Key Takeaways
Treat AI headlines as sales pitches until the evidence survives scrutiny.
Follow the incentives. Wilkins argues that much AI hype grows from ordinary commercial pressures, not a secret conspiracy. Ad-supported media benefit from alarming headlines, companies gain attention and investment by labeling products AI, and researchers can attract funding with ambitious claims. Each group amplifies the others, so speculation can start to feel like established fact. The book cites a survey of 2,830 European startups in which roughly 40 percent reportedly claimed AI despite not using it, and says the label could bring a funding advantage.
Check the underlying claim. A dramatic headline about robot armies, job losses, or AI lovers may describe a limited demonstration or a fictional scenario. The useful response is to ask what the system actually does, under what conditions, and what evidence supports the prediction.
The incentive map is a strong antidote to treating publicity as proof. It also avoids needing a single villain: attention-seeking and financial reward can produce exaggeration without coordination. Still, incentives explain why a claim spreads, not whether it is false. Some striking AI applications are real, and a skeptical reader should examine the demonstration, the independent evidence, and the gap between a laboratory result and routine use.
Most deployed AI is narrow pattern recognition, not a digital mind.
Separate capability from consciousness. The book contrasts narrow AI, built for a specific task, with general AI, which would perform across tasks with human-like flexibility. Examples of narrow systems include recognizing objects in images, sorting data, or producing a risk score. A system can perform these tasks usefully without understanding the world as a person does. Wilkins uses YouTube's copyright-matching tool, ContentID, to show how a system that works on clear matches can still misidentify material, including a cat's purring.
Keep the terms straight. Machine learning identifies patterns from data, and neural networks are one family of models used for this. Deep learning uses neural networks with many processing layers. The book's account blurs reinforcement learning with labeled examples; more precisely, reinforcement learning learns through feedback about actions and outcomes. These approaches are not interchangeable.
This distinction helps puncture the misleading idea that every improved classifier is a step toward a human-like mind. Capability is task-specific: excellent image recognition does not automatically yield common sense, flexible reasoning, or consciousness. At the same time, narrow systems can still shape decisions at scale. Calling a tool limited should not become a reason to ignore its consequences, especially when a seemingly small prediction determines what someone sees, earns, or receives.
A confident score can be wrong when the event is rare.
Ask how common the event is. Wilkins explains why a test's overall accuracy can hide a large number of false alarms. Imagine screening 1,000 drivers when one is drunk. If a test correctly detects that driver and mistakenly flags 1 percent of the other 999, about 10 people test positive, but only one is actually drunk. Under these assumptions, most positive results are false alarms. The point is not that every test behaves this way, but that accuracy alone is inadequate when the target event is uncommon.
Scores can mislead in court, too. The book describes two people who received sharply different repeat-offense risk scores. The person given the higher score did not reoffend in the account, while the person with the lower score did. A probability is not a verdict, and its limitations matter when liberty is at stake.
This is a practical lesson in base rates, the background frequency of whatever a system is trying to detect. It explains why a tool can seem accurate in testing yet produce many mistaken accusations in real use. The example also invites a harder question: who bears the cost of false positives, and who bears the cost of missed cases? Any high-stakes system needs clear error rates, appropriate thresholds, and a way to challenge an individual result.
Big Data creates leverage, but every connected device widens exposure.
Small traces add up. Big Data is the large, rapidly growing collection of information gathered from many interactions and devices. A message, location signal, or search may seem unremarkable by itself; assembled with other traces, it can reveal patterns about a person. The book uses a social-media fact-checking bot to illustrate the trade-off: to respond to a conversation, the bot needs access to its messages, creating both an information source and a privacy risk.
Reduce what can be lost. Wilkins recommends collecting only necessary data, deleting it when it is no longer needed, and separating records so a breach cannot expose everything. Encryption, restricted physical access, and offline backups add protection. Users can also limit permissions and close unused accounts. The book suggests entering false personal details, but that can break services or create other problems, so it is not a substitute for responsible data handling.
