AI literacy at work means knowing when a language model can help, what information it may receive and how to check its output before anyone relies on it. You do not need to build models or write code; you do need to judge evidence, risk and human responsibility.
A fluent answer is not proof that a tool understood your task or got the facts right. The professional skill is knowing what to check, who can approve the result and when not to use the tool at all.
Know what an LLM can and cannot establish
A large language model, or LLM, generates responses by drawing on patterns in data. It can help reshape language by drafting an outline, summarizing supplied material, translating a passage or suggesting ways to explain an idea.
It cannot guarantee that a statement is true. It may produce a plausible explanation without evidence, miss a condition in a document or make an error in a chain of reasoning. It can sometimes perform reasoning tasks well, but a coherent explanation does not prove its conclusion.
The distinction matters most when a task moves from drafting to deciding:
- For a draft internal agenda, you can use the model to suggest a structure and edit it yourself.
- For a summary of an approved document, compare the output with the original, especially around exceptions and qualifications.
- For a recommendation about a customer, employee or financial commitment, the model's answer is not a substitute for the evidence and professional judgment behind the decision.
Even an ordinary task can become high-risk if its output is copied into a contract, sent to a client or used repeatedly at scale. Judge the use by what happens next, not just by how simple the prompt looks.
Choose a tool for the task, not for its fluency
Before using an LLM for work, decide what a successful result looks like and what would make an error costly. Then check whether an organization-approved tool is suitable for the task and the information involved.
If you need a summary, for example, test whether the tool preserves exceptions and can point you back to the relevant source passages. If you need a translation, check whether someone qualified can review the terminology and meaning. For a recommendation, decide what evidence and human expertise would be required independently of the model.
Compare tools on work that resembles your actual task. Use material you are permitted to share, with known answers or an authoritative source for comparison. Include difficult cases, such as a document with an exception or a request where the correct answer is "there is not enough information."
Look beyond the best response:
- Does the tool consistently follow the requested constraints?
- Can you verify its claims and find what it omitted?
- Are its access, retention and data-use settings acceptable for the task?
- Can the output be reviewed in less time than doing the work another way?
A model that writes a polished answer but creates an expensive checking burden may be a poor choice. If you cannot verify the result or use the necessary data safely, choose another workflow.
Spot failures that polished writing can hide
Some errors announce themselves. Others look like reasonable professional prose. Watch for these patterns, but do not treat their absence as proof of accuracy.
Unsupported facts and invented sources
An LLM may produce a figure, attribution or citation that looks credible but does not exist, or that exists without supporting the claim. This is often called a hallucination: an output presented as fact without a reliable basis.
Open the source independently. Check that the passage supports the exact claim, including its scope and any qualifications. A real document title is not enough.
Missing exceptions and changed meaning
Summaries often compress the very detail that determines what someone should do. Watch for words such as "always," "never" or "guarantees" where the source is conditional.
In an illustrative example, a source says a service is available subject to approval. A generated summary says it is available to everyone. The sentence is easy to read and materially wrong. Comparing it with the source, rather than asking the model whether its summary is accurate, reveals the problem.
Instruction drift
A response may satisfy the most visible part of a request while ignoring a quieter constraint: omit names, preserve a caveat, avoid making a recommendation or use only the supplied document.
Keep the original instructions beside the output. Check the negative instructions explicitly; they are easy to miss when the overall answer looks useful.
Bias and missing context
An output can apply uneven assumptions about people or situations, even when it contains no obviously offensive language. It may describe the same behavior differently depending on who is involved, for example, or recommend a standard process that overlooks an accessibility need.
Check whether the output applies the same relevant criteria across comparable cases. When the task affects people, ask whether the context is complete enough for a fair judgment and whether an LLM should be involved at all.
Evaluate an output before you use it
Use the following review in order. For a low-consequence draft, it may be quick. An output that will affect people, money, rights or external commitments needs more careful examination of the evidence and an appropriate reviewer.
