Gen AI Innovation
Many of us are using large language models like ChatGPT, Claude, and Copilot to boost productivity. Far fewer are aware of all the risks sitting inside these tools. That gap is where the real damage happens — not because the technology is unsafe, but because it's used without knowing what to watch for.
Here are the six key risk areas of LLMs, and a mitigation prompt for each one, so you can handle them like a pro.
1. Hallucinations
The risk: the AI confidently states a "fact" that is completely fabricated.
Example: you ask for legal precedents, and the AI invents a court case that never happened — with a case name, a date, and a confident tone that gives you no reason to doubt it.
Mitigation prompt: "Check your output for factual accuracy. If you are unsure of a specific detail or date, state that you don't know rather than guessing. Provide source links where possible."
2. Bias and Ethical Issues
The risk: AI reflects the societal biases present in whatever it was trained on.
Example: a recruitment prompt that consistently suggests male candidates for leadership roles, without you ever asking for that skew.
Mitigation prompt: "Review this response for any gender, racial, or cultural biases. Ensure the tone is neutral and inclusive, and provide diverse perspectives on this topic."
3. Confidentiality and Data Security
The risk: inputting sensitive data into an LLM can expose that data, depending on the tool and settings involved.
Example: pasting a proprietary company strategy document in just to get a quick summary.
This one is less about a clever prompt and more about a behavioural fix: never paste personally identifiable information or proprietary data directly. Instead, use something like: "I will provide a sanitised version of a business problem. Please analyse the logic without requiring specific names or proprietary data."
4. Sycophancy — the "Yes-Man" Effect
The risk: the AI tends to agree with whatever opinion you've stated, even when that opinion is wrong.
Example: ask "why is [a flawed idea] actually genius?" and most models will find a way to justify it rather than push back.
Mitigation prompt: "Adopt the role of a 'Critical Challenger.' Don't just agree with my premise — identify the flaws in my logic and provide a counter-argument."
5. Copyright and Compliance
The risk: generating content that unintentionally mirrors copyrighted material.
Example: asking for a poem "in the style of" a living author, and getting back something that leans on their specific, protected phrasing rather than just their general style.
Mitigation prompt: "Ensure this output is original and does not quote verbatim from existing copyrighted works. Focus on the core concepts rather than mimicking a specific person's unique style."
6. Spotting AI-Generated Content and Deepfakes
The risk: mistaking AI-generated images, audio, or video for reality.
Example: a viral "leaked" audio clip of a CEO saying something controversial — that was never actually said.
This risk doesn't have a prompt-based fix, because you're not the one prompting it — someone else is, and you're on the receiving end. The mitigation is a habit: don't just trust what you're told, look at what you're shown. Watch for the "uncanny valley" signs — inconsistent lighting, extra fingers, distorted background textures, audio that doesn't quite match mouth movement.
The Pattern Across All Six
Notice that none of these six risks are solved by avoiding AI altogether. Each one has a specific, learnable countermeasure — a prompt to add, a habit to build, a question to ask before you trust the output. That's the whole point of AI literacy: not fear of the tool, and not blind trust in it either. Knowing precisely where the risk sits, and having the right move ready when it shows up.