[Day 7] Five Questions That Define Five Years of GenAI
[Day 7] Five Questions That Define Five Years of GenAI
Every year since ChatGPT launched, I've found myself asking a completely different question about generative AI—not the same question with better answers, but an entirely new question that only became possible because the previous one had been resolved. In 2022, I wondered what AI actually knew (and learned the hard way about hallucinated salary data in job descriptions). By 2025, I was producing interactive training decks and Excel templates without writing a single line of code. Now, at the start of 2026, I'm asking how to build automated workflows that free me up to focus on what humans do best. These five questions trace the extraordinary pace of change in just five years—and reveal where the real opportunities lie for knowledge workers ready to ask the next question
Every year since ChatGPT launched in November 2022, I've found myself asking a different fundamental question about generative AI. Not the same question with updated answers, but genuinely different questions, each one only possible because the previous year's question had been resolved.
Looking back, these questions trace the arc of how quickly this technology has evolved. What seemed like science fiction in 2022 became table stakes by 2024. What felt cutting-edge in 2024 is now basic functionality. The pace is disorienting, but the questions themselves tell a coherent story about what's changed and what it means for knowledge workers like me.
Today's Experiment
I mapped the five fundamental questions I asked about GenAI from 2022 to 2026, along with real examples of how each question played out in my work in the field of HR. This isn't a technical deep-dive. It's a practitioner's field guide to how fast the ground has shifted.
The Five Questions at a Glance
Year
The Question
What Changed
2022
What does GenAI actually know?
Discovered training data cutoffs and hallucination limits
2023
Can I give GenAI my own information to work with?
Document uploads and RAG made AI work with your data
What useful outputs can I actually produce with GenAI?
Artifacts: interactive HTML, Excel templates, Word docs, presentations
2026
How can I automate workflows and build scalable playbooks?
No-code integrations, reusable systems, efficiency at scale
2022: What Does GenAI Actually Know?
When ChatGPT launched in November 2022, the first question everyone asked was: what does this thing actually know? The answer turned out to be both impressive and limiting. ChatGPT had been trained on internet data up to 2021, which meant it couldn't tell you anything about recent events. Ask about 2022 trends or current best practices, and you'd get either a polite refusal or, worse, a confident fabrication.
Example: I asked ChatGPT to help me improve a job description for various roles that we were hiring for. The AI produced decent general language, but when I asked it to include current market salary benchmarks or reference recent skills in demand, everything it generated was outdated or simply made up. It cited salary ranges from 2020 data and mentioned tools that had since evolved significantly. The output looked professional but required complete verification against current sources.
The lesson: GenAI in 2022 was like a brilliant colleague who'd been in a coma since 2021 and had a habit of filling gaps in their knowledge with confident guesses. Useful for improving the structure and language of job descriptions, but you couldn't trust it on current market data or recent developments.
2023: Can I Give GenAI My Own Information?
Once we understood the knowledge limits, the natural next question was: what if I could give it my own data? 2023 answered this with document uploads, longer context windows, and a technique called RAG (Retrieval-Augmented Generation) that let AI pull from external sources.
Claude 2 launched with a 100,000-token context window, meaning you could upload entire policy documents and have the AI work with them. GPT-4 added document processing. Suddenly, the AI wasn't limited to its training data. It could work with your company's specific information.
Example: I was researching HR policy updates for new markets and needed to understand employment law changes will inform our employment agreement. Instead of spending hours parsing dense legal text, I copied and pasted the relevant legislation directly into Claude and asked specific questions: "Based on this law, are we required to provide paid leave for X situation?" and "What's the interpretation of 'reasonable accommodation' in this context?" The AI analyzed the actual legal text I provided and gave me clear interpretations with references to the specific sections. It wasn't legal advice, but it dramatically accelerated my research and helped me formulate the right questions to bring to our legal team.
The lesson: The 2023 breakthrough wasn't smarter AI. It was AI that could work with your specific source material instead of just its general training. This shifted GenAI from a general knowledge tool to a research assistant that could analyze the exact documents you needed to understand.
The lesson: The 2023 breakthrough wasn't smarter AI. It was AI that could work with your specific context instead of just its general training. This shifted GenAI from a general knowledge tool to a personalized work assistant that understood your company's way of doing things.
2024: What Formats Can GenAI Work With Beyond Text?
By 2024, text felt limiting. The question became: what else can AI understand and create? The answer was multimodal capabilities. GPT-4o processed images, audio, and video alongside text. Claude gained vision capabilities. AI image generators matured. Specialized tools emerged for presentations, diagrams, and data visualization.
Example: During a goal brainstorming session, I photographed a ideas handwritten on sticky notes and uploaded the image to Claude and asked it to interpret the image and generate a summary of the ideas listed. The AI correctly was able to accurately transcribe the handwritten text, saving time to manually transcribe a messy brainstorm into a structured planning document that I could use and share.
The lesson: 2024 broke GenAI out of the text-only box. The marked improved ability to process images, understand visual information, and work across formats made AI useful for a much wider range of HR work, from interpreting data to digitizing whiteboard sessions.
2025: What Useful Outputs Can I Actually Produce?
Understanding information is one thing. Creating polished, usable deliverables is another. In 2025, my question shifted to: what tangible artifacts can I produce with GenAI that I can actually use in my work?
The answer surprised me. Claude's artifacts feature and similar capabilities in other tools meant I could generate not just text, but functional outputs: interactive HTML pages, formatted Word documents, Excel templates with working formulas, and more. All without writing a single line of code.
