recognition via ReCatchpa, which ironically forces us to prove we are human.
Campaigns like We are all techworkers -to view the tech supply chain and its reach more broadly, seem sympathetic, but I wonder on the viability of these initiatives versus the massive scale of big tech.
Data is the next topic, with many companies espousing a regime of There is no data like more data
IBM is an example, with the company using an anti-trust case against it of 13 years and 100 million words as a basis of its speech recognition software.
Further down the line the Enron corpus, a major fraud, is used for natural language processing.
This all feels rustic versus the 315 million photos on Meta’s platforms and 500 million tweets are published every day in 2019.
Extraction is justified if it comes from a primitive and unrefined source - on the narrative of data as the new oil.
These sources of data are worrying, as a 2014 model of Amazon shows. The model lead it to reinforce hiring men and masculine language in resumes.
In this sense artificial intelligence is a register of power, with over 95% of AI models treating gender as binary and inalterable, reinforced by classifying people as objects in datasets which are used to train AI.
Affect, or emotion recognition, is a new frontier.
Job applicants being scored on their facial expressions for suitability for a job is disturbing, or Palantir using its software, with a point system, makes police move into intelligence roles, reminiscent to Minority Report in terms of self reinforcement.
Vigilant system selling mass surveillance of license plates to local governments, with these kind of companies even being contracted to collect fines against a 25% fee for police.
Amazon using its Ring cameras to monitor its delivery staff and signing agreements with 600 local police departments to provide them with access to recordings and providing credits for subscribers to its Neighbors app.
We kill people based on metadata e.g. we track them, you whack them, as a vendor of "killer AI" says on a military conference, is not that far away.
Even seemingly benign initiatives, like IBM creating a hypothetical “terrorist credit score” for Syrian refugees to prevent backlash against immigrants, lead to discrimination and prosecution of those already in a very weak position. Then we have 60.000 Michigan citizens being penalized by AI, taking away their social benefits based on model outcomes, costing much more than the "savings" realised, and eroding trust in society.
A grim and thought provoking book, on how (as Margaret Atwood already said so poignantly in the The Handmaid's Tale) Better is never better for everyone.
January 11, 2026
Atlas of AI: Power, Politics, and the Planetary Costs of Artificial Intelligence - Kate Crawford
طوال الأسبوع الماضي هاج تويتر وماج بسبب إضافة تحديث جديد
للذكاء الاصطناعي في التطبيق حيث يمكنه تعديل أيّ صورة بمجرد أن
يطلب منه أي شخص ذلك، وانتشرت تعديلات لصور فتيات ونساء
وأطفال معدلة بدون إذنهم؛ تم تعديلها إلى صور بملابس البحر
وغيرها -وهنا أنا ألطف من الكارثة التي وقعت وتم تغيير الأمر لاحقًا ليقتصر فقط على
على أصحاب الحسابات الموثقة-المعروفين بذوي العلامة الزرقاء.. هذه إحدى المشاكل الجديدة والمصاحبة لاستخدام الذكاء الاصطناعي .
ا
ورد في الكتاب: في العام 2014؛ أنشأت شركة أمازون برنامج للتوظيف
قال القائمون عليه: "هذا برنامج ترسل له 100 سيرة ذاتية ويختار أفضل
خمسة."
بعد تشعيل النموذج الأولي ظهرت مشكلة غريبة؛ بات البرنامج
يستثني النساء تلقائيًا، لأن النماذج الأساسية التي قام عليها وتعلمها
كانت قائمة ومؤسسة على سير ذاتية وقصص نجاح لرجال..
تعلم الذكاء الاصطناعي هذا التحيز واستخدمه لاستثناء النساء× استثناء ليس فقط كجندر بل كلغة كذلك؛ لأن ما قرأه في سيرهن وقصصهن لم يشبه أيًا مما تغذى عليه.
مثل هذه المعضلة وغيرها تناقش سريعًا في فصول هذا الكتاب.. التكاليف الخفيّة للذكاء الاصطناعي من التحيز والطبقيّة والخصوصية والحرية وكيف يتغذى على جهود عمالة رخيصة وثروات أمم مستعمرة ويسرقها وتدمير الكوكب بتلويث البيئة بمخلفات هذه الصناعة.
أكثر ما أحببته في الكتاب هو تسميته لكل شيء بمسماه الحقيقي، بلا
تلطيف كما يحدث دائمًا.
للأسف الكتاب نشر بـ 2020 ولذا أصبح كل المطروح قديمًا في مجال
سريع التطور مثل الذكاء الاصطناعي فيما عدا مايخص سرقة الليثيوم
وغيره من الموارد الطبيعية وتلويث البيئة، كان هذا عيب الكتاب لكن دعونا لا ننسى أنه مجال سريع التغيّر- الأسبوع الماضي قرأت مصطلح جديدًَا-الاقطاعية التكنولوجية؛ ما يعني هذه الكيانات التي تسيطر اليوم على عالم الانترنت...
طوال الأسبوع الماضي هاج تويتر وماج بسبب إضافة تحديث جديد
للذكاء الاصطناعي في التطبيق حيث يمكنه تعديل أيّ صورة بمجرد أن
يطلب منه أي شخص ذلك، وانتشرت تعديلات لصور فتيات ونساء
وأطفال معدلة بدون إذنهم؛ تم تعديلها إلى صور بملابس البحر
وغيرها -وهنا أنا ألطف من الكارثة التي وقعت وتم تغيير الأمر لاحقًا ليقتصر فقط على
على أصحاب الحسابات الموثقة-المعروفين بذوي العلامة الزرقاء.. هذه إحدى المشاكل الجديدة والمصاحبة لاستخدام الذكاء الاصطناعي .
