36 AI mistakes and safety risks at work in 6 groups, each with a case or example, one control and the course that teaches it.
Staff who use AI at work, and the managers, privacy officers and boards who set the rules, can use these 36 items as a checklist, starting with the group closest to their job.
Staff use personal AI accounts or public AI websites for work.
Control: Give staff an approved tool and a list of data that may go nowhere else.
Taught in: Shadow AI and Safe AI Use
Case: In December 2025 Eindhoven municipality made public that staff had uploaded youth care files and CVs to public AI websites (Omroep Brabant).
Reference: NIST AI 600-1, Data Privacy
Free and personal plans may keep your inputs and train future models on them.
Control: Use business plans whose contract excludes training on your inputs.
Taught in: AI Basics at Work
Case: In April 2023 Samsung engineers pasted internal source code into ChatGPT. Samsung banned generative AI tools on company devices from 1 May 2023 (TechCrunch).
Reference: OWASP LLM02:2025 Sensitive Information Disclosure
AI notetakers record everyone on a call and send the data to the vendor.
Control: Announce the notetaker, get everyone’s consent and know where recordings go.
Taught in: Shadow AI and Safe AI Use
Case: A proposed class action filed in August 2025 (NPR) alleges Otter.ai recorded calls without every participant’s consent; core claims proceed (Baird Holm).
Reference: NIST AI 600-1, Data Privacy
Unpublished price-sensitive facts typed into an outside AI tool may count as unlawful disclosure under the EU Market Abuse Regulation (MAR). This is legally untested.
Control: Keep inside information out of every AI tool not approved for it.
Example (constructed): An investor relations manager pastes draft quarterly results into a free chatbot 2 days before publication.
A share link can expose a whole conversation, sometimes to search engines.
Control: Never share a chat link that holds client, colleague or company data.
Taught in: AI Basics at Work
Case: In July 2025 Fast Company found nearly 4,500 shared ChatGPT chats in Google results. OpenAI removed the “discoverable” option on 1 August 2025 (Fast Company, TechRadar).
Your data is only as safe as the AI vendor’s own systems.
Control: Run every AI app through supplier security review before use.
Taught in: Shadow AI and Safe AI Use
Case: In January 2025 Wiz Research found a DeepSeek database open on the internet without authentication, holding chat histories and API keys (Wiz Research).
Reference: NIST AI 600-1, Value Chain and Component Integration
The model states false quotes, figures or references with full confidence.
Control: Check every quote, figure and citation against the original.
Taught in: AI Basics at Work
Case: In October 2025 Deloitte partly refunded an AUD 440,000 Australian government review that contained a fabricated court quote and non-existent references (Associated Press).
Reference: NIST AI 600-1, Confabulation
Unchecked, generic or wrong AI output shipped as finished work.
Control: Name a human reviewer for every AI-assisted deliverable, including bought-in content.
Taught in: Managing Teams That Work With AI
The model agrees with you because you seem to want agreement.
Control: Ask for the strongest case against your view.
Taught in: AI Basics at Work
Case: In April 2025 OpenAI rolled back a GPT-4o update that made ChatGPT overly agreeable, including validating doubts and urging impulsive actions (OpenAI).
Reference: NIST AI 600-1, Human-AI Configuration
Models miss events after their training cutoff unless they search or you supply the facts.
Control: For laws, prices and policies, give the model the current source.
Taught in: AI Literacy and the AI Act
Example (constructed): In October 2026 an HR adviser asks a chatbot what AI Act Article 4 requires and gets the “sufficient level” wording that Regulation (EU) 2026/1744 replaced in July 2026.
A model can do a hard task well and fail an easier one beside it.
Control: Test each new type of task on cases where you already know the answer.
Taught in: AI for Business Analytics
Example (constructed): An assistant summarises a 40-page supplier contract well, then totals its price table wrong.
A prompt that reveals the answer you hope for pushes the model towards it.
Control: Ask neutrally and keep your own view out of the prompt.