The key insight is that privacy risk is cumulative. A service need not collect a dramatic secret to become revealing when ordinary details are combined. Data minimization is therefore more than a technical preference: it limits what an organization can misuse, lose, or expose. Individual caution helps, but it cannot replace institutional responsibility. Users often cannot see the full data trail or meaningfully negotiate how it is used.
Automate a proven workflow, then count every human fallback.
Build before automating. Wilkins uses growth hacking to mean an iterative, data-led way of improving a process. His advice is to establish a repeatable method that works before adding AI, then use automation to speed up tasks that already produce reliable results. A vague goal such as making a business smarter gives a model little to act on; a clear task and a measurable outcome make experimentation more useful.
Include the hidden support. The book recounts a self-driving truck carrying 51,000 cans of beer on a 120-mile route. The trip involved police vehicles, support vehicles, tow trucks, and a human driver available to take over. That demonstration shows why counting only the machine can make automation look more independent than it is. Wilkins argues that new tools can create different work, but the example does not establish that every displaced worker will find a better job.
The workflow-first principle is sound: automation tends to expose unclear processes rather than magically repair them. The truck story makes a useful accounting point by revealing the human labor hidden around a machine. But the book's confidence that technology will create enough new jobs to offset losses goes beyond what one demonstration can show. Job creation, retraining, wages, and who captures productivity gains are separate questions, not automatic consequences of adoption.
Historical data can teach algorithms yesterday's discrimination as objective truth.
Patterns inherit their past. Amazon tested a recruitment tool trained on about a decade of resumes, many from men. According to the book's account, the system learned to penalize terms associated with women and graduates of women's colleges. It also gave greater weight to some forceful verbs, mistaking patterns in past hiring for reliable signs of future performance. The system did not need an instruction to discriminate; the examples it learned from carried the pattern into its recommendations.
Check the whole pipeline. A soap dispenser at Facebook's headquarters reportedly failed to detect a Black employee's hand while working for a white colleague. The book also cites a study reporting more errors in self-driving car tests involving darker-skinned pedestrians. These examples point beyond model code to sensors, training data, testing conditions, and the people included in design decisions.
The central warning is that prediction and fairness are not the same thing. A model can reproduce patterns accurately while preserving an unjust status quo. Removing explicit references to protected groups may not solve the problem if other details act as substitutes. Testing across different users and environments, publishing relevant error rates, and giving affected people recourse turn fairness from a promise into something that can be examined.
Keep consequential systems transparent, contestable, and backed by humans.
Make the evidence visible. Wilkins contrasts opaque risk scores used in legal settings with Video Assistant Referee, or VAR, in soccer. VAR gives officials high-speed camera views, including close analysis of goal-line decisions, while human referees remain involved. The system can be questioned because the underlying footage is available for review. In contrast, the book describes a proprietary criminal-risk tool whose calculations were not open to public inspection and whose scores influenced decisions about bail or sentencing.
Design for appeal. A person affected by an algorithmic decision should be able to see the relevant evidence, understand the decision's limits, and ask a qualified human to review it. VAR is not a template for every domain: sports decisions use narrower rules than judgments about justice or human welfare. Its value here is the combination of visibility, review, and human responsibility.
Transparency is not merely a courtesy. It allows errors to be detected, assumptions to be challenged, and responsibility to remain identifiable. The comparison with sports is useful because it turns an abstract principle into a familiar process. Its limits matter, though: a clear replay can settle whether a ball crossed a line, while social and legal decisions involve contested values. Review mechanisms must expose both evidence and the standards used to interpret it.
Use AI for support without mistaking simulated attention for relationship.