ا
ورد في الكتاب: في العام 2014؛ أنشأت شركة أمازون برنامج للتوظيف
قال القائمون عليه: "هذا برنامج ترسل له 100 سيرة ذاتية ويختار أفضل
خمسة."
بعد تشعيل النموذج الأولي ظهرت مشكلة غريبة؛ بات البرنامج
يستثني النساء تلقائيًا، لأن النماذج الأساسية التي قام عليها وتعلمها
كانت قائمة ومؤسسة على سير ذاتية وقصص نجاح لرجال..
تعلم الذكاء الاصطناعي هذا التحيز واستخدمه لاستثناء النساء× استثناء ليس فقط كجندر بل كلغة كذلك؛ لأن ما قرأه في سيرهن وقصصهن لم يشبه أيًا مما تغذى عليه.
مثل هذه المعضلة وغيرها تناقش سريعًا في فصول هذا الكتاب.. التكاليف الخفيّة للذكاء الاصطناعي من التحيز والطبقيّة والخصوصية والحرية وكيف يتغذى على جهود عمالة رخيصة وثروات أمم مستعمرة ويسرقها وتدمير الكوكب بتلويث البيئة بمخلفات هذه الصناعة.
أكثر ما أحببته في الكتاب هو تسميته لكل شيء بمسماه الحقيقي، بلا
تلطيف كما يحدث دائمًا.
للأسف الكتاب نشر بـ 2020 ولذا أصبح كل المطروح قديمًا في مجال
سريع التطور مثل الذكاء الاصطناعي فيما عدا مايخص سرقة الليثيوم
وغيره من الموارد الطبيعية وتلويث البيئة، كان هذا عيب الكتاب لكن دعونا لا ننسى أنه مجال سريع التغيّر- الأسبوع الماضي قرأت مصطلح جديدًَا-الاقطاعية التكنولوجية؛ ما يعني هذه الكيانات التي تسيطر اليوم على عالم الانترنت...
June 3, 2021
There is some good information buried in this book, but it's so hard to suss out with the repetitive and pedantic writing. The definition of AI appears to change from page to page, even from paragraph to paragraph and while that could be fine (because the author mentions that the definition of AI changes) it's not okay when the term is flexed to stand in for whatever topic they are arguing against. Sometimes AI means "technology" sometimes it means "the internet" sometimes it means "machine learning" and sometimes it means "computer programming" and so on. The argument loses impact when I can think of so many other nouns I could swap in with the same argument, such as "food production" or "garbage handling" or "housing". There was so little that seemed specific to AI, why did they pick that term to cluster their arguments around?
The layout was also frustrating. The whole book could have been reduced to the few pages in the end (right before the coda) where the actual intention of the previous chapters was laid out. It would have helped a lot to have that first so that I had some idea where the rambling previous 6 chapters were going and what the point of them was. Each chapter felt the same, just with different bad actors. Half of the books is references and bibliography, so just as soon as it starts pulling info together and discussing it it more analytically, the book ends. Reading in the notes that this was several presentations and articles that were pulled together and updated for the book that makes some sense but I think it would actually have been better served to include those as they were and write forwards and afterwords for them instead of trying to force them to be a coherent book.
But even with all those complaints, I could certainly put up with all that because the point of the book (technology is bad in ways many people don't realize), is one I'm quite sympathetic to. Unfortunately the book doesn't offer a way forward, or alternatives, or any examples of positive use of the technology and ways it is improving. There are brief dismissive mentions of regulation and awareness but they are brushed off as "the whole thing is rotten from the start so why bother just nuke it all from orbit and start over" but there's no suggestion of an alternative or what research is being done to replace what AI does.
The layout was also frustrating. The whole book could have been reduced to the few pages in the end (right before the coda) where the actual intention of the previous chapters was laid out. It would have helped a lot to have that first so that I had some idea where the rambling previous 6 chapters were going and what the point of them was. Each chapter felt the same, just with different bad actors. Half of the books is references and bibliography, so just as soon as it starts pulling info together and discussing it it more analytically, the book ends. Reading in the notes that this was several presentations and articles that were pulled together and updated for the book that makes some sense but I think it would actually have been better served to include those as they were and write forwards and afterwords for them instead of trying to force them to be a coherent book.
But even with all those complaints, I could certainly put up with all that because the point of the book (technology is bad in ways many people don't realize), is one I'm quite sympathetic to. Unfortunately the book doesn't offer a way forward, or alternatives, or any examples of positive use of the technology and ways it is improving. There are brief dismissive mentions of regulation and awareness but they are brushed off as "the whole thing is rotten from the start so why bother just nuke it all from orbit and start over" but there's no suggestion of an alternative or what research is being done to replace what AI does.
This entire review has been hidden because of spoilers.
