Technical AI Literacy Self-Assessment

My model of AI comes from media, vendor claims, or older chatbot experiences. I cannot say what has changed recently in capability, agent behavior, or reliability.

I use one chatbot and know it has gotten better over time. I do not yet distinguish model generations, reasoning modes, or enterprise versus free tiers.

I have compared at least two models on real tasks recently (i.e. Claude Opus 5 vs. Opus 4.8). I understand that model capability varies by task type, speed, cost, reasoning depth, and source-grounding, and I am thoughtful about when I select which model.

I keep a lightweight calibration habit: three to five recurring, work-like tasks with clear success criteria that I re-run when a major model or agent changes. I compare reliability, latency, cost, grounding, and failure modes, not just answer quality. I can say where the frontier has moved because I tested it.

When using AI, I often type a short question and may or may not be satisfied with the response.

I write conversational prompts and iterate when the first answer misses. I don't yet routinely provide source documents, examples, or context.

I routinely provide real context, including documents, strong examples, format requirements, and decision criteria. I use tools such as projects, custom instructions, memory, or knowledge bases to keep relevant context available, but I still assemble and steer the work task by task.

I turn recurring needs into reusable systems: standing instructions, curated knowledge, examples, decision rules, and saved workflows or skills with defined inputs, outputs, and checks. These systems apply my standards consistently across sessions or users, whether for day-to-day knowledge work or coding and agent work.

I am unfamiliar with agents.

My AI use is primarily synchronous. I do not assign tasks and "walk away."

I have delegated multi-step work, such as a deep-research task with citations or workflows that involve reading from, writing to, or organizing files. I have seen an agent plan, act, observe, iterate, and recover from at least one mistake.

I regularly delegate well-scoped or recurring tasks to agents. I may use the agent to clarify the goal and propose a plan, but before consequential actions I confirm the source boundaries, tools, permissions, approval points, stopping conditions, and verification criteria. For recurring workflows, I also define the trigger or cadence, exception handling, and escalation path. I supervise through checkpoints, evidence, and redirection rather than producing every step myself.

I have not used AI to create code, tools, automations, or working prototypes.

I have asked AI for a formula, script, or code snippet, for example generating an Excel formula, a short Python script, or a SQL query, but have not yet turned these into something I or others regularly use in our work.

I have used AI to create quick tools that I or others actually use: an internal calculator, dashboard, prototype, data-cleaning script, or form workflow. For each one I can answer three questions: what data does it touch, what happens when it breaks, and who maintains it. I treat these as prototypes and can say why they are not production infrastructure.

I have directed at least one consequential build and can review it technically. I document its goals, users, roles, data boundaries, permissions, and failure conditions. I inspect code changes, logs, test coverage, and critical security paths, and seek additional review when the stakes are high. Before the tool changes sensitive systems, I confirm backups or rollback options. I can explain what was tested, what remains uncertain, and why the evidence supports its use.

My use happens in a browser tab disconnected from my files, calendar, email, and systems.

I copy and paste between AI and my real tools. I am the integration layer.

I have connected AI to part of my safe stack (Drive, calendar, email, a document set) through built-in connectors. I understand that connectors, including those built with Model Context Protocol (MCP), allow AI systems to access external tools and data. I have felt the difference between an AI guessing about my work and one reading my actual documents.

I distinguish read-only connectors, write-capable tools, browser actions, file access, scheduled agents, and MCP-style integrations. I assess whether access exceeds task needs and understand risks such as prompt injection, data exfiltration, and combining internal and web access.

I have not yet developed a calibrated sense of when to trust output. I often dismiss it all or accept it all.

I know AI can hallucinate, but my skepticism is not yet methodical. My trust is based mostly on tone, confidence, or plausibility.

I verify proportional to stakes. I ground high-stakes work in provided documents rather than open-ended recall, spot-check citations, and treat fluent tone as non-evidence. I have caught AI being wrong (a fabricated citation, a confident miscalculation) in a way that changed my practice.

I design verification before relying on output. For agentic or high-volume work, I define "done and correct" up front and use tests, checklists, sampling, citations, or second-model review to check quality at scale. In work I review from others, I can recognize offloaded judgment: polished output the producer cannot explain or defend.

I cannot evaluate an AI product on its merits. A polished demo or an enthusiastic colleague's recommendation is indistinguishable from a good product.

I evaluate AI products mainly by features, use cases, cost, and privacy, but not by architecture, tools, permissions, evals, or failure modes.

When I encounter an AI product, whether in a demo, a colleague's recommendation, or a tool I am considering for my own work, I ask sharp questions because I can anticipate failure modes that marketing omits. I can distinguish a chatbot, workflow automation, app builder, coding assistant, connected assistant, and agent.

I interrogate claims at the architecture level. I ask which model is used; what tools and data it can access; whether access is read-only or action-taking; how context and memory are handled; how outputs are evaluated; how permissions, logs, and security are managed; what requires human approval; how cost and latency scale; and how the demo differs from production.

I do not know where my prompts and data go, what vendors retain or train on, or which of my tools are cleared for sensitive data.

I know sensitive data shouldn't go into consumer AI tools, but my decisions are ad hoc. I could not say which of the tools I use are covered by an enterprise agreement versus a consumer account, or what my organization's policy allows and prohibits.

I operate within my data boundaries deliberately. I know which surfaces I work on (consumer versus enterprise accounts, and why the distinction matters for retention and training), I know which classes of data may enter which of my tools, and I check before enabling a new connector or introducing a new class of data. For any tool I adopt on my own, I read the retention and training clauses first.

I verify rather than assume. I can check what a connector or agent I enable can actually access, not just what its documentation claims. I notice vendor policy changes that affect my work, such as a lab changing its retention terms. And I map model dependency in my own workflows: I know which depend on which models, and what I would do if one were deprecated, degraded, or pulled.

I am not aware of how AI pricing works or that model choice carries cost implications.

I know what a subscription costs per seat. When comparing options I think in sticker price or price per token.

I understand that a cheaper per-token model can be more expensive per completed outcome, that agentic workflows multiply token consumption, and that the real price is tokens × price × attempts.

I make cost-aware selections as a habit in my own work. I match model tier to task value rather than sending the most expensive model to do routine work, I set sensible model defaults or routing in my recurring workflows, and I notice when agentic work is multiplying consumption and adjust. I understand the tradeoff between seat-based and usage-based pricing well enough to explain when each favors the buyer.