Connection needs more than conversation. The book argues that isolation can damage social functioning and describes sensory-deprivation experiments in which participants became restless and experienced hallucinations after extended restriction of sensory input. Wilkins links this vulnerability to life increasingly mediated by screens, and imagines AI tools offering patients conversation, monitoring, and personalized encouragement. Such systems might help people access support or help clinicians organize information.
Do not confuse unlike experiences. Sensory deprivation is not the same as ordinary screen use, and a conversational program is not a human relationship. The book's proposal that an AI could act as a tireless counselor or authority figure captures both its promise and its danger: a system designed to soothe can also steer. In healthcare, patients should know when they are speaking to a machine and what information it records.
A responsive interface can reduce friction in seeking help, especially when human services are difficult to reach. Yet social connection involves reciprocity, shared responsibility, and the possibility of being changed by another person, not just receiving fluent replies. The book's account is most persuasive as a warning against treating loneliness as a software defect. AI may support care, but its convenience should not become a rationale for withdrawing human contact.
Treat AI as an instrument of power, not a neutral oracle.
Ask who sets the rules. Wilkins argues that technology companies can gain political influence when they control platforms, data, and the systems people rely on. He extends this concern through the idea of AI as a new authority, almost a secular religion whose claims are interpreted by companies and specialists. This is a speculative argument, but it directs attention to who defines the system's goals and who can challenge them.
Scores can govern daily life. The book describes China's social credit system as a way to rate citizens' behavior and restrict access to services such as transportation or education. It recounts a student who reportedly faced restrictions after her father caused a fatal car accident, illustrating how collective or inherited judgments can punish someone for another person's actions. Whether a system is called AI or administration, its rules and consequences deserve scrutiny.
The power analysis adds a political dimension often missing from stories about clever machines. Control over infrastructure, information, and decision rules can matter more than whether a system is genuinely intelligent. The religious metaphor dramatizes how authority can become difficult to question, but the book sometimes treats corporate motives as more settled than its evidence supports. The stronger test is institutional: who owns the data, sets the rules, audits outcomes, and provides an appeal?
Analysis
Wilkins's book is a polemical introduction to AI, not a neutral technical survey. Its most durable contribution is a shift in attention: away from imagined machine intentions and toward the institutions that build, sell, deploy, and govern automated systems. That move connects the book's otherwise sprawling subjects, from clickbait and hiring to surveillance, legal scoring, and self-driving vehicles. Across these cases, the central concern is not simply whether a machine can think. It is how claims about machine capability can redistribute money, authority, and risk.
The book is strongest when it asks readers to inspect the system around a demonstration. The escorted truck, opaque court score, and visible soccer replay each make a different point about hidden labor, unreviewable decisions, and accountability. Its basic caution about probability is also valuable: a score should not be mistaken for certainty, particularly when errors affect liberty or access to essential services.
Its weaknesses stem from the same urgency that makes it engaging. Some technical explanations are simplified, and the account occasionally presents speculative motives as fact. The argument that AI is unlikely to replace jobs is more confident than the examples justify. A technology can create new occupations while still eliminating particular jobs, depressing wages, or concentrating gains. Likewise, analogies between AI, religion, and political rule illuminate the authority problem but do not prove that technology leaders consciously seek priestly power.
The book is dated to 2019, so its reports and forecasts should be checked against current evidence. Its enduring reading method is more useful than its strongest predictions: distinguish a system's actual capability from its publicity, inspect the data and incentives behind it, ask who bears the error, and demand a meaningful way to contest consequential decisions.
Review Summary
The reviews for Artificial Intelligence are mixed, with an average rating of 3.18 out of 5. Some readers found the book outdated, especially given recent advances in AI technology. One reviewer expected more focus on AI concepts but found it to be more about AI's societal impact. Another reader felt the book started with too much criticism but improved later. A Spanish-language review suggests the reader expected more from the book. Overall, opinions vary on the book's depth, relevance, and ability to meet reader expectations.