November 1, 2022
In the “German Ideology”, Marx denounced the abstract, impersonal, ahistorical, and necessary pronouncements of metaphysics, philosophy, religion, law, economics, science, and so on - as power structures used by the capitalist system in its quest for domination and profit. These days it is quite fashionable to add gender, race, ecology, colonialism, externalities, war, along with a few others to the original economic class-category and to just repeat the Marxist critique to new fields like AI. Even if trite in this respect, this book is great in reminding us that AI is not some geek, idealist, objective, or wonderful fantasy - but the newest and most powerful power structure developed and employed these days.
December 30, 2025
So, in early twentieth-century Germany there was a horse called Clever Hans that seemed to ‘answer’ arithmetic questions by tapping his hoof. Psychologist Oskar Pfungst eventually figured out that Clever Hans was Copycat Hans.
Crawford explains that “the questioner’s posture, breathing, and facial expression would subtly change around the moment Hans reached the right answer, prompting Hans to stop there.”
In psychological terms, the Clever Hans Effect describes situations where a system appears to perform intelligently, but is actually exploiting unintended signals, correlations, or expectations supplied by humans or the experimental setup.
Ergo, Crawford uses the Clever Hans Effect to critique claims of AI intelligence, neutrality, or autonomy. Decolonial theory, which is surprisingly imbricated with AI, points out that the ‘emergence narratives’ surrounding AI reflect the ‘myth of modernity’.
Here, thinkers like Quijano and Mignolo have show how modernity is entangled with coloniality, to the extent that Mignolo called modernity the “darker side” of Western modernity. Crawford comments:
The big takeaway is that this idea of AI as being ethereal, mathematical algorithms in the cloud is absolutely not the case. In actual fact, the only way they work is by extracting vast amounts of data, human labor, and natural resources, which includes energy, water, and minerals. The book makes the case that AI is the extractive industry of the 21st century.
Crawford argues that AI is neither ‘artificial’ nor ‘intelligent’, but that the plethora of definitions it has generated frame how AI “will be understood, measured, valued, and governed.” The problem we are faced with currently is that “only continual technical progress matters, with no consideration of the computational cost of those approaches and their far-reaching impacts on a planet under strain.”
This is a measured and fascinating dive into the AI debate that is easily digestible and neither too abstract or technical. It extends beyond AI though to even look at space colonisation and frontier mining, described as “common corporate fantasies of tech billionaires.”
Crawford argues that this
underscores a fundamentally troubling relationship to Earth. Their vision of the future does not include minimizing oil and gas exploration or containing resource consumption or even reducing the exploitative labor practices that have enriched them. Instead, the language of the tech elite often echoes settler colonialism, seeking to displace Earth’s population and capture territory for mineral extraction.
Crawford explains that “the questioner’s posture, breathing, and facial expression would subtly change around the moment Hans reached the right answer, prompting Hans to stop there.”
In psychological terms, the Clever Hans Effect describes situations where a system appears to perform intelligently, but is actually exploiting unintended signals, correlations, or expectations supplied by humans or the experimental setup.
Ergo, Crawford uses the Clever Hans Effect to critique claims of AI intelligence, neutrality, or autonomy. Decolonial theory, which is surprisingly imbricated with AI, points out that the ‘emergence narratives’ surrounding AI reflect the ‘myth of modernity’.
Here, thinkers like Quijano and Mignolo have show how modernity is entangled with coloniality, to the extent that Mignolo called modernity the “darker side” of Western modernity. Crawford comments:
The big takeaway is that this idea of AI as being ethereal, mathematical algorithms in the cloud is absolutely not the case. In actual fact, the only way they work is by extracting vast amounts of data, human labor, and natural resources, which includes energy, water, and minerals. The book makes the case that AI is the extractive industry of the 21st century.
Crawford argues that AI is neither ‘artificial’ nor ‘intelligent’, but that the plethora of definitions it has generated frame how AI “will be understood, measured, valued, and governed.” The problem we are faced with currently is that “only continual technical progress matters, with no consideration of the computational cost of those approaches and their far-reaching impacts on a planet under strain.”
This is a measured and fascinating dive into the AI debate that is easily digestible and neither too abstract or technical. It extends beyond AI though to even look at space colonisation and frontier mining, described as “common corporate fantasies of tech billionaires.”
Crawford argues that this
underscores a fundamentally troubling relationship to Earth. Their vision of the future does not include minimizing oil and gas exploration or containing resource consumption or even reducing the exploitative labor practices that have enriched them. Instead, the language of the tech elite often echoes settler colonialism, seeking to displace Earth’s population and capture territory for mineral extraction.
November 14, 2021
if you're considering buying or reading this book, i encourage you to choose another title from the rapidly expanding genre given crawford's apparently widely known plagiarism.
i did not particularly enjoy 'atlas of ai', which reads more like a stream of disconnected anecdotes rather than a coherent whole. seriously, if you are on progressive tech twitter at all, there won't be anything new for you in here, i promise. it's not that the material in here isn't interesting or valuable, it's just that 'atlas of ai' highlights an enormous body of work from the sts and algorithmic fairness communities without performing any real synthesis.
i did not particularly enjoy 'atlas of ai', which reads more like a stream of disconnected anecdotes rather than a coherent whole. seriously, if you are on progressive tech twitter at all, there won't be anything new for you in here, i promise. it's not that the material in here isn't interesting or valuable, it's just that 'atlas of ai' highlights an enormous body of work from the sts and algorithmic fairness communities without performing any real synthesis.
