Start free trial
Trending New Genres Authors
EnglishEnglish
EspañolSpanish
简体中文Chinese
繁體中文Chinese (Traditional)
FrançaisFrench
DeutschGerman
日本語Japanese
PortuguêsPortuguese
ItalianoItalian
한국어Korean
РусскийRussian
NederlandsDutch
العربيةArabic
PolskiPolish
हिन्दीHindi
Tiếng ViệtVietnamese
SvenskaSwedish
ΕλληνικάGreek
TürkçeTurkish
ไทยThai
ČeštinaCzech
RomânăRomanian
MagyarHungarian
УкраїнськаUkrainian
IndonesiaIndonesian
DanskDanish
SuomiFinnish
БългарскиBulgarian
עבריתHebrew
NorskNorwegian
HrvatskiCroatian
CatalàCatalan
SlovenčinaSlovak
LietuviųLithuanian
SlovenščinaSlovenian
СрпскиSerbian
EestiEstonian
LatviešuLatvian
فارسیPersian
മലയാളംMalayalam
தமிழ்Tamil
اردوUrdu
Searching...
SoBrief
Artificial Intelligence

Artificial Intelligence

A Guide for Thinking Humans
A machine can defeat a Go champion and still misunderstand the world beyond the board.
by Melanie Mitchell 2019 336 pages
4.33
4k+ ratings
Listen
Immersive
V2.3
Amazon Kindle Audible
Try Full Access for 3 Days
Unlock listening & more!
Continue

Key Takeaways

Separate narrow competence from general intelligence before calling a machine smart.

One unsettling performance. At a Google meeting, AI pioneer Douglas Hofstadter described how a program called EMI had composed music resembling Chopin closely enough to fool music experts. He had expected creativity to resist mechanical imitation. The encounter unsettled him not because the program was conscious, but because a limited set of rules could produce something that sounded deeply human. The book uses this tension to distinguish narrow AI, which performs a defined task, from general intelligence, which flexibly combines abilities across many situations.

A test can mislead. In 2014, a chatbot posing as a thirteen-year-old fooled one-third of judges in brief conversations. That result did not establish broad understanding: the bot could dodge questions, change topics, and rely on a deliberately limited persona. A machine can succeed at a task without possessing the wider capacities people associate with intelligence.

Analysis

What's striking is how easily an impressive result becomes a claim about the whole mind. The distinction between task performance and general capacity resembles the difference between passing a single exam and being able to use knowledge in unfamiliar settings. A chatbot's ability to sustain a short exchange is real, but it is weak evidence about its grasp of the world. The same caution applies to human expertise: skill in one domain rarely proves competence everywhere. The challenge is not to dismiss specialized systems, but to describe precisely what they do and what their success does not establish.

Treat each AI boom as a promise to test, not a prophecy.

A recurring cycle. At the 1956 Dartmouth workshop, researchers proposed that machines could simulate every feature of intelligence and predicted major progress within a short time. Some early forecasts promised human-level capability within a generation. The field instead met hard limits, and funding fell after broad promises about language, reasoning, and robotics failed to materialize. Similar rounds of enthusiasm and disappointment came to be called AI springs and AI winters.

The pattern persists. Mitchell describes optimism around neural networks, expert systems, and later deep learning as part of this history. New methods attract investment and striking headlines; when performance falls short of the promises, support recedes. This history does not show that progress is impossible. It does show why claims about imminent general intelligence deserve evidence about actual capabilities, not just confidence, investment, or a memorable demonstration.

Analysis

The boom-and-bust pattern is more than a cautionary tale about overconfident scientists. It shows how public expectations, funding, and research priorities can reinforce one another. A field may pursue commercially valuable tasks while headlines imply that it is solving intelligence itself. That mismatch can obscure genuine advances and create needless backlash when broad promises fail. Skepticism is most useful when it asks for measurable claims and clear timelines, rather than treating every forecast as either prophecy or nonsense.

Faster chips do not automatically produce exponentially broader minds.