Taught in: AI Basics at Work
Case: An October 2023 study of 5 AI assistants found they “consistently exhibit sycophancy”, and people sometimes preferred convincing agreeable answers to correct ones (Sharma et al.).
Without audience, purpose and constraints, the model guesses and builds on its guesses.
Control: Put all context in the first prompt, and restart a chat that went wrong.
Case: A May 2025 study of 200,000+ simulated conversations found top models performed 39% worse on average when requirements came step by step (Laban et al.).
Without a stated format and length, the model picks its own, often long and generic.
Control: State format, length and reader in every prompt.
Taught in: AI Basics at Work
Example (constructed): A manager types “summarise this report” and gets 900 words. The board needed 5 sentences and a table of 3 risks.
Asking for figures without context or sources brings unverifiable assumptions into your work.
Control: Ask for the source of every figure, then open and check it.
Taught in: AI for Business Analytics
Case: In a 7 June 2024 damages ruling, a Dutch court used ChatGPT to estimate solar panel lifespan and electricity prices. Commentators raised reliability concerns, though no error has been shown (CMS).
Several jobs in one prompt let the model shorten or skip some, and make checking hard.
Control: Give one task per prompt and check each result.
Example (constructed): One prompt asks for a market summary, 3 competitor profiles, a pricing table and a CEO email. The pricing table comes back with invented figures.
Asking the same model to confirm its answer adds no independent check.
Control: Check against an independent source or a person.
Taught in: AI Basics at Work
Case: In Mata v. Avianca a lawyer asked ChatGPT “Is varghese a real case”, and it said yes. Sanctions over 6 fake cases followed on 22 June 2023 (SDNY sanctions order).
Many tools start each new chat from zero, so earlier instructions are gone.
Control: Paste a saved instruction block at the start of each chat, or use project instructions.
Taught in: AI Basics at Work
Example (constructed): A consultant sets a client’s style rules on Monday, opens a new chat on Wednesday and sends a draft that breaks them.
A user types instructions that override an AI system’s rules.
Control: Cap what customer-facing AI may promise and send binding matters to a person.
Taught in: Shadow AI and Safe AI Use
Case: In December 2023 a user told a Chevrolet dealer’s chatbot to agree with everything, and it called a USD 1 offer for a Tahoe “legally binding” (GM Authority).
Reference: OWASP LLM01:2025 Prompt Injection
Hidden text in emails, files or web pages can steer an AI assistant that reads them.
Control: Keep sensitive data out of reach of assistants that read outside content.
Taught in: Shadow AI and Safe AI Use
Agents with broad access can act on bad instructions.
Control: Give agents minimum access and require human approval to send, pay or delete.
“AI helper” extensions can read every page you open.
Control: Install only extensions your IT team approves.
Taught in: Shadow AI and Safe AI Use
Case: In December 2025 OX Security found 2 copycat Chrome AI sidebar extensions, one with 600,000+ users and a “Featured” badge, sending users’ ChatGPT and DeepSeek chats to an attacker (OX Security).
Fraudsters fake executives’ faces and voices to order payments.
Control: Confirm urgent payment requests on a known separate channel, even after a video call.
Taught in: Shadow AI and Safe AI Use
Case: In January 2024 an Arup employee in Hong Kong made 15 transfers totalling HK$200 million after a video call with a deepfaked CFO and colleagues (CNN).
Reference: NIST AI 600-1, Information Integrity
Organisations that build or use AI must take measures to support staff AI literacy. Since 27 July 2026 (Regulation (EU) 2026/1744) no guaranteed level per person is required, and the AI Act does not require a certificate.
Control: Train by role and keep records of who learned what.
Taught in: AI Literacy and the AI Act
Example (constructed): A company emails staff a video link and keeps no record.
Since 2 February 2025 the AI Act bans 8 practices, including inferring people’s emotions at work, except for medical or safety reasons.
Control: Check HR and monitoring tools against the Article 5 list before purchase.