July 3, 2022
Mandatory reading for all “tech dudebros” who think AI is the ultimate answer to life and the universe. Actually, mandatory reading for everyone.
January 4, 2022
This is probably my favorite of all the books tech-and-politics bookclub (even though I'm reading it half a year later whoops). I haven't yet found another book in this area with as much intersectional knowledge. Crawford has a robust background in tech journalism, and her acknowledgements further demonstrate her care in looping in experts from outside her own domain to paint a expansive picture of AI in our world at large, including the environmental impacts of computation, the near-invisible sociopolitical implications of the way ML practices format classification, and the impacts of machine decision making on the realm of human life.
Also bonus points for the drawings at the beginning of each chapter
[image error]
Also bonus points for the drawings at the beginning of each chapter
[image error]
November 12, 2021
I have been in the middle of this for so many months so honestly good to just finish a book LOL. I found the first two chapters (Earth and Labor, specifically the first bits about mining and the really detailed description of Mechanical Turk) really interesting but felt like, as the book went on, it was pretty general/I didn't gain a ton from it. Glad I read though
Read
May 16, 2021Accessible and engaging, with a compelling political agenda
January 10, 2025
Het boek belicht een belangrijk en doorgaans genegeerd aspect van AI; de vervuiling.
Energie, water, mineralen, AI is niet een cloud, niet disembodied en geen software. De ruggengraat van AI is de aarde en haar grondstoffen. Het heeft zijn wortels in de materiële realiteit, op de meest letterlijke manier.
Ook zonder menselijke arbeid zou AI niet functioneren. Het is niet zelfstandig.
Het boek behandelt de mensenrechtenschendingen in de mijnen en ook hoe het er aan toe gaat achter de schermen van Amazon. Arbeid in de 21ste eeuw verschilt structureel niet veel van de eeuwen ervoor.
De klok tikt en de camera filmt de werknemer.
AI is niet artificieel en ook niet intelligent, de hoeveelheid zieldodende taken die er door hoogopgeleide (en slecht betaalde) mensen gedaan wordt is om van te schrikken. Tijdens het uitvoeren van deze zieldodende taken moeten werknemers vaak doen alsof ze robots zijn. Er worden lange shiften gedraaid om de illusie te wekken dat de service van een platform 7/24 beschikbaar is. Bezos calls it “artificial artificial intelligence”, schaamteloos.
En dan hebben we het nog niet gehad over de data zelf waarmee machine-learning systemen getraind worden. De classificaties zijn problematisch, de afwezigheid van consent en het subject zijn dat ook.
De relatie tussen AI (of technologie in het algemeen) en militaire technieken (of destructie in het algemeen) lijken mij niet meteen te leiden tot een democratisering van AI. AI heeft een grond in het formaliseren wat niet te formaliseren valt, en het doet dit doordrongen door een kapitalistische logica.
(Schrijfstijl was niet echt aangenaam, heel droog, en onnodig lang, ze mist een soort helderheid, een scherpte die een boek als dit nodig heeft. Waar het filosofisch interessant werd begon ook het onnodig obscuur te worden. Één of twee vage zinnen over power politics op het einde van elk deeltje lieten me een beetje onverzadigd achter. Werk argumenten uit in de plaats van intuïties neer te pennen. En als je intuïtieve ideeën neerpent zonder dit met argumenten te onderbouwen, then have the heart to do it properly.)
Energie, water, mineralen, AI is niet een cloud, niet disembodied en geen software. De ruggengraat van AI is de aarde en haar grondstoffen. Het heeft zijn wortels in de materiële realiteit, op de meest letterlijke manier.
Ook zonder menselijke arbeid zou AI niet functioneren. Het is niet zelfstandig.
Het boek behandelt de mensenrechtenschendingen in de mijnen en ook hoe het er aan toe gaat achter de schermen van Amazon. Arbeid in de 21ste eeuw verschilt structureel niet veel van de eeuwen ervoor.
De klok tikt en de camera filmt de werknemer.
AI is niet artificieel en ook niet intelligent, de hoeveelheid zieldodende taken die er door hoogopgeleide (en slecht betaalde) mensen gedaan wordt is om van te schrikken. Tijdens het uitvoeren van deze zieldodende taken moeten werknemers vaak doen alsof ze robots zijn. Er worden lange shiften gedraaid om de illusie te wekken dat de service van een platform 7/24 beschikbaar is. Bezos calls it “artificial artificial intelligence”, schaamteloos.
En dan hebben we het nog niet gehad over de data zelf waarmee machine-learning systemen getraind worden. De classificaties zijn problematisch, de afwezigheid van consent en het subject zijn dat ook.
De relatie tussen AI (of technologie in het algemeen) en militaire technieken (of destructie in het algemeen) lijken mij niet meteen te leiden tot een democratisering van AI. AI heeft een grond in het formaliseren wat niet te formaliseren valt, en het doet dit doordrongen door een kapitalistische logica.
(Schrijfstijl was niet echt aangenaam, heel droog, en onnodig lang, ze mist een soort helderheid, een scherpte die een boek als dit nodig heeft. Waar het filosofisch interessant werd begon ook het onnodig obscuur te worden. Één of twee vage zinnen over power politics op het einde van elk deeltje lieten me een beetje onverzadigd achter. Werk argumenten uit in de plaats van intuïties neer te pennen. En als je intuïtieve ideeën neerpent zonder dit met argumenten te onderbouwen, then have the heart to do it properly.)