Exponential curves impress. Ray Kurzweil’s forecasts of near-term human-level AI draw heavily on exponential technological growth, including the historical doubling of components on computer chips every one to two years. A fable in the book makes the pattern vivid: doubling rice grains on successive squares of a chessboard seems manageable at first, then quickly exceeds what a kingdom can supply.

Progress has different speeds. The argument assumes that gains in computing power can unlock broader intelligence, including through brain reverse engineering. Mitchell notes that hardware has advanced rapidly, but software has not shown the same kind of steady exponential growth. Nor does processing power by itself explain how a machine would acquire common sense, learn from limited experience, or integrate abilities. Extrapolating one curve across these different challenges risks confusing faster computation with a solution to intelligence.

Analysis

What's useful here is the separation of a genuine trend from the conclusion someone wants to draw from it. Exponential growth can make change feel invisible until its effects become dramatic, but a curve in one domain does not guarantee a matching curve in another. Computing capacity is an input to AI progress, not a theory of cognition. The argument also leaves room for surprise: a successful new method could change the trajectory. The disciplined response is neither to extrapolate blindly nor to assume that present limits are permanent.

Deep learning's leap depended on data, computing power, and human setup.

A breakthrough with foundations. Deep learning trains neural networks with multiple hidden layers to detect patterns in data. In computer vision, these networks became especially effective after researchers could train them on huge collections of labeled images using faster computers. ImageNet supplied 1.2 million training pictures across a thousand categories. In 2012, AlexNet, a convolutional neural network, raised the competition's top-five accuracy from 74 percent to 85 percent.

Learning still needs labor. ImageNet's pictures required human workers to check and label them, and researchers had to design the network and tune its settings. Supervised learning means that a system learns from examples paired with human-provided answers. The network adjusts its internal numerical weights to reduce its mistakes. The result can be powerful, but describing it as a machine teaching itself hides the human work that makes the training possible.

Analysis

The story corrects a common misconception: data-driven systems are not free of human judgment. People select examples, define categories, choose what counts as success, and adjust the machinery. This resembles scientific measurement, where instruments extend observation but still depend on choices about what to measure. The performance gains are substantial, yet the labor and design decisions are part of the technology, not incidental scaffolding. That matters when assessing cost, accountability, and whose assumptions become embedded in a system.

Demand failure tests alongside benchmark scores.

A score has a scope. By 2017, the leading ImageNet system reached 98 percent accuracy under the competition's top-five measure: its correct label only had to appear somewhere among five guesses. Its top-one accuracy was about 82 percent. Neither number guaranteed that the system recognized objects as people do. A network trained to spot animals, for example, partly relied on blurred backgrounds because those often appeared in animal photographs.

Errors can be strange. An adversarial example is an input altered in a way that looks unchanged to people but can make a neural network confidently misclassify it. Researchers demonstrated such errors in images, including traffic signs. Training data can also carry social biases: the book reports that face-recognition systems made more errors for some racial groups, and cites a test in which Amazon's system mismatched 28 members of Congress with people in an arrest database. High benchmark scores do not settle questions of reliability or fairness.

Analysis

Benchmarks are valuable because they make comparison possible, but they are measurements of designed tasks, not universal certificates of competence. The gap between top-five and top-one accuracy shows how much a metric's definition can shape a headline. Robust evaluation should also ask what happens under altered inputs, unfamiliar settings, and across demographic groups. A system's failure pattern may matter more than its average score when the decision affects safety or access to services. Metrics need context, and deployment needs scrutiny beyond the leaderboard.

A game champion can still be an intellectual specialist with no transfer.

Rewards teach actions. Reinforcement learning trains an agent through actions and rewards rather than labeled examples. In the book's simple robot-soccer illustration, a robot gradually learns that moving toward the ball can lead to a reward for kicking it. DeepMind combined this approach with neural networks to train systems on Atari games and Go. AlphaGo Zero learned Go through self-play, starting with the rules rather than human game examples, and defeated earlier versions of AlphaGo.

