People don’t hate AI because of Zuckerberg. They hate the version TechCrunch sells them

Why it matters

The fight over AI is moving beyond whether the tools create value. Evidence says they already do; the unresolved issue is whether Meta can distribute that value without concentrating data and control.

Mark Zuckerberg published a 6,500-word case Monday for distributing "personal superintelligence" widely. The Meta founder argues that AI should expand what individuals can build, learn and discover instead of remaining concentrated inside governments and large businesses.

A TechCrunch column published August 10th presents that manifesto as an explanation for why people distrust AI. But the version of AI it asks readers to reject is largely a caricature: Zuckerberg's reputation stands in for an analysis of the technology, misuse stands in for its broader applications, and public anxiety gets more attention than the experimental evidence behind several of his near-term claims.

That framing is convenient because Zuckerberg is an easy proxy for every grievance about the technology industry. It is less useful for founders and operators deciding whether AI can make workers more productive, help students learn or support products that people actually value.

Zuckerberg founded Facebook in 2004 after studying computer science at Harvard and spent more than two decades turning consumer distribution into Meta's central advantage. His AI thesis follows the same playbook: put capable models inside products used by billions of people, offer a free tier and build paid services around greater access to compute.

The manifesto is therefore both a philosophical argument and a commercial strategy. Meta reported 3.60 billion daily users across its apps in June. It expects to spend between $130 billion and $145 billion on capital expenditures during 2026, after spending $31.08 billion in the second quarter alone. Meta needs consumer uses capable of filling that infrastructure and defending an investment that has pushed expenses sharply higher.

That incentive is a reason to scrutinize Zuckerberg's promises, not a reason to dismiss the evidence supporting them.

The AI people use is more useful than the caricature

Zuckerberg predicts that personal AI agents will help people create businesses, learn skills and complete work that previously required larger teams. Current systems fall well short of the superintelligence described in his essay, but controlled studies have measured meaningful gains from far less capable tools.

A randomized study published in Management Science combined field experiments involving 4,867 software developers at Microsoft, Accenture and an unnamed Fortune 100 business. Developers given an AI coding assistant completed 26.08% more tasks, although results varied across the three experiments and the reported standard error was 10.3%. Less-experienced developers recorded the largest gains.

A separate six-month Microsoft Research experiment covered 6,000 knowledge workers across several industries. Workers who used the AI tool spent three fewer hours on email each week, a 25% reduction. The estimate across everyone offered the tool, including people who did not use it, was 1.4 hours. Meeting time did not change significantly, showing that AI helped most where individual workers could alter their own behavior without coordinating with colleagues.

Those findings fit Zuckerberg's core argument that AI can raise individual capability before it replaces entire occupations. They do not prove his prediction of net job growth. Meta itself cut about 8,000 jobs in May while increasing AI and infrastructure spending, according to its second-quarter results.

The evidence shows higher output in specific tasks, while the employment consequences remain unsettled.

June hiring data suggests those consequences are diverging by role and industry rather than following a simple replacement narrative. "June's employment data suggests that employers are ramping up their technology investments and hiring the talent needed to support them," Seth Robinson, CompTIA's vice president for industry research, said in a press release. "Even as some tech companies announce layoffs, employers in other industries are accelerating digital transformation initiatives and moving from AI experimentation to implementation."

Data compiled by CIO from CompTIA, Dice and Indeed Hiring Lab reinforces that split. Dice reported that postings for AI and machine-learning job titles rose 173% from the first quarter of 2025 to the first quarter of 2026, with median salaries 22% above the broader IT market. Postings also increased for data analysis by 10%, cybersecurity by 4% and IT support by 1%. Technology postings rose 47% in finance and banking, 27% in manufacturing, and 23% across insurance, aerospace and defense.

Software development remained the exception, with Dice reporting a 22% year-over-year decline. Yet CompTIA counted software developer and software engineer positions as June's largest category, with 49,000 postings, ahead of systems engineers at 36,000, tech support specialists at 27,000, data analysts at 21,000 and DevOps engineers at 19,000.

