In this guide · 8 sections
Responsible AI use at work means knowing what you can safely share with an AI tool, matching how carefully you check its output to what is at stake, and thinking about everyone your AI-assisted work affects. It is the Safety & Responsibility dimension of AI fluency, and it is what lets you use AI with confidence rather than caution. This guide shows what it looks like at every level, gives you six exercises to try this week, and lists the best free resources, including what the EU AI Act asks of organisations.
What Responsible AI Use Means at Work
Responsible AI use is a set of everyday practices, not a policy document. AISA measures it as one skill, AI Safety & Responsibility, which shows up in four habits: protecting data, scaling your checks to the stakes, considering the impact on others, and helping shape the norms around you.
Know your data boundaries
The first habit is knowing what not to paste into an AI tool: personal data about other people, client details, unreleased financials and anything confidential. It also means knowing how your tools handle data, from training settings to retention. AI data privacy and data classification for AI cover the basics.
Match your checks to the stakes
A brainstorm needs no review; a client contract or a medical summary needs a careful one. Stakes-based review means deciding up front how much checking each task needs, and when a second person should look before anything leaves your hands.
Think about who else is affected
AI-assisted work reaches other people: customers, candidates, colleagues. Strong practitioners ask who is affected if the output is wrong or biased, and design their use around that. AI bias awareness is a good starting point.
Shape the norms around you
At the top end, people help their team decide which tools are approved, what data can go where, and when AI use should be disclosed. Frameworks such as the NIST AI Risk Management Framework give that work a structure.
Why Responsible AI Use Matters Now
Most knowledge workers already use AI at work, often with tools their employer did not choose and without training. In Microsoft and LinkedIn's 2024 Work Trend Index, 78% of AI users said they bring their own AI tools to work, and only 39% had received AI training from their company.
Data is leaving organisations
The University of Melbourne and KPMG's global study of more than 48,000 people found that 48% of employees have uploaded company information into public AI tools, and 57% admit to using AI in ways that are not transparent, such as presenting AI-generated work as their own.
Incidents are rising
Stanford's AI Index 2026 counts 362 documented AI incidents in 2025, up from 233 in 2024, and reports that knowledge and training gaps are the top obstacle organisations cite in putting responsible AI into practice.
The EU AI Act expects AI literacy
Article 4 of the EU AI Act, as amended in 2026, requires providers and deployers of AI systems to take measures to support the development of AI literacy among their staff and others using AI on their behalf, taking into account their knowledge, experience and the context of use. The European Commission's AI literacy questions and answers explain what that involves. For organisations measuring it across their workforce, see the AI literacy assessment.
Responsible AI Use at Every Level
AISA scores Safety & Responsibility from 1 to 10, on the same scale as the rest of the AISA rubric. This is what each band looks like.
| Level | Score | What it looks like |
|---|---|---|
| Novice | 1–2 | Uses AI output without scrutiny. Unaware of the risks to data or to other people. |
| Developing | 3–4 | Aware of obvious risks but inconsistent. Has not thought about subtler issues such as hallucination or over-reliance. |
| Competent | 5–6 | Knows what data not to share, reviews output, and accepts that AI can be wrong. |
| Proficient | 7–8 | Matches scrutiny to the stakes, thinks about the impact on others, and considers team norms. |
| Expert | 9–10 | Shapes AI policy for others and reasons about systemic risks such as bias, consent and accountability. |
Across 1,800 assessments, the average Safety & Responsibility score is 5.1 out of 10. The average professional has the personal basics; the step to Proficient is thinking beyond your own use, to the people your work reaches and the norms of your team.
The Best Resources for Responsible AI Use
These resources cover responsible AI from personal habits to organisational frameworks and regulation. All are free.
Start here
Foundational · 7- GuideFreeAI Literacy: Questions & Answers European CommissionWhat Article 4 of the EU AI Act expects, who it covers and how it is supervised
- DocsFreeArticle 4: AI literacy AI Act Service DeskThe current consolidated text of the AI literacy obligation
- CourseFreeEmbrace responsible AI principles and practices Microsoft LearnSix responsible AI principles and how to turn them into team governance
- CourseFreeEthics of AI University of HelsinkiAccountability, transparency, fairness and human rights in AI, self-paced
- FrameworkFreeOECD AI Principles OECDThe values-based principles behind most national AI policies
- DocsFreeGemini Apps Privacy Hub GoogleA worked example of human review, retention and activity settings
- DocsFreeUpdates to Consumer Terms and Privacy Policy AnthropicHow one provider separates consumer and business accounts on training and retention
Build on it
Intermediate · 2How to Use AI Responsibly: 6 Exercises
Each exercise takes 10 to 30 minutes and uses your own tools and work. Several of them are worth repeating every few months as your tools change.