Success stays bounded. Transfer learning means using experience from one task to help with another. Mitchell notes that these systems generally do not do this as people do: a network trained for one Atari game must be trained separately for another, and AlphaGo cannot apply its Go skill to unrelated tasks. A system can develop a winning strategy inside a game while lacking concepts that would help it adapt to small changes in the game, let alone to everyday life.

Analysis

The achievement is remarkable precisely because it is not a general-purpose mind. Self-play allows a system to improve without a human supplying correct moves, yet its learning remains shaped by the rules, reward, and boundaries of its environment. This distinction is useful beyond games: optimization can be extraordinary inside a clearly defined problem while failing to travel outside it. Generalization is therefore not a bonus feature. It is a separate capacity that must be tested rather than inferred from a spectacular win.

Fluent outputs can conceal a machine's lack of world knowledge.

Language rests on context. In Mitchell's restaurant story, a customer orders a hamburger cooked rare, receives a burned one, and leaves angrily. People infer that he probably did not eat it. A language system would need to connect restaurant customs, sarcasm, idioms, and the sequence of events to reach that answer. Modern speech recognition can transcribe speech accurately without understanding what the words mean.

Patterns are not the whole story. Neural translation systems learn from large collections of human-translated sentences. They can be impressively useful, yet the book's examples show them confusing different meanings of words such as bill, or translating idioms literally. A system may also perform well on reading tests that ask it to locate an answer explicitly stated in a paragraph, without demonstrating that it can reason about implied meaning. Smooth language is evidence of useful pattern handling, not proof of comprehension.

Analysis

The gap between fluent form and grounded meaning is easy to miss because people instinctively treat language as evidence of a speaker's understanding. This makes linguistic systems especially prone to anthropomorphism. Yet the distinction should not erase their practical value: transcription and translation can save time even when they fail on nuance. A sensible standard is purpose-specific trust. Casual searches tolerate errors that legal, medical, or diplomatic communication cannot, especially when tone and implication carry the message.

Ground intelligence in common sense, abstraction, and experience of the world.

People bring hidden knowledge. Core knowledge is the early-developing understanding of objects, living things, and other people. Humans use intuitive physics to predict how objects move, intuitive biology to distinguish a dog from a stroller, and intuitive psychology to anticipate goals and feelings. These abilities support mental models: representations that people can use to imagine what may happen next or how a situation might change.

Abstraction links situations. Bongard problems present small sets of diagrams and ask what shared feature separates one group from another. People can often identify the relevant pattern from only a dozen examples, while a study described in the book found that neural networks trained on 20,000 examples of a simple same-or-different task performed only slightly better than guessing. Mitchell argues that analogy helps people recognize what matters across different situations. A machine's ability to classify familiar examples does not guarantee this flexible abstraction.

Analysis

This account makes intelligence less like a storehouse of facts and more like a system for modeling situations and mapping one situation onto another. That is a powerful explanation for why children learn concepts from so few examples. It also raises an open question: must machines share human bodily experience to gain useful common sense, or could other forms of interaction supply an equivalent grounding? The book makes embodiment a compelling possibility, not a settled requirement. That distinction keeps the hypothesis ambitious without treating it as proven.

Keep humans accountable when AI decisions carry serious consequences.

Benefits do not erase risks. AI can support navigation, accessibility, medical research, and repetitive work. But systems used for driving, face recognition, or consequential decisions can also be brittle, biased, difficult to explain, or vulnerable to manipulation. A 2018 survey of technology experts cited in the book found that 63 percent expected AI to leave people better off by 2030, while 37 percent did not.

Governance is part of design. Face recognition illustrates the stakes: a wrong match can deny access or lead to suspicion, and unequal error rates can intensify existing discrimination. Mitchell argues that technical experts and companies should not be left to regulate themselves. Government, universities, companies, and public-interest groups need to help set rules. The relevant question is not simply whether an AI system is accurate, but whether its errors, uses, and effects are acceptable in the situation where it operates.