Read
September 28, 2022After a very long hiatus, I'm going to work through my backlog of Goodreads reviews, though I'll probably write far less than before.
I interviewed Kate Crawford about her book for the New Books Network!
Apple Podcasts: https://podcasts.apple.com/us/podcast...
Spotify: https://open.spotify.com/episode/7DmZ...
The thing I liked most about Atlas of AI is that it focused on the physical materials that make up our computers and AI systems. Computers are made of stuff. That stuff has to be mined and manufactured and transported around the world. AI software runs on massive servers. Those servers are physical behemoths that are sitting in the desert. Kate Crawford has the chutzpah to genuine consider what an AI system is, from top to bottom: the miners, the underwater internet cables, the shipping containers, the software engineers, the click workers who label data, the corporations that acquire data from users, etc. It's all here: https://anatomyof.ai/.
This is such an ambitious and frankly baller way of doing things. Imagine how much we could understand the workings of the world if we had diagrams like this for every technology. It really sends the mind through a major tizzy to think about the extent and scope of all of this. It really makes me think: what would it mean for Anatomy of AI to be "obvious"? Would every single person working in every job remotely related to the AI field have to, like, write blog posts about their role and how it relates to everything else? Would everyone need a Go Pro on their head blasting a live feed of everything they're doing all day? I'm not sure this would be useful. We'd be inundated in a morass of information. We would need a Kate Crawford to go sift through it all and figure it out for us. So there's no easy way out here.
Let me try to articulate this better. What I want is a deep understanding of how the world works. I get that things are complicated, but surely it's not too much to ask in the 21st century, that I can see, in full, all of the people and processes involved in a given technology. But this is actually very hard. Most of the work being done in the world is happening in private. There are no journalists or TV cameras in the vast majority of workplaces. A great deal of work involves extracting materials from the earth or coordinating the way groups of people operate. None of the people doing this work have any incentive to describe the complex networks of relationships between themselves and the people they work with. And if everyone somehow _did_ start to do that—say, everyone in the AI industry had a blog—there would be way too much information, an overwhelming panopticon that would become meaningless.
Given this state of affairs, what we need is armies of Kate Crawfords. I'm enlisting.
I interviewed Kate Crawford about her book for the New Books Network!
Apple Podcasts: https://podcasts.apple.com/us/podcast...
Spotify: https://open.spotify.com/episode/7DmZ...
The thing I liked most about Atlas of AI is that it focused on the physical materials that make up our computers and AI systems. Computers are made of stuff. That stuff has to be mined and manufactured and transported around the world. AI software runs on massive servers. Those servers are physical behemoths that are sitting in the desert. Kate Crawford has the chutzpah to genuine consider what an AI system is, from top to bottom: the miners, the underwater internet cables, the shipping containers, the software engineers, the click workers who label data, the corporations that acquire data from users, etc. It's all here: https://anatomyof.ai/.
This is such an ambitious and frankly baller way of doing things. Imagine how much we could understand the workings of the world if we had diagrams like this for every technology. It really sends the mind through a major tizzy to think about the extent and scope of all of this. It really makes me think: what would it mean for Anatomy of AI to be "obvious"? Would every single person working in every job remotely related to the AI field have to, like, write blog posts about their role and how it relates to everything else? Would everyone need a Go Pro on their head blasting a live feed of everything they're doing all day? I'm not sure this would be useful. We'd be inundated in a morass of information. We would need a Kate Crawford to go sift through it all and figure it out for us. So there's no easy way out here.
Let me try to articulate this better. What I want is a deep understanding of how the world works. I get that things are complicated, but surely it's not too much to ask in the 21st century, that I can see, in full, all of the people and processes involved in a given technology. But this is actually very hard. Most of the work being done in the world is happening in private. There are no journalists or TV cameras in the vast majority of workplaces. A great deal of work involves extracting materials from the earth or coordinating the way groups of people operate. None of the people doing this work have any incentive to describe the complex networks of relationships between themselves and the people they work with. And if everyone somehow _did_ start to do that—say, everyone in the AI industry had a blog—there would be way too much information, an overwhelming panopticon that would become meaningless.
Given this state of affairs, what we need is armies of Kate Crawfords. I'm enlisting.
March 25, 2023
This feels like a bad time to review this book, because I'm a bit oversaturated on the "let's discuss AI" moment that's happening right now, but I'll do it anyways.
Starting with the good things: I really liked the 'atlas' framing. An atlas is not a map that purports to objectively convey its territory, but rather a particular viewpoint or interpretation of reality, imbued with its own preoccupations and biases. It felt like a helpful shorthand for the thesis of Crawford's critiques: that technologists have been (willfully?) ignorant of the politics imbedded in their work, and thoughtless about the broader social and historical contexts they operate within. I also found this book to be a useful overview of some of the considerations and history that are frequently overlooked in discussions about AI, and I think I’ll dive into some of the authors that Crawford referenced.
Despite having some interesting content, this book bored me more than it should have, and I thought that the writing style was too academic. The other thing that bothers me in general about this specific genre of book is that it's uninspiring. I think I'm just tired of the "here's what people did that was bad" pattern, and what I actually want is a compelling vision for how we could build and use new technologies to make the world better.