People Also Read
Glossary
Narrow AI
Task-specific artificial intelligenceArtificial intelligence designed to perform a limited range of tasks, such as classifying images, matching records, or making recommendations. A narrow system can perform well within its intended setting without possessing broad human-like understanding or the flexibility to handle unrelated tasks.
General AI
Broad human-like capabilityA hypothetical form of AI able to learn and perform across a wide variety of domains with flexibility comparable to human intelligence. The book contrasts this aspiration with present-day task-specific systems. It should not be confused with a system that is merely very capable at one narrow task.
Machine learning
Learning patterns from dataA family of computational methods in which a system adjusts its behavior based on examples or experience rather than relying only on hand-written instructions. Its results depend on the data, training process, and task. Learning patterns does not by itself mean the system understands those patterns as a person would.
Neural network
Connected computational unitsA machine-learning model made of linked computational units arranged to transform input into output. During training, the model adjusts the strength of connections to improve performance on a task. Neural networks can identify complex patterns, but their outputs remain dependent on training data, model design, and the conditions in which they are used.
Deep learning
Multi-layer neural network learningA group of machine-learning approaches that use neural networks with many layers of computation. These layers can learn increasingly complex representations from data, making deep learning useful for tasks such as image and speech recognition. Its internal complexity can make results difficult to interpret, but it does not automatically make a system generally intelligent.
Reinforcement learning
Learning from action feedbackA machine-learning approach in which a system takes actions and learns from feedback about their outcomes, often represented as rewards or penalties. It differs from training that simply supplies the correct label for each example. The book discusses reinforcement learning but describes it in terms that overlap with other training methods.
Big Data
Large-scale collected informationVery large, varied, and rapidly generated collections of data, often gathered from digital services, sensors, and connected devices. Such information can be analyzed for patterns that would be difficult to detect in smaller datasets. Its scale can create useful insights, but also raises concerns about privacy, security, and the ability to control or interpret the collection.
Internet of Things
Networked physical devicesA network of physical objects equipped with sensors, software, or communication capabilities that exchange information with other devices or services. Examples include connected household appliances and industrial sensors. Linking devices can enable coordination and automation, while increasing the number of potential security weaknesses and sources of personal data.
Attack surface
Potential points of system entryThe set of interfaces, devices, accounts, and other points through which an attacker might try to access a system. Connecting more components can create additional paths into a network, especially when one weak device can affect others. Limiting unnecessary connections and protecting each component can reduce exposure.
Algorithmic fairness
Fairness in automated decisionsThe problem of ensuring that algorithmic systems do not produce unjustified or systematically unequal outcomes for different people or groups. Fairness depends on what a system predicts, how its data was collected, which errors matter, and how its outputs are used. There is no single technical test that resolves every ethical or legal question.
Security by design
Build security in from the startAn approach to product and system development that considers security during design rather than treating it as a later repair. The book emphasizes limiting data collection and unnecessary features, deleting information that is no longer needed, and separating records so a breach cannot expose everything at once.
Social credit system
Behavior-based citizen scoringA system for assessing people according to recorded behavior and using those assessments to influence access to services, opportunities, or privileges. The book presents China's system as a warning about surveillance and collective punishment. Specific systems and policies can differ, so claims about how they operate should be examined in their particular context.
Video Assistant Referee
Video review for soccer decisionsA soccer officiating system that lets video officials review selected incidents and communicate relevant footage to the on-field referee. It uses camera views to help examine decisions, while human officials retain responsibility for applying the rules. The book presents VAR as an example of technology paired with visible evidence and human review.
Growth hacking
Iterative, data-led improvementAn entrepreneurial approach that uses repeated experimentation and analysis of results to improve a product or process. In the book, the method means establishing a reliable workflow and clear goal before applying AI to automate it. The term is broader in business practice, but this is the role it plays in Wilkins's discussion.
Download PDF
Download EPUB
.epub digital book format is ideal for reading ebooks on phones, tablets, and e-readers.