Analysis

The book's practical warning is about misplaced trust, not just future machine motives. Once a system is embedded in institutions, its design choices can affect people who never chose to use it. Human oversight is meaningful only when a person has the authority, information, and time to challenge a recommendation. Otherwise, a human in the loop can become a ceremonial shield for automation. Regulation should therefore address the full decision process, including who is answerable when the system fails.

Judge creativity by discernment, not output alone.

Imitation can move us. EMI, David Cope's music program, generated pieces modeled on classical composers. Cope selected musical patterns, built rules for combining them, and chose which results to release. Some pieces sounded convincing enough to confuse trained musicians. Their effect unsettled Hofstadter, who feared that human creativity might prove mechanically simple.

Creation includes evaluation. Mitchell argues that producing a pleasing work is not the same as understanding or judging it. In EMI's case, Cope supplied the musical knowledge and selected the successful compositions; the program could not assess their emotional or artistic value. She allows that computers could be creative in principle, but says existing systems do not meet this fuller standard. Her view leaves an important distinction: a machine can contribute novel material to a creative process even if the human still supplies taste, interpretation, and responsibility.

Analysis

The distinction between generating and judging is illuminating, though it may set a demanding test for machine creativity. Human artists also work through inherited forms, tools, collaborators, and feedback from audiences. That does not make their work uncreative, but it complicates the idea that creativity requires solitary self-understanding. A useful question is what kind of agency a system contributes: Does it merely produce variations, or can it select, revise, and explain choices in relation to a goal? The answer may be a spectrum rather than a simple yes or no.

Analysis

Mitchell's central achievement is to replace the vague question of whether machines are intelligent with more answerable questions: What task was defined? What examples trained the system? What exactly counts as success? How does it fail outside the test conditions? This shift turns AI from a contest over metaphysical labels into an empirical subject. A program can outperform people at a bounded task without thereby reproducing the flexible capacities that make human competence travel across settings.

The book's strongest evidence comes from contrast. ImageNet accuracy sits beside misclassification, bias, and adversarial vulnerability. AlphaGo's striking strategic play sits beside its inability to use that learning elsewhere. Neural translation's fluency sits beside its difficulty with idiom and context. These contrasts do not negate achievement. They reveal that performance is multidimensional, and that a single metric can hide precisely the limitations that matter in deployment.

The argument also has a productive tension. Mitchell treats common sense, abstraction, analogy, and embodied experience as likely ingredients of general intelligence, while acknowledging that the terms themselves remain imperfectly understood. That uncertainty is not a flaw to be papered over. It marks the boundary between a useful research program and a settled theory. Still, humanlike understanding may not be the only route to reliable systems: a specialized tool could be dependable through constraints, verification, and human review without possessing a human mental model. The practical standard should therefore be fitness for purpose, not resemblance to people.

The book's enduring lesson is methodological. Treat AI claims as claims about systems operating under specific conditions, not as declarations that intelligence has been solved or is about to arrive. Measure not only average performance but also robustness, transfer, error distribution, bias, and accountability. The most urgent danger is not necessarily an artificial mind with limitless power. It is the ordinary institutional decision to grant fallible systems authority that their tested competence does not justify.

Summary in 30 Seconds
AI systems can excel at tightly defined tasks without gaining abilities that transfer to new settings. Deep learning depends on large datasets, computing power, and substantial human design and labeling. High benchmark scores can hide brittle performance, biased errors, and susceptibility to altered inputs. Fluent language does not establish understanding, which relies on context, common sense, and abstraction. Consequential systems need independent scrutiny and accountable human oversight.
Contains spoilers

Last updated:

Report Issue

Review Summary

4.33 out of 5
Average of 4k+ ratings from Goodreads and Amazon.

Artificial Intelligence: A Guide for Thinking Humans provides a balanced overview of AI's current state and future potential. Readers appreciate Mitchell's clear explanations of complex concepts, her skepticism towards AI hype, and her exploration of AI's limitations. The book covers AI history, recent breakthroughs, and ethical concerns. While some find certain sections challenging, most praise Mitchell's accessible writing style. Some readers note that parts of the book are already outdated due to rapid AI advancements, particularly in language models, but still find value in its foundational content.

