Super Intelligence master guide › Energy, work, learning and safety series › Article 0015
AI literacy is the ability to use an AI system to complete useful work while understanding its evidence, limits and authority. It includes choosing a suitable task, explaining the desired result, checking consequential details and deciding which actions remain under human control. A well-written request helps, but literacy extends across the entire job: preparation, assistance, review, delivery and correction.
This matters because an assistant can produce something that looks finished before the underlying work is reliable. A polished comparison can omit an important condition. A helpful schedule can conflict with an appointment it never received. A confident recommendation can rest on an assumption the user did not intend. Becoming skilled at using AI therefore means making the work inspectable, rather than treating fluency as a completion signal.
The Super Intelligence series uses SI as a broad public-facing theme for machine intelligence. Current AI assistants should be assessed through demonstrated capabilities on particular tasks. Hypothetical technical superintelligence would involve capabilities far beyond those established by an ordinary successful interaction. The practical methods below do not require that hypothetical threshold: they help a reader evaluate the assistance available now.
Start with the result, then choose the tool
The first decision is what should exist when the task is complete. A useful result might be a meeting brief, a comparison of supplied proposals, a revised application, a checklist or a set of questions for further investigation. Each result has different requirements. A brief needs faithful compression. A comparison needs consistent criteria. A checklist needs completeness within a defined scope.
Beginning with a tool instead can encourage activity without a clear purpose. Repeated requests for more detail may create a longer document without making the decision easier. The remedy is to state the intended use. A comparison for choosing a supplier should explain the tradeoffs that affect that choice; it need not reproduce every sentence from every proposal.
Consider a hypothetical administrator preparing a meeting. The immediate problem is not lack of prose. It is uncertainty about which decisions must be made and what evidence supports each option. The useful AI task is to extract decision points, connect them with supporting documents and identify unresolved questions. A beautifully worded summary that hides those questions would fail.
This way of defining work also reveals when another tool is better suited. A calculator can provide a deterministic total. A spreadsheet can expose a formula. A calendar can show a confirmed commitment. An AI assistant may help organize the surrounding explanation, while those tools supply the result that must be exact.
An everyday application is delegation with explicit completion conditions; the brief should identify the accepted result, permitted actions and circumstances requiring clarification.
Understand the three kinds of assistance
AI literacy becomes clearer when assistance is divided into producing information, recommending a choice and taking an action. These categories can appear in one conversation, but they deserve separate treatment.
Producing information includes summarizing supplied notes, drafting text or explaining a concept. The output remains available for inspection before use. Recommending a choice adds a judgment about what should happen. That judgment depends on criteria, priorities and missing information. Taking an action changes something outside the conversation: a file, account, appointment, message or business record.
A request to prepare an email belongs to the first category. A request to suggest whether it should be sent belongs to the second. A request to send it belongs to the third. Completion of one does not automatically establish authority for the others.
The practical benefit of this distinction is precise delegation. A user can ask an assistant to prepare alternatives and explain their implications while retaining the final decision. Where action is authorized, the task can specify its destination, scope and stopping conditions. This makes the system’s responsibility easier to review.
The distinction also improves diagnosis. If the draft is inaccurate, the problem concerns information quality. If the recommendation uses the wrong priorities, the problem concerns judgment. If the assistant acts beyond permission, the problem concerns authority. Those failures require different repairs, even when they appear in the same workflow.
A useful mental model starts with the difference between models, retrieval and tools; different components explain why an assistant may produce information, rely on a source or perform an action.
Write a task brief that reduces hidden assumptions
A strong task brief explains the objective, available material, relevant constraints and form of the deliverable. It also identifies what the assistant should do when information is missing. That final instruction prevents a gap from silently becoming an invented answer.
For example, a hypothetical procurement request might ask for a comparison of three supplied maintenance proposals. It can specify that price, response time, exclusions and renewal terms matter, and that absent terms should be marked as unknown. The requested result could be a short decision brief supported by a detailed comparison.
The brief should distinguish a constraint from a preference. A fixed submission date cannot be traded away casually. A preference for a shorter report can be adjusted if an important exception needs explanation. When these are mixed together, an assistant may optimize the presentation while overlooking a binding condition.
It can also identify the intended reader. A specialist may need technical distinctions; a general reader may need an explanation of why those distinctions matter. This is not a request to distort the underlying evidence. It is a way to make the same evidence usable.
