AI demos are good at creating momentum. They are less good at proving that a system belongs inside your business. That gap is the most useful commercial lesson in The AI Con, Emily M. Bender and Alex Hanna’s 2025 critique of the stories told around AI.
Bender, a linguist, and Hanna, a sociologist, argue that “AI” often works as an umbrella term: it encourages us to imagine intelligence while hiding the actual system, data, labour and incentives underneath. The book is intentionally combative; readers may reject some of its conclusions. Yet its practical challenge is hard to dismiss: stop evaluating AI as a magical capability and start evaluating the specific automation being sold.
Five questions before you buy or build
- What precise task is being automated?
- What inputs does it use, and what outputs does it produce?
- How was it tested against conditions like ours?
- Who checks mistakes, and what recourse exists when it fails?
- What measurable result should improve within a defined pilot?
This is not anti-innovation. It is how serious adoption begins. IEEE’s 2025 procurement standard treats AI purchasing as a lifecycle covering problem definition, vendor evaluation, contracts and monitoring. NIST likewise frames trustworthy AI as something that must be designed, used and evaluated — not merely claimed.
Good AI strategy is not belief. It is disciplined selection.
The book is strongest when exposing language that turns a tool into an actor: AI “understands”, “decides” or “replaces”. Precise language restores accountability. A system classifies enquiries; a manager decides which classification is safe enough to use. A model drafts a response; the business owns the promise sent to a customer.
Its weakness is the mirror image of its strength. By concentrating on hype and power, it gives less space to well-scoped deployments that already deliver useful value. Founders should therefore read it as a stress test, not a ban.
The best takeaway is simple: do not buy AI because the demo is impressive. Buy — or build — a tightly defined workflow because evidence shows it improves an outcome you care about, with humans, costs and failure modes made visible.
Sources
- HarperCollins — The AI Con
- Official book site — The AI Con
- Kirkus Reviews — The AI Con
- IEEE 3119-2025 — AI procurement standard
- NIST — AI Risk Management Framework
Review note: this article is a sourced key-lessons review based on publisher, author, standards and independent-review materials. It does not reproduce the copyrighted book text.