Your rating:
4.63
595 ratings
Want to read the full book?

Glossary

narrow AI

Task-specific artificial intelligence

An AI system designed to perform one narrowly defined task or a small set of related tasks. A program may become exceptionally good at Go, image classification, or translation while lacking the ability to apply its skills broadly. The term contrasts with general AI, which would integrate abilities across many kinds of situations.

deep learning

Training multilayer neural networks

A family of machine-learning methods that train neural networks with multiple hidden layers. The word deep refers to the number of hidden layers, not necessarily to the sophistication or meaning of what the system learns. Deep learning became highly successful as large data sets and powerful computing hardware became available.

supervised learning

Learning from labeled examples

A training method in which a system receives examples paired with correct answers, called labels. It compares its predictions with those answers and adjusts internal parameters to reduce errors. A system trained to identify handwritten digits, for example, learns from images that people have labeled with the corresponding digit.

convolutional neural network

Network for spatial image patterns

A neural network architecture commonly used for image processing. It applies learned filters across local regions of an image, building layered representations of features that can become increasingly complex. Convolutional neural networks have powered major advances in object recognition, though their classifications can rely on misleading image correlations.

adversarial example

Input engineered to fool AI

An input changed in a carefully chosen way that causes a machine-learning system to make a confident error. In image examples, the change may be imperceptible to people while making a neural network assign a different category. Researchers have also demonstrated analogous attacks involving speech and text.

reinforcement learning

Learning through actions and rewards

A machine-learning approach in which an agent acts in an environment and receives rewards or penalties. The agent uses these signals to improve its choices over time, rather than learning from a set of examples labeled with the correct action. It has produced notable results in games, where rules and rewards are clearly defined.

transfer learning

Reusing learning across tasks

The ability to apply knowledge or skills acquired in one task to improve performance on another, related task. People often transfer experience naturally, such as using familiarity with one racket sport when learning another. Mitchell emphasizes that many AI systems instead need to be trained separately for each task.

Winograd schemas

Tests of contextual language reasoning

Short paired sentences that differ by a small change in wording, where that change determines what a pronoun or phrase refers to. Solving them generally requires everyday knowledge about people and objects. The challenge was proposed as a way to test language understanding without allowing a program to rely only on conversational tricks.

core knowledge

Early intuitive world knowledge

Basic knowledge about objects, living things, and other people that humans possess or develop early in life. It includes expectations about how objects behave, how living things differ from inanimate ones, and how people act in pursuit of goals. This knowledge supports learning, prediction, and common-sense interpretation.

Bongard problems

Puzzles requiring visual abstraction

Visual puzzles consisting of two groups of diagrams. The solver must identify the shared concept that distinguishes the examples on one side from those on the other. The relevant feature may be simple, such as size, or require inventing an abstract description. Mitchell uses them to illustrate how difficult flexible abstraction is for AI.

Copycat

Program for letter-string analogies

A computer program developed by Melanie Mitchell and Douglas Hofstadter to solve analogy puzzles involving alphabetic strings. It combines processes that notice features in a problem with concepts such as successor, sameness, and letter group. Its design explored how analogy and abstraction might arise through interaction between perception and prior concepts.

barrier of meaning

Gap between performance and understanding

Mitchell's term for the distance between successful task performance and the richer understanding humans bring to perception, language, and reasoning. Current AI systems can excel at narrowly defined tasks while lacking common-sense knowledge, robust abstraction, and flexible transfer. Their unusual errors and vulnerability to manipulation help expose this gap.

FAQ

What's Artificial Intelligence: A Guide for Thinking Humans about?

  • Exploration of AI: The book provides a comprehensive overview of artificial intelligence, examining its history, current capabilities, and limitations.
  • Human vs. Machine Intelligence: It highlights the differences between human intelligence and AI, emphasizing AI's proficiency in narrow tasks but its lack of general understanding.
  • Complexity of Intelligence: The author delves into the complexities of human intelligence, comparing it to AI and raising questions about creativity, emotions, and consciousness.