A task brief need not be long. Its value comes from resolving uncertainty that would otherwise affect the result. If a sentence does not change the work, it may be unnecessary. If a missing sentence could change the conclusion, it belongs in the brief.
Build a small evidence pack
An evidence pack is the set of materials the assistant is allowed to use for the task. It might contain current instructions, original documents, relevant records and a note explaining which versions supersede others. The aim is to make the basis of the work visible.
Suppose a hypothetical team has two versions of a policy. The older version contains a permission that the newer one removes. Supplying both without indicating their relationship creates avoidable ambiguity. A short note that identifies the current version can be more useful than adding another page of instructions.
The same principle applies to numerical data. A revenue figure should travel with its period, unit and definition. A percentage should identify its denominator. A price should indicate whether tax, installation or recurring charges are included. The assistant cannot reliably compare quantities whose meanings remain hidden.
For everyday use, the evidence pack can be modest. A meeting agenda, three background documents and a list of decisions may be sufficient. Adding unrelated material can make important details harder to locate and increase the review burden.
This practice also gives a clear way to challenge an answer. If a conclusion cannot be connected with a document, stated assumption or transparent calculation, it should be treated as unresolved. The evidence pack therefore improves both preparation and verification.
The evidence pack supports a traceable research and verification process; preserving the claim, source and uncertainty makes later review more than a judgement of writing style.
Define acceptance criteria before inspecting the answer
Acceptance criteria describe what makes the result usable. They should concern the work itself, rather than how impressive the output looks. A report can be elegant and still fail if it ignores the requested period or compares unlike categories.
For a proposal comparison, useful criteria include covering every proposal, using the same criteria throughout, preserving material exclusions and distinguishing unknown terms from unfavorable terms. For a meeting brief, criteria might include identifying each required decision and connecting it to an authoritative source.
A hypothetical learner requesting an explanation can use different criteria: the explanation should resolve a named confusion, include an example and leave the learner able to solve a related problem. The quality of the answer then depends on what becomes possible afterward.
Criteria should be proportionate. A personal packing checklist does not need the evidence process of a financial commitment. The relevant question is what failure would matter and how it could be detected before use.
Setting criteria in advance reduces a common temptation to accept an answer because it arrived quickly. The review becomes a comparison between requirements and results. It also makes revisions more precise: instead of asking the assistant to improve everything, the user can identify the requirement that remains unmet.
Worked example: prepare a meeting without manufacturing agreement
Imagine a fictional operations team preparing a meeting about a delayed project. The available material contains a timeline, two status notes and an email identifying a missing dependency. The goal is to support a decision about the next stage.
A weak request asks for a positive meeting summary. That framing risks suppressing the very uncertainty the meeting needs to resolve. A stronger request asks for the current status, confirmed causes of delay, assumptions that remain uncertain and the decisions requiring approval.
The assistant’s output should distinguish a recorded fact from an interpretation. A document may establish that a delivery arrived late. It may not establish why. The brief can list a possible cause as a question for investigation without making it the official explanation.
Review then focuses on specific details. Does the timeline match the records? Are the decision owners correctly identified? Is a proposed date described as proposed? Have contradictory statements been preserved rather than blended into an artificial consensus?
The finished product is valuable when it supports a better meeting. It can shorten the time spent finding information and improve the clarity of decisions. It should not create agreement that the team never reached. The useful assistance lies in organizing evidence, leaving actual commitments to the authorized process.
Worked example: compare purchases using consistent terms
Consider a fictional household comparing three appliances. One price includes delivery, another includes installation and a third is a promotional price with conditions. An immediate ranking by headline price would create an uneven comparison.
The task should ask for a common cost basis and identify missing information. If the installation cost is unknown, the assistant can produce a conditional range or flag the gap. It should not invent a charge simply to complete the comparison.
The decision also depends on priorities. A smaller appliance might fit the available space; a cheaper one might have a longer delivery window. Asking for the cheapest option without stating the constraints can lead to a recommendation that is impractical.
A useful output therefore separates facts, assumptions and preferences. Facts come from the supplied specifications and prices. Assumptions concern unresolved charges or usage patterns. Preferences concern convenience, space and acceptable waiting time.
The user can then change a preference without rewriting the facts. If speed becomes more important than price, the ranking can be reconsidered transparently. This illustrates a durable benefit of AI literacy: the assistant helps build a decision structure that remains understandable when circumstances change.