Verify the claims against an independent source
Identify each fact, number, quotation and attribution that matters to the result. Compare it with an authoritative source you can access. Do not count the model's repetition of a claim as confirmation.
If you cannot find support, remove the claim, label it as unverified where appropriate or stop until someone can verify it. Ask the model to identify relevant passages if useful, but check those passages yourself.
Check every constraint
Compare the response with the task you actually set: audience, scope, required content, exclusions and format. Look especially for instructions the answer appears to have followed only partly.
A useful prompt can ask the model to flag uncertainty and distinguish supplied facts from suggestions. That makes review easier, but it does not guarantee compliance. The guide to advanced prompting can help you make instructions clearer.
Trace the logic and omissions
For a recommendation, separate the evidence from the conclusion. Ask: What facts support this? What assumptions connect them? Would the conclusion change if an assumption were false?
For a summary, compare the original's purpose, exceptions and unfavorable findings with the output. A summary can contain only true statements and still mislead by leaving out the decisive qualification.
Check relevance, tone and audience
Decide whether the answer fits the actual situation, not just the words in the prompt. A reassuring tone may be inappropriate if it hides an unresolved problem. A client-facing draft may sound authoritative while making a commitment nobody approved.
Edit the output as a professional document. Remove claims of certainty or promises that the evidence and your authority do not support.
Test a difficult variation
Where the task will be repeated, change one relevant assumption and see whether the output remains appropriate. For example, supply an approved test case with an exception, an incomplete source or a conflicting instruction.
A failed variation exposes a weakness to address. A successful one is not proof that the tool will handle every case. Keep the test material safe to share, and use the result to decide whether to revise the workflow, add a review gate or stop using the model for that task.
Set data boundaries before you paste anything
Data governance means the rules and responsibilities for deciding what information can be used, by whom and under what conditions. Apply those rules before entering a prompt, uploading a file or connecting a work system.
Start with the approved tool and account. Approval of a tool does not automatically mean every category of information may be entered into it. Check your organization's rules for customer information, personal data, confidential material, trade secrets and intellectual property. If the classification or permission is unclear, do not upload the material until the owner or responsible team confirms it.
Send the minimum information needed. For a general writing task, you might replace identifying details with placeholders and remove unnecessary background. But masking a name does not necessarily make a record anonymous: a combination of role, event and other details may still identify someone. Masking also does not override a rule that prohibits sharing the underlying information.
Check the settings and agreement for the specific workplace tool before use. Relevant questions include who can access prompts and outputs, how long they are retained, whether the provider may use them for model training and whether an approved opt-out is available and enabled. Turning off training does not by itself settle access, retention or permission to share.
Keep an appropriate record of the workflow: the source and classification of the information, the tool used, the purpose, the reviewer and the final decision. This is sometimes called lineage: a record of where information came from and how it was used or changed. Store that record in an approved place, without copying sensitive content into an unnecessary log.
Clarify who owns the source data, who can approve its use and what labels such as "confidential" mean in your organization. If a prompt may have exposed information improperly, stop using the conversation and follow your organization's incident-reporting process; do not assume deleting the chat resolves the issue.
For individual questions about legal obligations or data handling, consult qualified legal counsel and your organization's responsible data governance team.
Make review a decision, not a formality
The person who accepts or publishes an output remains responsible for using it appropriately. A reviewer needs access to the relevant evidence, enough expertise to challenge the result and authority to reject it.
Match the review to the consequences:
- Working draft with limited impact: The task owner checks accuracy, constraints and tone before use.
- Client-facing or consequential material: The task owner verifies the evidence and obtains review from the person authorized to approve the communication or decision.
- Sensitive, regulated or difficult-to-reverse use: Pause until the relevant specialist or formal approval process has assessed both the data handling and the intended use.
These are practical review gates, not a replacement for your organization's policy. If the person asked to approve an output cannot verify it, do not sign off more quickly; obtain the needed expertise or use another method.
For team-wide assessment of fairness, accountability and oversight, use the AI ethics auditing framework.
The members' bonus contains a reusable task brief and output audit record for applying this process.