Why should I read Artificial Intelligence: A Guide for Thinking Humans?

  • Informed Perspective: The book offers a balanced view of AI, discussing both its potential benefits and ethical concerns.
  • Accessible Language: Written in an engaging manner, it makes complex topics understandable for readers without a technical background.
  • Critical Thinking: It encourages readers to think critically about AI's implications in society, prompting discussions about trust, ethics, and future human-AI interactions.

What are the key takeaways of Artificial Intelligence: A Guide for Thinking Humans?

  • AI's Limitations: AI has made significant strides but still lacks the general intelligence and common sense that humans possess.
  • Human-AI Relationship: The book explores the evolving relationship between humans and AI, emphasizing the need for ethical considerations.
  • Future of AI: It discusses the potential for superhuman intelligence and the societal impacts that may arise from AI advancements.

What are the best quotes from Artificial Intelligence: A Guide for Thinking Humans and what do they mean?

  • "AI was harder than we thought.": This quote underscores the complexity of replicating human-like intelligence in machines.
  • "We tend to overestimate AI advances...": It serves as a caution against the hype surrounding AI, reminding readers of the complexity of human intelligence.
  • "Machine stupidity creates a tail risk.": This highlights the potential dangers of relying on AI systems that lack true understanding.

How does Artificial Intelligence: A Guide for Thinking Humans define artificial intelligence?

  • Symbolic vs. Subsymbolic AI: The book discusses symbolic AI, which uses human-understandable symbols, and subsymbolic AI, which relies on neural networks.
  • Complexity of Intelligence: Intelligence is defined as encompassing various dimensions, including emotional, verbal, and logical intelligence.
  • AI's Goals: AI aims to replicate or simulate aspects of human intelligence, but achieving true general intelligence remains a challenge.

What are the main methods of AI discussed in Artificial Intelligence: A Guide for Thinking Humans?

  • Deep Learning: This method uses deep neural networks to learn from vast amounts of data, dominating current AI research.
  • Reinforcement Learning: Involves training agents to make decisions based on rewards and penalties, allowing them to learn optimal behaviors.
  • Monte Carlo Tree Search: Used in game-playing AI, it simulates possible future moves to evaluate the best action.

How does Artificial Intelligence: A Guide for Thinking Humans address the ethical implications of AI?

  • Bias in AI Systems: The book discusses how AI can perpetuate societal biases, particularly in applications like facial recognition.
  • Need for Regulation: It advocates for thoughtful regulation to ensure AI technologies are developed and used responsibly.
  • Moral Machines: Raises questions about whether machines can possess moral reasoning and how to align AI systems' values with human values.

What are the challenges of achieving general AI according to Artificial Intelligence: A Guide for Thinking Humans?

  • Complexity of Human Intelligence: Human intelligence involves understanding context, emotions, and social interactions, which AI struggles to replicate.
  • Transfer Learning Limitations: Current AI systems cannot easily apply knowledge from one task to another, unlike humans.
  • Long-Tail Problem: AI systems often fail in rare or unexpected situations, highlighting the need for robust learning methods.

How does Artificial Intelligence: A Guide for Thinking Humans illustrate the difference between human and machine learning?

  • Learning from Few Examples: Humans can learn new concepts from few examples, while AI requires vast amounts of labeled data.
  • Exploration vs. Exploitation: Humans naturally balance exploring new options with exploiting known strategies, a balance AI struggles to achieve.
  • Common Sense Knowledge: Humans possess a wealth of common sense knowledge that informs decisions, a quality current AI lacks.

What is the significance of reinforcement learning in Artificial Intelligence: A Guide for Thinking Humans?

  • Learning from Interaction: Reinforcement learning allows AI systems to make decisions by interacting with their environment.
  • Applications in Gaming: Demonstrated by AlphaGo, it shows the potential of reinforcement learning in mastering complex tasks.
  • Challenges in Real-World Applications: Transferring reinforcement learning from games to real-world applications remains challenging.