Worked example: organize a week around real commitments
A hypothetical weekly planner contains work deadlines, a class, travel time and several optional activities. The assistant can help arrange these into a workable sequence, but it needs to know which entries are confirmed and which are flexible.
The first pass can identify conflicts and missing information. A meeting and a class may overlap. A task may have an estimated duration but no deadline. A journey may require a buffer. These are planning questions, not errors that should be hidden in a tidy calendar.
The second pass can propose alternatives. Moving an optional activity may resolve a conflict; shifting a confirmed appointment may require another party’s agreement. The output should distinguish what can be changed privately from what needs coordination.
A useful review checks the plan against reality. Does it include preparation time? Are transitions possible? Is every available hour filled, leaving no room for variation? A schedule that optimizes theoretical capacity can be less useful than one that preserves a margin.
The AI contribution is the arrangement and examination of options. The user supplies the actual commitments and decides which tradeoffs are acceptable. If the assistant has calendar access, permission to inspect entries should remain distinct from permission to alter them.
Learn to verify according to the kind of claim
Different claims require different checks. A statement about a supplied document can be checked against that document. A calculation can be recomputed. A current external fact needs an appropriate current source. A recommendation needs examination of its criteria and assumptions.
Asking the same system to confirm its own answer can help reveal a contradiction, but it does not necessarily provide independent evidence. If the original error came from a missing source, another fluent explanation may preserve it. A stronger check returns to the source or uses a method that does not share the same weakness.
For numerical work, preserve the intermediate reasoning in a form that can be inspected. If a total comes from several expenses, list the items and identify exclusions. If a percentage describes growth, show the starting and ending quantities. This makes a mistake easier to locate.
For prose, inspect consequential changes in meaning. A summary might turn a possibility into a commitment or remove an exception. Those changes can matter more than a spelling error.
Verification is therefore selective attention guided by consequence. Check the parts on which a decision depends, then broaden the review when the evidence suggests a deeper problem. A uniform demand to recheck every word can consume time without improving the most important result.
Diagnose weak results before changing the request
A disappointing result does not always mean the model lacks capability. The task may be ambiguous, the evidence incomplete, the comparison unfair or the requested format unsuitable. Diagnosis begins by identifying where the result diverged from the intended work.
If the assistant answered the wrong question, revise the objective. If it missed a document, improve the evidence pack. If it produced unsupported claims, require a distinction between supplied facts and inference. If it followed the instructions but the result is still unusable, reconsider the task design.
This prevents an endless cycle of adding adjectives such as detailed, accurate or comprehensive. Those words describe an ambition without identifying the missing mechanism. A concrete correction explains what should change and how success will be assessed.
A fictional report may be too long because it includes every available detail. The repair is to define which decisions the report supports and move secondary material into a supplement. Another report may be too short because it omits tradeoffs. The repair is to specify the comparison criteria.
Repeated failures also justify reducing scope. A system that struggles with a complete workflow might still assist with one bounded stage. Appropriate delegation preserves useful capability while keeping unresolved weaknesses visible.
Measure time saved after review and correction
A fast first answer does not establish a productivity gain. The complete cost includes preparing the task, inspecting the output, correcting errors and integrating the result into the actual workflow. Measuring only generation time hides those stages.
A hypothetical comparison takes twenty minutes to prepare and forty minutes to complete manually. With assistance, preparation takes fifteen minutes, generation takes two and review takes twenty. The total gain is meaningful, but it is smaller than a comparison between forty minutes and two minutes would suggest. These invented numbers illustrate accounting, rather than a measured result.
The quality of the final work also matters. If the assisted process produces a better comparison at the same total time, it may still be valuable. Conversely, a faster result with missing exclusions can increase the cost of the eventual decision.
A practical trial can record a few representative tasks, the time required and the repairs needed. The aim is to discover where assistance fits, rather than manufacture a universal performance score.
Such a record helps distinguish learning costs from recurring costs. Preparing a reusable brief may take effort initially and save effort later. Repeatedly repairing the same error suggests a weakness that should be addressed before expanding responsibility.
Preserve context without surrendering control
Recurring tasks benefit from continuity. A note that records definitions, decisions and unresolved issues can help the next session begin from the right place. However, stored context should remain reviewable. An old assumption can become misleading when circumstances change.