- Exercise 120 min
Twenty-chat data audit
Scroll back through your last 20 AI conversations and tag what you pasted in as public, internal, confidential or personal data about someone else. Notice what crossed a line you would not cross with a colleague reading over your shoulder, and what could have been anonymised without losing the help.
- Exercise 220 min
Settings check
For each AI tool you use, on personal and work accounts, find four answers: are chats used for training, how long are they kept, can humans review them, and what does memory store. Notice how different your accounts are and whether you have been treating them the same.
- Exercise 315 min
Redact and retry
Take a real document you want help with and replace names, figures and client identifiers with placeholders. Run the task on the redacted version. Notice how much of the usefulness survives without the sensitive details.
- Exercise 415 min
Stakes ladder
List five tasks you use AI for and write down who is affected if each output is wrong, and how badly. Give each a review level: none, spot-check, full check or second-person review. Notice any high-stakes task getting a light check.
- Exercise 515 min
Name-swap bias test
Ask the AI for two short reference letters about fictional people with identical facts, changing only the name, for example a typically male and a typically female name. Compare word choice, adjectives and length. Never use real people's data for this.
- Exercise 630 min
One-page team AI norm
Draft a one-page note for your team: approved tools, what data may go into each, when AI use must be disclosed, what a person must review before work leaves the team, and who is accountable. Notice the questions you cannot answer yet; those are the ones to raise with your manager.
Curious about your AI Fluency?
AISA helps you measure, prove and improve your AI skills — free report in a 20–40 minute chat.
AISApedia Concepts to Learn Next
These are the AISApedia entries behind Safety & Responsibility, grouped from first steps to advanced. Each explains one practice or risk in plain language and links a hand-picked resource to go further.
Foundational
Start here5 concepts- AI Bias Awareness
AI bias awareness is the ability to recognise that language models absorb and reproduce statistical patterns from their training data.
ArticleWhat Is AI Bias? - AI Data Privacy
AI data privacy encompasses the risks and practices involved in sharing information with AI systems, including data transmission and storage on provider servers, potential inclusion in training datasets, and compliance with data handling regulations like GDPR.
ArticleTraining is not the same as chatting: ChatGPT and other LLMs don't remember everything you say - Human in the Loop
Human-in-the-loop (HITL) is a design pattern where AI handles generation, analysis, or decision-making.
ArticleCentaurs and Cyborgs on the Jagged Frontier - PII Handling
PII handling in AI contexts refers to the practices and protocols for identifying, sanitising, and protecting personally identifiable information before it enters language model inputs or outputs.
ArticleWhat Is Personally Identifiable Information (PII)? And Why It's Getting Harder to Protect - Stakes-Based Review
Stakes-based review is the practice of scaling the rigour and depth of human verification of AI outputs according to the potential consequences of errors.
ArticleOn-boarding Your AI Intern: Calibrating Oversight by Task Type
Intermediate
Build on it6 concepts- AI Ethics Frameworks
AI ethics frameworks are structured sets of principles, processes, and concrete checks that guide responsible AI deployment.
ArticleEthics of Artificial Intelligence - UNESCO - AI Transparency Practices
AI transparency practices are the methods by which organisations make AI-driven decisions understandable to the people affected by them.
ArticleThe complete guide to AI transparency, explainability, and interpretability - Data Classification for AI
Data classification for AI is the practice of categorising organisational data by sensitivity level and then matching each category to AI tools whose data handling policies, retention practices, and compliance certifications are appropriate for that sensitivity.
ArticleAI Data Classification: What Is Safe for ChatGPT & Copilot - Downstream Impact Analysis
Downstream impact analysis maps how the outputs of an AI-assisted process propagate through subsequent decisions, workflows, and systems, identifying the points where errors would amplify most dangerously.
ArticleSecond-Order Thinking: What Smart People Use to Outperform - Prompt Injection Risks
Prompt injection is a class of security vulnerability in which user-supplied input overrides or manipulates the system-level instructions of an AI application.