Starting with the good things: I really liked the 'atlas' framing. An atlas is not a map that purports to objectively convey its territory, but rather a particular viewpoint or interpretation of reality, imbued with its own preoccupations and biases. It felt like a helpful shorthand for the thesis of Crawford's critiques: that technologists have been (willfully?) ignorant of the politics imbedded in their work, and thoughtless about the broader social and historical contexts they operate within. I also found this book to be a useful overview of some of the considerations and history that are frequently overlooked in discussions about AI, and I think I’ll dive into some of the authors that Crawford referenced.
Despite having some interesting content, this book bored me more than it should have, and I thought that the writing style was too academic. The other thing that bothers me in general about this specific genre of book is that it's uninspiring. I think I'm just tired of the "here's what people did that was bad" pattern, and what I actually want is a compelling vision for how we could build and use new technologies to make the world better.
Read
April 27, 2021Excerpt from book...
The Misreading of Human Emotion
https://www.theatlantic.com/technolog...
The Misreading of Human Emotion
https://www.theatlantic.com/technolog...
April 29, 2022
OK SO. I'm not saying this book was not very well researched and does not analyze critically and thoroughly facets of AI that most of us never really consider. It is. And it does. And I highlighted quite a lot of paragraphs where I loved the phrasing, the ideas being conveyed or the shocking examples provided.
BUT...
On multiple occasions, I caught myself comparing it to User Friendly: How the Hidden Rules of Design Are Changing the Way We Live, Work, and Play. Both these books set out to cover a complex topic exhaustively (AI and design, respectively). And both build their narratives around examples, events and characters. However, where User Friendly made me care about a topic I hadn't previously given any thoughts, Atlas of AI made me lose interest in a topic I was already invested in. Somehow Atlas of AI seems to fall short of creating an interesting storyline and really bringing to life the people and events it's talking about. At least for me. It could be the long sentences or the abstruse language (see what I did there?), but I had trouble connecting to it and getting through it, even though it's relatively short.
In sum, this book didn't reaaally do it for me. Have I not had to read it for a book club I'm organizing, I would have DNF'ed it.
BUT...
On multiple occasions, I caught myself comparing it to User Friendly: How the Hidden Rules of Design Are Changing the Way We Live, Work, and Play. Both these books set out to cover a complex topic exhaustively (AI and design, respectively). And both build their narratives around examples, events and characters. However, where User Friendly made me care about a topic I hadn't previously given any thoughts, Atlas of AI made me lose interest in a topic I was already invested in. Somehow Atlas of AI seems to fall short of creating an interesting storyline and really bringing to life the people and events it's talking about. At least for me. It could be the long sentences or the abstruse language (see what I did there?), but I had trouble connecting to it and getting through it, even though it's relatively short.
In sum, this book didn't reaaally do it for me. Have I not had to read it for a book club I'm organizing, I would have DNF'ed it.
May 1, 2021
This is called “Atlas of AI” for a good reason. We know technology is ubiquitous and we rely on it and enjoy it but somehow miss, or overlook the fact that we may all be victims as a consequence. Whether via environmental degradation, increasingly controlled workplaces, or abuse of our data. The stakes are huge, and not popularly understood and the gift of this work is to illuminate how the path ahead is now inevitably tied to degradation of our lives, and lifestyles, that in the case of climate change is existential, but does not end there. Eloquent and alarming this book is a classic for these troubled times.
January 11, 2022
One of the most sobering--verging on depressing--necessary books I've read in a while. The choice of Atlas in the title is precise: it's a compilation of maps at different scales and of different areas of the territory shorthanded as Artificial Intelligence. As Crawford points out, the term is often used to indicate a set of technical problems related to algorithms and computer engingeering, as if they were disconnected from the material realities of resource extraction, climate change, labor and political/economic exploitation. They aren't and to pretend otherwise is to engage in the game of roulette we've so gleefully accepted as "reality."
This is a necessary book for anyone who wants to have a realistic idea of the complexity of the problems facing anyone who cares about minor issues like justice and the sustainability of human life.
A slightly academic tinge to the prose at times, but Crawford's done her homework and it reads quickly. Crucial.
This is a necessary book for anyone who wants to have a realistic idea of the complexity of the problems facing anyone who cares about minor issues like justice and the sustainability of human life.
A slightly academic tinge to the prose at times, but Crawford's done her homework and it reads quickly. Crucial.
packed-away
November 5, 2024Abandoned after 30%
The book might as well be called Power, Politics, and the Planetary Costs of Capitalism, or Power, Politics, and the Planetary Costs of Petroleum, of Cars and Airplanes, of Global Supply Chain, or of any other modern industrial element. I thought it was common sense that data centers require a very large amount of electricity to run. Indeed the power grid is not green, or not green enough yet, but making the power grid green is a part of the fight against climate change. Do people still need to be reminded that capitalism, since inception, disregards environmental costs, care little about long term impact on people, let alone animals and plants?
I am for AI regulation, but I have trouble following the author’s arguments. The book is not well organized. If you make too broad claims, your claims become generic and you lose the target.
The book might as well be called Power, Politics, and the Planetary Costs of Capitalism, or Power, Politics, and the Planetary Costs of Petroleum, of Cars and Airplanes, of Global Supply Chain, or of any other modern industrial element. I thought it was common sense that data centers require a very large amount of electricity to run. Indeed the power grid is not green, or not green enough yet, but making the power grid green is a part of the fight against climate change. Do people still need to be reminded that capitalism, since inception, disregards environmental costs, care little about long term impact on people, let alone animals and plants?