A useful continuity note explains what is current, why a decision was made and what would trigger reconsideration. It can also identify material that was deliberately excluded. This reduces the risk that the next assistant treats every old statement as an active instruction.
Privacy belongs in this decision. The relevant question is what information the task actually requires and whether the chosen service is suitable for handling it. Workplace and school policies may determine the available tools and permitted data. Familiarity with an interface does not establish permission to upload confidential material.
For a hypothetical recurring budget task, category definitions and approved formulas may be worth retaining. Full personal account records may be unnecessary for explaining the method. Separating method from sensitive inputs can reduce exposure without preventing useful work.
Control also requires a usable exit. The final result should be understandable outside the conversation, with essential assumptions recorded. If the work cannot be continued without reconstructing a long exchange, the assistance has created a new dependency.
Build AI literacy through short, inspectable practice
Practice is most useful when the learner can judge the result. Begin with material already understood: summarize a familiar document, compare two known alternatives or rewrite a draft while preserving its meaning. This makes errors visible and develops the habit of examining them.
The next stage introduces uncertainty deliberately. Supply a document with a missing term and observe whether the assistant identifies it. Ask for a comparison containing different units. Request a summary of two versions with a change in policy. These exercises test attention to evidence rather than admiration for fluent prose.
Keep a short record of what went wrong and which correction worked. Over time, the record becomes a personal collection of useful task patterns. It should include cases where assistance was unnecessary, because choosing not to delegate is also a skill.
Workplaces can develop shared examples using fictional or appropriately permitted material. A demonstration should show the original request, the output, the review and the corrected result. Showing only the polished answer conceals the judgment required to produce it.
The aim is independent competence. A capable user should become better at defining work, recognizing uncertainty and deciding when another source or method is required. Those abilities remain useful as interfaces and model capabilities change.
The human consequence should include capability that remains when assistance is removed; purposeful AI use is stronger when the user can still recognise errors and apply the underlying idea independently.
Make workplace expectations explicit
AI literacy has an organizational dimension. A team needs a shared understanding of what assistance is permitted, what disclosure is expected and who remains accountable for the final result. Otherwise, similar tasks can receive inconsistent treatment.
For example, a fictional service team may permit assistance with drafting responses while requiring review before external delivery. The rule should explain which information may be used and which commitments need authorization. A vague instruction to use AI responsibly leaves those practical questions unresolved.
Teams also need a route for reporting failures. An incorrect summary might reveal a task-design problem that affects colleagues. A repeated omission might justify changing the evidence pack. Treating every repair as a private inconvenience prevents useful learning.
Evaluation should include the work that was displaced as well as the work created. Assistance may reduce drafting effort while increasing review effort. A team can adjust roles accordingly, rather than assuming the initial task boundary remains optimal.
Clear expectations support experimentation. When the permitted scope is understandable, staff can try useful approaches without guessing whether a successful draft authorizes further action. The organization gains evidence about actual value while preserving control of its commitments.
Recognize when the problem needs clarification
Some tasks fail because the desired result cannot be determined from the available request. A missing preference is different from a missing fact. If two options are equally supported but serve different priorities, more research may not settle the choice. The user must decide which priority matters.
A fictional travel plan illustrates this distinction. The assistant may know the available routes and approximate journey times, yet cannot infer whether the traveler values fewer changes, lower cost or a later departure. Asking for the best route hides that preference inside a seemingly factual question.
A useful response can present the tradeoff and request the one missing decision that changes the result. It need not interrupt work for every minor detail. Preparation that remains useful across alternatives can continue while the consequential preference is unresolved.
The same method applies at work. If a proposal requires a decision about service level, producing increasingly detailed technical summaries will not substitute for that decision. The assistant can explain what each level implies and identify the information needed, while keeping the choice visible.
This is a valuable part of literacy because it prevents false certainty. Recognizing a question that only the user or organization can answer is often more useful than generating another plausible recommendation.
Treat revisions as changes in meaning
Revision is a common AI task, and it deserves more attention than a simple before-and-after impression. A rewritten document may become clearer while also changing what it promises. Review should therefore examine meaning as well as style.
Suppose a fictional draft says a service can usually respond within two working days. A smoother revision might say the service responds within two days. The revised sentence has removed both a qualifier and a definition of the period. It now communicates a stronger commitment, even though the wording is shorter.
