“Generative AI” covers any model that produces new content, text, images, audio, video, or structured data like protein sequences, rather than just classifying or predicting a label for existing content (if you’re new to AI concepts generally, our beginner’s guide to AI covers the foundational terms this post assumes). The category exploded into mainstream awareness with image generators and ChatGPT, but the applications that are actually sticking in production tend to look less flashy than the viral demos.
Content Creation
The most visible application is generating text, images, and video: marketing copy drafts, product descriptions at scale, image assets for ads and social posts, background music, and increasingly, short video clips generated directly from a text prompt. Specific tools now anchor each of these sub-categories in production use: Midjourney and Stable Diffusion-based tools dominate image generation, ElevenLabs has become a default choice for AI voiceover and dubbing work, and Suno can generate a complete song, vocals, lyrics, and instrumentation, from a single text prompt in under a minute.
The realistic production pattern here is rarely “AI generates the final asset unsupervised.” It’s AI generating a first draft or a wide set of variations quickly, with a human editor selecting, refining, and approving before anything ships. Teams that have gotten real value from this have generally built an editorial or design review step into the workflow rather than treating the model’s output as final, since generated content still fails in ways that are easy to miss on a quick skim (subtly wrong facts, inconsistent brand voice, images with garbled text or anatomically odd details).
Code generation deserves its own mention here since it’s arguably the single most commercially significant generative AI application right now: AI coding assistants integrated into editors, GitHub Copilot chief among them, have gone from novelty autocomplete to a standard part of professional developer workflows in a few years, handling everything from boilerplate to full feature implementation from a natural-language description, again with a human reviewing before merging.
Generative AI by Modality
The underlying model architectures differ a lot by content type, but the production pattern across all of them, human review before publishing rather than unsupervised output, holds fairly consistently:
| Modality | Representative Products | Typical Production Use |
|---|---|---|
| Text | ChatGPT, Claude, Gemini | Drafting, summarization, research |
| Code | GitHub Copilot, Claude Code | In-editor autocomplete, full feature implementation |
| Image | Midjourney, Stable Diffusion | Ad creative, concept art, product mockups |
| Audio/Voice | ElevenLabs | Voiceover, dubbing, IVR/customer service voices |
| Music | Suno | Background/production music, jingles |
| Molecular/Biological | AlphaFold, Claude Science | Protein structure prediction, drug candidate design |
Drug Discovery
Generative models are used to propose novel molecular structures with specific target properties, binding to a particular protein, having favorable solubility, avoiding known toxicity patterns, rather than searching through existing chemical libraries. This flips the traditional drug discovery workflow: instead of screening millions of existing compounds hoping one works, a generative model can propose new candidate molecules designed toward a target from the start, which pharmaceutical companies and biotech startups use to narrow the search space before the still-necessary and still-lengthy process of lab synthesis and clinical trials. It’s worth being precise about the limits here: generative models compress the early candidate-generation phase, they don’t shortcut the trial and regulatory process, which remains the actual bottleneck in bringing a drug to market. Anthropic’s Claude Science platform, launched in mid-2026, is a current real-world example of a lab betting on exactly this kind of workflow compression.
Related work in protein structure prediction and design (building on the same transformer architecture used in language models, applied to amino acid sequences instead of text) has become one of the more scientifically significant applications of the broader deep learning wave. Google DeepMind’s AlphaFold predicts a protein’s 3D structure directly from its sequence, and has now been used to help predict structures across virtually the entire universe of proteins known to science, materially accelerating structural biology research that previously depended on slow, expensive lab techniques like X-ray crystallography.
Design and Architecture
In industrial design, product design, and architecture, generative tools are used for rapid concept exploration: given a set of constraints (materials, load requirements, dimensions, aesthetic direction), generate a wide range of candidate designs to explore before committing to detailed engineering work. This is sometimes called “generative design” and predates the current LLM/diffusion wave by several years in CAD software, but has become considerably more capable and accessible as the underlying models improved.
The pattern is consistent with content creation: these tools are best understood as expanding the space of options a human considers early in a process, not as replacing the engineering judgment and constraint-checking that has to happen before anything gets built or manufactured. Autodesk’s Fusion generative design tools are a concrete, widely-deployed example of this pattern: an engineer defines constraints (material, budget, manufacturing method, load requirements), the software generates and ranks a large set of candidate designs against those constraints, and the engineer selects and refines from there rather than starting from a blank page (Autodesk Fusion, Generative Design for Manufacturing).
How Much of This Is Actually Adoption vs. Experimentation
It’s worth grounding all of the above against real enterprise-adoption numbers rather than only anecdotes, because the gap between “using generative AI somewhere” and “generative AI is producing measured business value” is larger than headlines suggest. McKinsey’s 2025 State of AI survey found 71% of organizations now regularly use generative AI in at least one business function, up from 65% a year earlier, and 78% use AI in some form across at least one function. But the same survey found only about a third of organizations are actually scaling those AI programs enterprise-wide, and more than 80% report no measurable impact yet on enterprise-level earnings from their gen AI use (McKinsey, The State of AI in 2025). That gap, widespread trial use against a much smaller base of programs actually scaled and measured, is the practical, unglamorous reality behind most “AI transformation” headlines, and it’s exactly the distinction the “measured outcome vs. demo” filter later in this piece is meant to help a reader spot.
Where the Value Actually Is (and Isn’t)
A useful filter for evaluating any generative AI product claim in the news: does it change how many options a human can consider before making a decision (genuinely valuable, low risk), or does it remove human review from a decision that has real consequences if wrong (much higher risk, and the source of most generative AI failure stories that make headlines). Content drafts, design exploration, and candidate molecule generation all fall in the first category. Fully autonomous decision-making in high-stakes domains without review does not, and that gap is exactly where a lot of the current caution and regulatory attention around generative AI deployment is concentrated.
When the Filter Gets Ignored: A Real Example
The Air Canada case is the clearest illustration of what goes wrong when that filter is skipped. In 2022, the airline’s customer-service chatbot told a passenger, incorrectly, that he could apply for a bereavement fare discount retroactively after booking. Air Canada refused to honor it and argued in a Canadian small claims tribunal that it “cannot be held liable for the information provided by the chatbot,” treating the model’s output as somehow separate from the company’s own website. The tribunal disagreed: it ruled the chatbot’s statements were negligent misrepresentation, that Air Canada owed its customers a duty of care to keep the chatbot accurate, and awarded the passenger damages (Moffatt v. Air Canada, 2024 BCCRT 149, CanLII). The case is now widely cited precisely because it’s mundane rather than dramatic: no exotic failure mode, just a generative system given an unsupervised, customer-facing decision role it wasn’t reliable enough for, exactly the risk category the filter above is meant to flag.
What to Watch For
The clearest signal of whether a generative AI application in a given industry is maturing past hype is whether it’s being measured against a concrete production metric, cost per asset produced, time saved per task, defect/error rate against a human-only baseline, rather than described only in terms of what the model is theoretically capable of. When you see coverage of a new generative AI product, that’s the detail worth looking for: is there a measured outcome, or just a demo. And per the Air Canada case above, it’s worth asking a second question too: is a human actually reviewing this output before it reaches a customer, or has that step quietly been skipped because the model is “good enough” most of the time.
