Generative AI: the big picture — recap & real-world examples
Real-world examples & war stories
The Air Canada chatbot (2024). A customer asked Air Canada's website chatbot about bereavement fares. The chatbot invented a refund policy that did not exist. Air Canada argued in court that the chatbot was "a separate legal entity" responsible for its own words. A tribunal disagreed and made the airline honor the invented policy. 🎯 Takeaway: a company is responsible for what its generative model says, the same way it is responsible for what an employee says. Probabilistic output that reaches a customer with no check in between is not a technical detail — it is a policy your company now holds.
GitHub Copilot's growth (2021 onward). Code generation became one of the fastest, widest adoptions of any generative AI tool, inside engineering teams that are usually skeptical of new tools. 🎯 Takeaway: code is checkable by running it — tests pass or fail, the type checker accepts or rejects — which gives engineers a fast, cheap way to catch a bad output before it ships. That tight feedback loop is exactly the value-creating pattern this module names.
Google's Gemini image generation (2024). An image-generation feature, tuned to represent diversity across historical images, produced clearly inaccurate results — including historically White figures depicted with different skin tones. Google paused the feature within days. 🎯 Takeaway: each modality carries its own failure pattern, and a fix applied broadly, without enough testing on the specific modality's edge cases, can create a new, more visible failure than the one it was meant to solve.
Klarna's AI customer service reversal (2024–2025). In February 2024, Klarna said its AI assistant was doing the work of seven hundred customer service agents. In May 2025, its CEO told Bloomberg the company was hiring humans again. He said AI was cheaper but gave lower quality, and that customers must always be able to reach a person. 🎯 Takeaway: a use case can clear the "slow and expensive by hand" gate and still fail the "is checking or correcting it acceptable" gate — the choice between an AI-only, a human-only, and a blended service was a build-or-buy-shaped decision the company had to revisit in public.
AI-generated content corrections at major outlets (2023). CNET, in January 2023, published machine-written articles that later needed public corrections. 🎯 Takeaway: text generation is mature and cheap, but maturity is not the same as zero error rate — publishing generated text with no fact-check step treats a probabilistic tool as if it were deterministic.
Fast food AI drive-thru pilots (2024–2025). In June 2024, McDonald's said it would end its test of IBM's voice ordering at about 100 restaurants. It had aimed for 95% order accuracy before a wider rollout. In August 2025, Taco Bell said it was rethinking where it uses voice AI after viral ordering failures, even though the system had handled millions of orders. Both chains kept voice ordering on the roadmap. 🎯 Takeaway: a demo run in a quiet room and a product run in a loud drive-thru lane are different probabilistic environments — the failure rate that looked acceptable in testing did not hold in the real, noisy input the product actually faced.
Module recap
| Lesson | The one idea | The question it makes you ask |
|---|---|---|
| What makes AI "generative"? | A judge and a maker do different jobs — know which one your task needs | Does this task need a judgment, or something new to be made? |
| The five modalities | Five factories, five maturity levels, five failure patterns | Which modality does this feature need, and is it mature enough? |
| Probabilistic software | The same input can give a different output — testing becomes measuring | What is the acceptable range of outcomes, and have we measured it? |
| The generative AI product stack | The model is one of four layers — price all four | Have we budgeted grounding, action, and operations, not just the model? |
| Build, buy, or fine-tune | Buy by default; build or fine-tune only with a clear reason | Is the model itself our product, or the system we build around it? |
| Where value is created and destroyed | Three gates decide the outcome: slow by hand, cheap to check, safe to sometimes get wrong | Does this use case actually clear all three gates? |
The through-line: every lesson in this module is one piece of a single shift. Generative AI does a different job than the predictive AI that came before it. That job spans five modalities, each with its own maturity. The output is probabilistic, which changes what testing, support, and trust mean. The model is one layer in a larger stack, most of which a product team builds itself. Buying the model and building the system around it beats training your own, in almost every real case. And the use cases that actually pay off are the ones that clear three honest gates — slow by hand, cheap to check, safe to sometimes get wrong — not the ones with the best demo.
Walk-away question: "For our next generative AI proposal: which modality does it need, what is its measured pass rate across many runs, have we priced all four layers of the stack, can we justify build or fine-tune over buy, and does it clear all three value gates?"
If yes, you are scoping a real product decision. If no, you now know exactly which lesson in this module to reread — and where the deeper mechanics for each layer live, one module away in RAG & vector databases and Agentic AI.