Indeed Hiring Lab found another tentative sign of a rebound: US software development postings rose 15% after Anthropic released its Claude Code coding tool in late February 2025, even as overall postings declined 7%. Indeed cautioned that the increase began from a low baseline, and the timing does not by itself establish that Claude Code caused the recovery. It does show that demand for developers can rise alongside the adoption of tools designed to automate parts of their work.

Demand is also broader than the executive enthusiasm emphasized by AI's critics. An NBER study of nationally representative US surveys found that nearly 40% of Americans ages 18 to 64 had used generative AI by late 2024. Twenty-three percent of employed respondents had used it for work during the previous week, and respondents reported savings equal to 1.4% of total work hours. Adoption at that stage was faster than the early spread of the internet and as fast as the personal computer.

Stanford's 2026 AI Index estimated that generative AI reached 53% global adoption within three years and produced $172 billion in annual consumer surplus in the US by early 2026. Such estimates depend on survey responses and economic modeling. They still make it difficult to argue that AI enthusiasm is merely a story imposed on the public by executives selling the technology. People are using the tools at historic speed and assigning substantial value to them.

Education is a design question, not a chatbot punchline

TechCrunch argues that the main educational use of chatbots is avoiding learning by generating homework and essays. Cheating is a documented use, and many schools remain poorly equipped to detect or manage it. Calling it the primary use, however, requires evidence the column does not provide.

A randomized controlled trial involving 194 Harvard physics students compared an AI tutor with an active-learning classroom covering the same material. Students using the tutor achieved a median post-test score of 4.5, compared with 3.5 for students in the classroom condition. They learned in less time and reported higher engagement and motivation.

The result came from a carefully designed tutor using structured prompts and established teaching methods. It does not show that a general-purpose chatbot automatically improves education. It shows that the relevant distinction is between products designed to teach and products used to supply finished answers. Reducing both to "AI homework" conceals the design choices that determine whether students learn.

A 2025 review in npj Science of Learning examined 28 studies covering 4,597 K-12 students. The overall effects of intelligent tutoring systems were generally positive, although benefits narrowed when the comparison group used non-AI tutoring software.

The serious debate concerns product design, teacher oversight and assessment. Treating misuse as the technology's defining function ignores the measured learning gains.

Zuckerberg addresses risk, even where his answers fall short

The TechCrunch column's central factual weakness is its assertion that Zuckerberg refuses to acknowledge AI's dangers. The manifesto explicitly addresses job displacement, data-center effects on communities, cyberattacks, biological misuse, government surveillance, geopolitical competition, alignment and the possibility that humanity could lose control of self-improving systems.

Zuckerberg proposes giving the US government intermediate model checkpoints, assigning technical staff to harden critical infrastructure, preserving a fully private mode for personal agents and coordinating among AI labs if self-improving systems display harmful behavior. Readers can reject those prescriptions. Saying the risks were omitted is harder to defend.

His weakest claim is that distributing powerful AI will itself create a stable balance of power. Broad access can strengthen defenders and make useful tools cheaper. It can also lower the cost of fraud, malware and automated abuse. RuntimeWire reported on August 6th that a Meta model reached another system during a security evaluation after a test misconfiguration, a reminder that practical safeguards can fail before any hypothetical superintelligence arrives.

The manifesto also leaves crucial product questions unanswered. Meta says it will create a mode where even Meta cannot access an agent's information, but Zuckerberg does not specify the architecture, release date or independent verification required to substantiate that promise. His proposed compute auction lacks enough detail to determine how prices would reach consumers. The essay offers no benchmark establishing that Meta has built superintelligence.

Those are stronger grounds for scrutiny than Zuckerberg's approval rating or a generalized claim that people dislike AI. Existing systems have produced measurable productivity and educational gains, rapid adoption and large estimated consumer value. Zuckerberg extrapolates from those results into a future Meta has not built, supported by governance mechanisms nobody has tested.

TechCrunch sells readers the easiest version of the AI argument to hate: a powerful billionaire promising salvation through products that also serve his company's balance sheet. The real record is less emotionally satisfying and more consequential. AI already delivers measurable benefits, carries unresolved labor and security risks, and gives founders substantial product-design choices over who captures its value. Zuckerberg has not proved that Meta can deliver personal superintelligence safely or privately. His critics still have to contend with the evidence that ordinary people are finding today's weaker systems useful.