I am for AI regulation, but I have trouble following the author’s arguments. The book is not well organized. If you make too broad claims, your claims become generic and you lose the target.
May 6, 2021
Enjoyed it over all. I read it in a breeze and learned a lot. Accessible, informative, synthetic and encompassing. Chapters on labor, classification and the earth were especially informative.
It did leave me wondering a) whether the book buys into some of the bullshit that it is supposedly trying to demystify b) if what is called ai now is just a continuation and extension of colonialism and capitalism (which the historical narrative seems to imply) than why not just call it that? What's "AI" about AI apart from being a marketing gimmick?
It did leave me wondering a) whether the book buys into some of the bullshit that it is supposedly trying to demystify b) if what is called ai now is just a continuation and extension of colonialism and capitalism (which the historical narrative seems to imply) than why not just call it that? What's "AI" about AI apart from being a marketing gimmick?
January 30, 2025
*4,5 ⭐️ Super spannend und interessant. Mein Mund hat mehrmals offen gestanden, weil ich nicht glauben konnte, was ich da lese. Ein absolutes Must Read in der heutigen Zeit. Ich fand’s nur manchmal schwer zu erkennen, welche der Dinge heute nicht mehr stattfinden.
October 27, 2025
Klare These - KI ist weder künstlich, noch intelligent und steht im Dienste mächtiger ökonomischer und politischer Akteure. Wird im Hinblick auf verschiedene Bereiche belegt (Ausbeutung von Rohstoffen und Arbeitskraft, Extraktion und Klassifizierung von Daten und Affekten ect.). Ein schöner Überblick mit illustrativen Beispiele, bei dem ich mit manchmal mehr Tiefe gewünscht hätte.
March 3, 2023
"Refusal requires rejecting the idea that the same tools that serve capital, militaries, and police are also fit to transform schools, hospitals, cities, and ecologies, as though they were value neutral calculators that can be applied everywhere."
April 3, 2025
377 pages, plus de 50 pages de bibliographie et de notes. Donc près de 300 pages, six chapitres abordant les impacts de l’IA sur notre environnement (démesurés et brutaux), le monde du travail (surveillance, impartition), les données (nos droits et l’appétit sans fin des AI), la classification (les biais, les banques de données d’apprentissage, etc), les influences (les origines, les études, les clients et les promoteurs de l’IA), les gouvernements (l’utilisation des IA et les dérives possibles). Passionnant et troublant.
Des bases solides qui me permettre d’aborder l’IA avec un esprit curieux et sérieux
Des bases solides qui me permettre d’aborder l’IA avec un esprit curieux et sérieux
September 27, 2021
This is an important book that should be read not only by AI/ML practitioners but also by Crypto and in general anybody working on cloud and internet technologies.
Although well written, it is a bit depressing how Craword describes the hidden costs, not only of energy-wise, but also in terms of labor (mining, production of hardware and generation of training sets) as well as the consequences of the application of these models (discrimination, bias, misrecognitions).
At the end of the day, many of these AI models are simply amplifiers of current oppressive systems, and they only further the unjustices that has existed always.
The only section I missed in this book was about how to fix all these issues: I don't think anybody has a clear idea of what is the solution. Crawford provides a few general ideas that seems to go in the right direction, but they are only that, general policies that may work.
Nevertherless, I think this is a fundamental book for everybody who works with a computer.
Although well written, it is a bit depressing how Craword describes the hidden costs, not only of energy-wise, but also in terms of labor (mining, production of hardware and generation of training sets) as well as the consequences of the application of these models (discrimination, bias, misrecognitions).
At the end of the day, many of these AI models are simply amplifiers of current oppressive systems, and they only further the unjustices that has existed always.
The only section I missed in this book was about how to fix all these issues: I don't think anybody has a clear idea of what is the solution. Crawford provides a few general ideas that seems to go in the right direction, but they are only that, general policies that may work.
Nevertherless, I think this is a fundamental book for everybody who works with a computer.
June 11, 2021
This is a fantastic introduction to the foundations of AI. I loved it all, but it felt like it was building toward something that never came.
This is the type of book that sits alongside books like Race After Technology, Don't Be Evil, Ruined by Design and Blockchain Chicken Farm in that it can change the way you actually process the world, and the motivations, goals and methods of the leaders of today's tech CEOs and their apologists.
The destructive potential of large tech companies, when enabled by spineless or witless governments is terrifying and needs to be checked. This book makes a clear case as to why, and raises plenty of alarms about what other unintended consequences might arise if private companies continue to dominate tech - in ways that expand the surveillance and security powers of the state in frightening ways.
Highly recommend. I just with the Conclusion and Coda sections were expanded and explored further.
This is the type of book that sits alongside books like Race After Technology, Don't Be Evil, Ruined by Design and Blockchain Chicken Farm in that it can change the way you actually process the world, and the motivations, goals and methods of the leaders of today's tech CEOs and their apologists.
The destructive potential of large tech companies, when enabled by spineless or witless governments is terrifying and needs to be checked. This book makes a clear case as to why, and raises plenty of alarms about what other unintended consequences might arise if private companies continue to dominate tech - in ways that expand the surveillance and security powers of the state in frightening ways.