Example: I needed to create a training deck for one of our Sales teams. Instead of spending hours in PowerPoint, I described the content to Claude and asked for an interactive HTML presentation. In under 30 minutes, with some iteration, I had a fully functional training module, less than 1 MB in size, with clickable navigation, embedded timers, and professional styling. It ran in any browser with no special software required.
Another example: I asked Claude to create Excel templates for customized onboarding checklists and manager coaching guides in Word format. The Excel template included properly formatted cells, dropdown menus for status tracking, and conditional formatting that highlighted overdue items. The Word coaching guide had a professional layout with space for notes, reflection questions, and action items. What would have taken me half a day of formatting work was done in minutes. The templates needed some refinement to match our branding, but the heavy lifting was complete.
The lesson: 2025 was the year GenAI stopped being just a writing assistant and became a production tool. The ability to generate functional artifacts, not just drafts, changed the equation from "AI helps me write" to "AI helps me build."
2026: How Can I Automate Workflows and Build Scalable Playbooks?
Now that I can produce useful outputs, my question has evolved again: how do I systematize this? How do I build workflows and playbooks that let me scale both efficiency and accuracy, not just for one task, but for repeatable processes?
We're just at the start of 2026, and I'm genuinely excited to experiment with and document how to use AI to design systems that automate workflows. The tools are ready: Claude now connects natively to Google Drive, Notion, and other work tools. Zapier's integration lets you connect AI to over 8,000 applications without code. The Model Context Protocol (MCP) has standardized how AI connects to external services.
What I'm exploring: Imagine a playbook where a manager submits a coaching request through a simple form. The submission automatically triggers AI to pull relevant context, generate a customized coaching framework based on our methodology, and draft a preparation guide, all delivered to my inbox for review before I meet with the manager. The system handles the repetitive assembly work; I focus on the judgment calls about what this specific situation actually needs.
Another possibility: An onboarding workflow where new hire information flows through automatically, generating customized welcome materials, first-week schedules, and stakeholder introduction emails, all templated to our standards but personalized to each role and individual. What currently takes hours of manual coordination could become a system that runs with human oversight rather than human assembly.
The opportunity: The goal isn't to remove humans from the process. It's to free up time from repetitive assembly work so I can focus on what humans do better: strategy, judgment, relationship-building, and thoughtful design. If AI can handle the templating and drafting, I can spend more time actually coaching managers, understanding team dynamics, and thinking strategically about talent.
What I Learned Today
Each year's question only becomes possible because the previous year's question was answered. You can't ask "what outputs can I create?" until you've solved "what formats can AI handle?" You can't ask about automation until you know what's worth automating. The questions build on each other.
The pace is genuinely unprecedented. Five years from "what does it know?" to "how do I build automated workflows?" represents a compression of capability that normally takes decades. Understanding this pace matters for anyone planning their skills development.
The human role keeps moving upstream. In 2022, humans verified facts. In 2023, humans provided context. In 2024, humans chose formats. In 2025, humans refined outputs. In 2026, humans design systems. The pattern is clear: as AI handles more execution, humans focus more on strategy, judgment, and design.
The competitive advantage now is in asking the right question for the current moment. Someone still asking 2022's question ("what does AI know?") while others are asking 2026's question ("how do I build scalable workflows?") will fall behind. Staying current with what's possible matters more than mastering any single capability.
Try It Yourself
Identify which question you're currently asking about GenAI, then try experimenting one level ahead:
If you're still at "what does it know?" → Try uploading a work document and asking AI to analyze it. Experience the shift from general knowledge to your specific context.
If you're at "can I give it my data?" → Try uploading an image or screenshot and asking AI to interpret it. Experience the shift from text-only to multimodal.
If you're at "what formats work?" → Try asking Claude to create an artifact: an HTML page, an Excel template, or a formatted Word document. Experience the shift from drafts to deliverables.
If you're at "what can I produce?" → Try connecting AI to another tool via Zapier or native integration. Experience the shift from manual to automated.
Where I'm headed: This year, I'll be documenting my experiments in building these automated workflows and playbooks. The goal is to share what works (and what doesn't) so others can learn from the process. If you're interested in following along, that's exactly what this blog is for.
I've tried using Claude, ChatGPT, and Gemini to create presentations and the results were always underwhelming. The content was decent, but the formatting? A manual nightmare. So when I tested Gamma.app with a genuinely complex use case (employee offboarding compliance across six Asia-Pacific countries), I wanted to see if a specialized tool could actually deliver what general-purpose AI assistants can't: polished, presentation-ready output in minutes, not hours. Here's exactly what happened, step by step, including the limitations they don't tell you about upfront.
What happens when you give three AI tools the same complex virtual training scheduling problem and one of them walks straight into a cultural landmine? I ran an experiment with Claude, ChatGPT, and Gemini: identical prompts asking each to recommend the best approach for scheduling a 100-person virtual training across six APAC countries during Lunar New Year season. One tool recommended a culturally tone-deaf date to save $2,500, and one framed the time zone constraint as a "Golden Window" that actually helped me think through the problem differently. This is what AI collaboration looks like in practice and why the human still needs to be in the loop.
I uploaded one PDF chapter on biomechanics and five minutes later had a polished infographic, a 12-question quiz, 51 flashcards, and a podcast episode—all generated by AI, all grounded exclusively in my source material with zero hallucinations. No Canva, no Anki, no quiz app, no piecing together outputs from different tools. Today’s experiment explores NotebookLM, Google’s quietly powerful (and still free) learning tool that handles production while you handle curation, and what it reveals about the shifting role of humans in AI-augmented learning.