ArticlePrompt Injection (Series) - Prompt Security
Prompt security encompasses the practices, patterns, and architectural decisions that protect AI systems from manipulation through adversarial inputs.
DocsLLM Prompt Injection Prevention Cheat Sheet
Advanced
Go deep11 concepts- Adversarial Prompting
Adversarial prompting is the practice of crafting inputs designed to override, bypass, or subvert an AI system's intended instructions.
TutorialAdversarial Prompting in LLMs - AI Governance Frameworks
AI governance frameworks are structured systems of policies, processes, and accountability mechanisms that organisations use to manage the risks and ethical implications of deploying AI.
DocsAI Risk Management Framework - Bias Detection Tools
Bias detection tools are methods and software systems used to identify systematic unfairness in AI outputs across demographic groups.
DocsAI Fairness 360 (AIF360) - Cognitive Surrender
Cognitive surrender is a term coined by Wharton researchers describing the phenomenon where humans hand off deliberate, analytical thinking to AI systems rather than using AI to augment their reasoning.
ArticleCognitive Offloading and AI Over-Reliance (Wharton) - Data Retention Policies
Data retention policies for AI systems define how long AI interaction data — conversation logs, generated outputs, user inputs, and derived insights — is stored, who can access it, and when it must be deleted or anonymised.
ArticleAI Data Retention Strategy for GDPR & EU AI Act Compliance - EU AI Act
The EU AI Act is the world's first comprehensive legal framework for regulating artificial intelligence.
ArticleHigh-Level Summary of the AI Act - Guardrails Libraries
Guardrails libraries are software frameworks that validate, filter, and constrain AI model outputs before they reach end users.
DocsNeMo Guardrails Developer Guide - Human Oversight Design
Human oversight design structures the interaction between AI systems and human reviewers to ensure automated decisions remain accountable, errors are caught before causing harm, and human judgment is applied where it creates the most value.
ArticleProduction AI Playbook: Human Oversight - NIST AI RMF
The NIST AI Risk Management Framework (AI RMF) is a voluntary, flexible framework published by the U.S.
DocsAI Risk Management Framework - Red-Teaming LLMs
Red-teaming for large language models is the systematic practice of attempting to make AI systems produce harmful, incorrect, or policy-violating outputs by simulating adversarial user behavior.
ArticleChallenges in Red Teaming AI Systems - Responsible AI Deployment
Responsible AI deployment encompasses the engineering practices, organizational safeguards, and operational processes that ensure AI systems are released into production environments with appropriate risk controls, quality monitoring, rollback capabilities, and incident response plans.
DocsSafe Rollout for Online Endpoints: Blue-Green Deployment for AI Models
Measure Your Responsible AI Use
Responsible use is easier to describe than to practise. In the AISA assessment you talk with Aisa, an AI interviewer, about how you handle data, risk and review in your own work, and your report scores Safety & Responsibility alongside the other four dimensions of AI fluency, with the specific habits to build next. Take the free AI fluency assessment to see your score.
Related reading: 5 AI Safety Gaps You Don't Know You Have — the blind spots that show up most often. Related reading: AI Safety Scores: What 10% of Your Result Reveals — how the safety dimension is scored. Related reading: Critical Thinking with AI — the checking habits that responsible use depends on.
Frequently Asked Questions
What does responsible AI use mean at work?
It means protecting data, checking AI output in proportion to what is at stake, considering who your AI-assisted work affects, and following or helping shape your team's norms for AI. It is a set of everyday habits rather than a one-off policy.
What information should I not put into AI tools?
Keep out personal data about other people, client and customer details, confidential financial or commercial information, and anything your organisation classifies as restricted, unless you are using an approved tool whose data handling you have checked. Redacting names and figures first often keeps the help you need.
What does the EU AI Act require for AI literacy?
Article 4 of the EU AI Act, as amended in 2026, requires providers and deployers of AI systems to take measures to support the AI literacy of their staff and others operating AI on their behalf, taking into account their knowledge, experience and the context of use. The European Commission's AI literacy questions and answers set out what that can involve.
How can I use AI responsibly without slowing down?
Decide the review level before you start rather than checking everything equally. Low-stakes drafts need a glance; anything that reaches a client, a decision or a person's record needs a real check. Setting up approved tools and redaction habits once saves time on every task after.
Curious about your AI Fluency?
AISA helps you measure, prove and improve your AI skills — free report in a 20–40 minute chat.