How does Artificial Intelligence: A Guide for Thinking Humans address natural language processing (NLP)?

  • Complexity of Language: NLP systems face challenges in understanding human language due to its ambiguity and context-dependence.
  • Advancements and Limitations: While deep learning has advanced NLP, these systems often lack true understanding.
  • Future Directions: Improvement requires incorporating commonsense knowledge and reasoning abilities for human-like responses.

What role does common sense play in AI according to Artificial Intelligence: A Guide for Thinking Humans?

  • Foundation of Understanding: Common sense is critical for human intelligence, which AI currently lacks.
  • Challenges in AI Development: Instilling common sense in AI is difficult as much of this knowledge is subconscious.
  • Research Efforts: Ongoing efforts aim to develop AI systems with commonsense reasoning capabilities.

About the Author

Melanie Mitchell is a computer science professor at Portland State University with extensive experience in AI research. She received her PhD from the University of Michigan under Douglas Hofstadter and John Holland, developing the Copycat cognitive architecture. Mitchell has made significant contributions to analogical reasoning, complex systems, genetic algorithms, and cellular automata. She has authored several influential books, including "An Introduction to Genetic Algorithms" and "Complexity: A Guided Tour," which won the 2010 Phi Beta Kappa Science Book Award. Her work has been widely cited, and she has also critiqued other prominent works in the field of artificial intelligence and complex systems.

Download PDF

To save this Artificial Intelligence summary for later, download the free PDF. You can print it out, or read offline at your convenience.
Download PDF
File size: 0.27 MB     Pages: 21

Download EPUB

To read this Artificial Intelligence summary on your e-reader device or app, download the free EPUB. The .epub digital book format is ideal for reading ebooks on phones, tablets, and e-readers.
Download EPUB
File size: 1.37 MB     Pages: 28
Want to read the full book?
Follow
Listen10 mins
Home
Swipe
Library
Get App
Try Full Access for 3 Days
Listen, bookmark, and more
Compare Features Free Pro
📖 Read Summaries
Read unlimited summaries. Free users get 3 per month
🎧 Listen to Summaries
Listen to unlimited summaries in 40 languages
—
❤️ Unlimited Bookmarks
Free users are limited to 4
—
📜 Unlimited History
Free users are limited to 4
—
📥 Unlimited Downloads
Free users are limited to 1
—
Risk-Free Timeline
Today: Get Instant Access
Listen to full summaries of 26,000+ books. That's 12,000+ hours of audio!
Day 2: Trial Reminder
We'll send you a notification that your trial is ending soon.
Day 3: Your subscription begins
You'll be charged on Nov 24,
cancel anytime before.
Consume 2.8× More Books
2.8× more books Listening Reading
Our users love us
600,000+ readers
Trustpilot Rating
TrustPilot
4.6 Excellent
This site is a total game-changer. I've been flying through book summaries like never before. Highly, highly recommend.
— Dave G
Worth my money and time, and really well made. I've never seen this quality of summaries on other websites. Very helpful!
— Em
Highly recommended!! Fantastic service. Perfect for those that want a little more than a teaser but not all the intricate details of a full audio book.
— Greg M
Save 62%
Yearly
$119.88 $44.99/year/yr
$3.75/mo
Monthly
$9.99/mo
Start a 3-Day Free Trial
3 days free, then $44.99/year. Cancel anytime.
Unlock a world of fiction & nonfiction books
26,000+ books for the price of 2 books
Read any book in 10 minutes
Discover new books like Tinder
Request any book if it's not summarized
Read more books than anyone you know
#1 app for book lovers
Lifelike & immersive summaries
30-day money-back guarantee
Download summaries in EPUBs or PDFs
Cancel anytime in a few clicks

Why one summary?

Scanner
Find a barcode to scan

⭐
⭐
⭐
We have a special gift for you
Open
38% off
DISCOUNT FOR YOU
⭐
⭐
⭐
$79.99
$49.99/year
Only $4.16 per month
Continue
2 taps to start, super easy to cancel
Settings
General
Widget
Loading...