Highly recommend. I just with the Conclusion and Coda sections were expanded and explored further.
April 18, 2023
This book had a lot of great information, however I feel Crawford lost her credibility when she spoke on facial recognition technology and how trans and non-binary people were being "misidentified," when in fact the AI is simply correctly sexing them. Trans and non-binary people still have a subjective biological SEX. It measures the human face and categorizes according to male or female and does this with great accuracy. To claim AI is "wrong" for not taking into account a person's personal self-idenity is unscientific and absurd. Just as AI cannot see a person's religion or political leanings in their face, it cannot see a person'spersonal gender ideology. It sees facts, it interprets that into data and can therfore correctly determine a person's sex (which is immutable), and it does this with great accuracy, regardless of the person's subjective self-identity. This take lacked scientific integrity and stood out like a sore thumb in a book based on data and facts.
May 27, 2021
I loved this book so much. I have been frustrated by many tech analyses, because it feels like writers don't quite get how things work in practice, and therefore have a lot of naive ideas for root problems or solutions. That is not the case here.
Crawford instead contextualizes AI into historical environmental destruction, racism, colonialism (etc), things that we all now understand to be bad. She then paints a picture of how the same destructive assumptions we made then, show up in today's practices.
Truly insightful! Everybody who works in tech needs to read this book!
Crawford instead contextualizes AI into historical environmental destruction, racism, colonialism (etc), things that we all now understand to be bad. She then paints a picture of how the same destructive assumptions we made then, show up in today's practices.
Truly insightful! Everybody who works in tech needs to read this book!
November 4, 2021
Demasiado descriptivo, los ejemplos están bien traídos y la estructura funciona pero no llega a entrar realmente en profundidad a analizar nada. Puede que yo esperase otra cosa y es verdad que este libro sí sirve para desmontar muchos mitos sobre la inteligencia artificial, pero creo que explica mejor todo lo que no es que todo lo que es realmente (lo que me gustaría entender mejor a mí). Además he leído que la autora es una imbécil que les robaba las ideas a investigadoras más jóvenes.
December 21, 2023
My main mistake on purchasing this book was not reading past the title "Atlas of AI." Seeing just that title, I figured that this would just be a survey covering the various technical aspects of machine learning techniques and their applications, maybe with a bit of history tied in. What I wound up reading instead happened to be perhaps the most pessimistic/apocalyptic view of AI that I've ever encountered.
The book aims to follow the building of AI systems from the ground-up, highlighting where people are being exploited along the way, all while offering little to no solution whatsoever to the problem apart from just stopping the development of AI altogether. Crawford devotes some time to things that also have almost nothing to do with AI in particular and are instead just general problems of capitalism: the mining of materials for making batteries (mainly for phones and electric cars, unless I'm mistaken), the use of under-paid workers in other countries for mining datasets, poorly chosen terminology (master-slave in clock synchronization techniques, git [side note: what does clock synchronization have to do specifically with AI?????]), the advancement of technology for military purposes. Another one of these criticisms was how classification tasks are using discretized information that abstracts away hidden meaning from things, which again seems like a far reach considering that there needs to be some form of discretization if it's going to work on a computer. Perhaps the biggest fault is that all of this criticism takes place with next to no praise for the actual advancements achieved by advancements in AI.
Still, I suppose some of the overarching arguments are valid, specifically the ones having to do with the chapter following machine learning techniques for surveillance or for machine learning techniques used in tasks for which model bias could have devastating consequences, perhaps the biggest of these being security screenings or welfare "credit scores."
With just the ideas presented in this book, I would probably have given this a 3-star rating. Alas, the writing itself was pretty poorly done, with random jumps to first-person narration that just really have no place in this story. Also, the combination of footnotes with citations is extremely annoying, as you never know whether its worth it to flip all the way to the back of the book or not.
TL;DR: Not what the title says. Missed opportunity.
The book aims to follow the building of AI systems from the ground-up, highlighting where people are being exploited along the way, all while offering little to no solution whatsoever to the problem apart from just stopping the development of AI altogether. Crawford devotes some time to things that also have almost nothing to do with AI in particular and are instead just general problems of capitalism: the mining of materials for making batteries (mainly for phones and electric cars, unless I'm mistaken), the use of under-paid workers in other countries for mining datasets, poorly chosen terminology (master-slave in clock synchronization techniques, git [side note: what does clock synchronization have to do specifically with AI?????]), the advancement of technology for military purposes. Another one of these criticisms was how classification tasks are using discretized information that abstracts away hidden meaning from things, which again seems like a far reach considering that there needs to be some form of discretization if it's going to work on a computer. Perhaps the biggest fault is that all of this criticism takes place with next to no praise for the actual advancements achieved by advancements in AI.
Still, I suppose some of the overarching arguments are valid, specifically the ones having to do with the chapter following machine learning techniques for surveillance or for machine learning techniques used in tasks for which model bias could have devastating consequences, perhaps the biggest of these being security screenings or welfare "credit scores."
With just the ideas presented in this book, I would probably have given this a 3-star rating. Alas, the writing itself was pretty poorly done, with random jumps to first-person narration that just really have no place in this story. Also, the combination of footnotes with citations is extremely annoying, as you never know whether its worth it to flip all the way to the back of the book or not.
TL;DR: Not what the title says. Missed opportunity.
Displaying 1 - 30 of 381 reviews