AI Strategy6 min read

AI Risk Terms: Hallucination, Workslop, Shadow AI and More

Part 3 of a plain-English AI terms series: hallucination, workslop, shadow AI, data poisoning, prompt injection, jailbreaking and LLMjacking, with a question for each.

Haojun See

Founder & Director, On The Ground

Updated 22 October 2026

Why AI risk terms matter to business leaders

Most AI risks don't look like a hacker in a hoodie. They look like a confident wrong answer, a polished report nobody checked, or a staff member pasting client data into a free tool. This is part 3 of our plain-English series on AI terms, covering the risks leaders should be able to name. The terms follow the vocabulary of the Singtel AI.dea – AI bridge programme, run with SIM Academy (Programme Architect: Michelle Yeo), whose cybersecurity sessions involve Singtel's Cyber Security Institute. The explanations are our own.

Hallucination: confident answers that are false

A hallucination is when generative AI states something untrue as if it were fact: a figure that was never published, a clause that isn't in the contract, a court case that doesn't exist. It happens because language models produce the most likely-sounding words rather than looking facts up. The danger is the tone. A wrong answer arrives in the same polished, confident style as a right one, so readers have no signal to slow down. Ask: for anything that leaves the building, who checks the facts and the sources before it goes?

Workslop: AI output that looks done but isn't

Workslop is AI-generated work that looks finished but lacks the substance to be useful, so the effort of fixing it lands on whoever receives it. The term was coined in 2025 by researchers writing in Harvard Business Review. Think of a slide deck with tidy headings and generic bullet points, or a summary that misses the one decision that mattered. The sender saved time; the receiver now has to work out what is wrong and redo it. Across a team, that cost can cancel out the time AI was meant to save. Ask: do we have a shared standard for AI-assisted work, such as checking facts and stating what AI drafted, before it is passed on?

Shadow AI: tools staff use without approval

Shadow AI is staff using AI tools the company hasn't approved, usually with good intentions. Someone uploads a customer list to a free tool to tidy it up, or pastes a contract in to get a summary. The work gets done, but company and personal data now sits with a provider you have no agreement with. For Singapore businesses, that can become a PDPA problem as well as a confidentiality one. Banning AI outright tends to drive the habit out of sight rather than stop it. Ask: which AI tools are our staff actually using today, and have we told them which ones are safe for which kinds of data?

Data poisoning: corrupting what the AI learns from

Data poisoning is feeding false or junk data into the material an AI learns from or analyses, so that it draws the wrong conclusions. The poison can be deliberate, such as a batch of fake product reviews planted by someone with a motive, or accidental, such as duplicate records and bot traffic nobody cleaned out. The result is AI output that looks reasonable but rests on a bad foundation, so decisions get made on patterns that aren't real. Ask: where does the data feeding this tool come from, and who checks it for junk before it is used?

Prompt injection and jailbreaking: tricking the AI

Prompt injection is hiding instructions in content an AI will read, so that it follows the attacker's instructions instead of yours. The instructions can sit in an email, a supplier document, a web page or a customer form, sometimes in white text or tiny fonts that people can't see. This is called indirect prompt injection, and it is most dangerous when an AI agent can act, such as sending emails, moving files or approving payments. Jailbreaking is the direct version: a user talks the AI out of its own safety rules through role-play, persistence or clever wording. The OWASP Top 10 for LLM applications ranks prompt injection first among risks to AI applications. Ask: what can this AI do with what it reads, and what stops it from acting on instructions hidden in a document?

LLMjacking: someone else using your AI account

LLMjacking is when attackers gain access to a company's AI service, usually through a leaked API key or an exposed server, and use it at the company's expense. The first sign is often a usage bill far above normal, and the stolen access can also be sold on or used to reach data connected to the AI. The usual causes are ordinary: keys pasted into code or shared documents, test servers left open to the internet, and no alerts on unusual usage. Ask: where are our AI keys stored, who can see them, and will we be alerted if usage suddenly jumps?

Next in the AI terms series

Part 4 turns to AI strategy terms: AI ambition, readiness, quick wins and RICE. Earlier parts cover core AI terms and the types of AI; part 5 covers architecture and governance. To learn these with your own team, you can register your interest on the AI.dea programme page.

Frequently asked questions

What is an AI hallucination?

An AI hallucination is when a generative AI tool states something false as if it were true, such as an invented statistic, policy or citation. It happens because the model predicts likely-sounding words rather than checking facts. Hallucinations can be reduced by grounding answers in trusted documents and having people check important outputs, but not eliminated.

What is shadow AI and why is it a risk?

Shadow AI is staff using AI tools their company hasn't approved. It is a risk because confidential or personal data can end up with a provider the company has no agreement with, which may breach confidentiality duties or the PDPA. A short AI use policy with approved tools and clear data rules usually works better than a ban.

What is the difference between prompt injection and jailbreaking?

Jailbreaking is a user directly persuading an AI to ignore its safety rules. Prompt injection hides instructions inside content the AI reads, such as an email, document or web page, so the AI follows the attacker without the user knowing. Prompt injection is the bigger business risk when AI agents can send messages or move data.

AI Strategy

Types of AI Explained: Predictive, Generative, Vision and More

AI Strategy

AI Terms for Leaders: Hyperscalers, Agents and Observability

AI Strategy

AI Strategy Terms: Ambition, Readiness, Quick Wins and RICE

PreviousAI Strategy Terms: Ambition, Readiness, Quick Wins and RICE Next AI Architecture and Governance Terms for Business Leaders