A good revision brief can identify protected details: obligations, exceptions, quantities, dates and statements of uncertainty. The assistant may improve their presentation while preserving their substance. Where a change is proposed, it should be distinguishable from a stylistic edit.
For an important document, compare the final version with the original at the level of claims. Which statements were added? Which were removed? Which became more definite? This approach can reveal changes that a quick reread overlooks.
The method also protects the author’s intention. Clearer prose should make the intended meaning easier to understand. It should not quietly replace that meaning with one that sounds more persuasive. Skilled use of AI includes controlling that difference.
Design assistance for the reader’s actual situation
A technically complete answer can still be unusable if its presentation demands more time, attention or background knowledge than the reader has. AI literacy includes adapting the result to the situation in which it will be used.
A fictional maintenance note written for a specialist may contain terminology that a household reader cannot interpret. Simplifying it can improve access, but the simplification must preserve the distinctions that affect the next action. Replacing an unfamiliar technical term with a misleading everyday word would create a new problem.
Ask for a layered result where appropriate: a short explanation of the decision, followed by the detail needed to examine it. The concise layer helps the reader orient; the deeper layer preserves the evidence. Neither should contradict the other.
Accessibility can also require a change in format. A dense block of instructions may be easier to use as a short sequence, while a nuanced comparison may require connected prose. The best format follows the work, rather than a universal preference for brevity or length.
Finally, examine whether the output assumes resources the reader lacks. A plan that requires unavailable software, specialist knowledge or uninterrupted time may need adaptation. The assistant is useful when it helps fit a method to real constraints, instead of presenting an idealized workflow that cannot be followed.
Frequently asked questions about AI literacy
Is AI literacy the same as writing better prompts?
Prompting is one part of AI literacy. A request can improve the chance that an assistant understands the task, but the user still needs suitable evidence, a way to inspect the answer and clarity about action authority. A beautifully phrased request cannot supply information that was never available.
A useful comparison is project preparation. Explaining the job matters, yet the result also depends on the materials, acceptance criteria and review process. Developing all of these together makes assistance more dependable than collecting elaborate request templates without understanding their purpose.
How can a beginner decide what to delegate?
Choose a bounded task with an inspectable result and a manageable cost of error. A familiar document summary is often easier to evaluate than an unfamiliar technical recommendation. A draft is easier to review before use than a completed external action.
State what success means and compare the result with that standard. If review requires knowledge the user does not possess, obtain an appropriate independent check or reduce the task. Expand delegation when repeated experience shows that the output is useful after review, rather than because the system sounds confident.
Should every answer be checked with another AI system?
Another system can provide a different perspective, but agreement alone does not establish correctness. Two systems may rely on similar information or share a missing assumption. The strongest check depends on the claim: original records for a summary, an independent calculation for arithmetic and an authoritative source for an external fact.
Use additional AI assistance when it helps identify questions or discrepancies. Preserve a route to evidence that can resolve them. The purpose is to improve the basis of the answer, not merely increase the number of answers that resemble it.
What happens when an assistant becomes more autonomous?
Greater autonomy increases the importance of defining scope, permissions and stopping conditions. A system that can edit, send or schedule may complete more stages of work, but each external effect creates something that must be controlled and, where necessary, reversed.
Begin by separating preparation from execution. Specify the exact action authorized and the conditions under which it should pause. Inspect the evidence of what happened afterward. Autonomy becomes useful when it reduces effort while keeping authority and consequences understandable.
Does using AI reduce the need to learn the underlying skill?
The answer depends on the objective. If the task is routine production, assistance may appropriately reduce manual effort. If the objective is to develop competence, replacing all practice can remove the experience needed to evaluate future work.
Preserve opportunities to explain decisions and complete representative tasks independently. An assistant can support learning through examples, feedback and questions, but a finished output does not demonstrate the user’s understanding. Durable literacy includes the ability to recognize when the system’s assistance is weak and to continue without it.
Continue through the series
The wider system appears in Super Intelligence: Energy, Work, Learning and Safety. Explore task completion in 0012 — Progress Measured by Work You Can Delegate and evidence handling in 0014 — Research at the Speed of a Question. For a setting where evidence requirements are particularly consequential, continue to 0016 — Healthcare Promise and the Evidence It Requires.
Previous: 0014 — Research at the Speed of a Question · Next: 0016 — Healthcare Promise and the Evidence It Requires