Taught in: AI Literacy and the AI Act
Example (constructed): A contact centre scores agents’ frustration from their voices and uses the scores in reviews.
Since 2 August 2026, chatbots must disclose they are AI and deepfakes need labels (Al Jazeera). Fines reach EUR 15 million or 3% of turnover; no AI Act fine had been published by 15 October 2026, as far as we found.
Control: Label chatbots and deepfakes before anyone sees them.
Taught in: AI for Video Production
Example (constructed): A webshop chatbot poses as a named human agent.
Personal data in an unapproved chatbot can be a breach, reportable within 72 hours. Health and other special category data need an exception under GDPR Article 9.
Control: Report suspected leaks to your privacy officer at once.
Taught in: Shadow AI and Safe AI Use
Case: On 7 August 2024 the Dutch privacy regulator cited a GP practice employee who entered patient data into a chatbot (DutchNews.nl, AP).
GDPR Article 22 limits solely automated decisions with significant effects, such as AI-only CV rejections.
Control: Keep a trained person making every hiring, pay or dismissal decision.
Taught in: Managing Teams That Work With AI
Case: In October 2018 Amazon was reported to have dropped a CV-scoring tool that penalised CVs containing “women’s” (Reuters via Euronews).
Reference: NIST AI 600-1, Harmful Bias or Homogenization
AI-only output may have no copyright owner and can copy protected work.
Control: Document human creative input in assets you need to own.
Taught in: AI for Video Production
Case: A US appeals court held in March 2025 that copyright registration needs a human author, and the Supreme Court declined review on 2 March 2026 (Thaler v. Perlmutter, SCOTUSblog, Skadden).
Reference: NIST AI 600-1, Intellectual Property
Consumer plans often come without a GDPR Article 28 processor contract.
Control: Use only tools with a processor agreement for personal data.
Case: Italy’s privacy regulator fined OpenAI EUR 15 million in December 2024 (Euronews), partly over training ChatGPT on personal data without an adequate legal basis. A Rome court annulled it on jurisdiction grounds on 18 March 2026 (Reuters via Freevacy).
MAR restricts inside information. Under DORA, contracted AI tools at financial entities belong in the register of information. NIS2 requires supply chain security and board training.
Control: Put every AI tool through your existing third-party checks.
Taught in: AI Strategy and Change for Decision-Makers
Example (constructed): An insurer’s claims team buys an AI summary tool on a company card, and it never reaches the DORA register.
Blocking tools staff rely on, without an approved alternative, pushes work to workarounds.
Control: Pair any block with an approved tool from day one.
Case: After Italy blocked ChatGPT in March 2023, Italian developers’ output fell about 50% for 2 business days and Tor use rose, an April 2023 working paper found (Kreitmeir and Raschky).
Nobody is accountable for AI policy, tools or incidents, so decisions stall.
Control: Name one accountable owner, with a deputy, for AI policy, tools and incidents.
Taught in: AI Playbook Workshop
Example (constructed): IT approves a chat tool, legal writes a policy, and 6 months later nobody can say who approves a new AI notetaker.
Tools and rules keep changing after a single launch session.
Control: Run short refreshers yearly and after each major tool or rule change.
Taught in: Managing Teams That Work With AI
Example (constructed): A firm trained staff in March 2025. In October 2026 its slides still show the old Article 4 wording.
Without a baseline and quality measures, cost savings can hide falling quality.
Control: Measure quality and customer outcomes before and after, with cost.
Taught in: Managing Teams That Work With AI
Case: In February 2024 Klarna said its AI assistant did “the work of 700 customer service agents”. In May 2025 its CEO said the push went too far and quality dropped (Entrepreneur).
Roles are cut on AI claims nobody checked against real workload data.
Control: Test AI claims on your own workload data before any job decision.
Taught in: AI Strategy and Change for Decision-Makers
Case: Commonwealth Bank of Australia planned to replace 45 customer service roles with an AI voice bot, then reversed and apologised on 21 August 2025 after the union said calls were rising (Information Age).