Governance

AI Guardrails: What's important and what could go wrong?

Written on July 30, 2025 · by Tony Lan

Artificial Intelligence (AI) will transform how we work, shop, and solve problems in our daily lives. However, there could be serious mistakes if not properly identified early on with appropriate mitigations in place to eliminate serious mistakes.

Without appropriate AI guardrails, AI systems can inflict real harm on humans — by giving wrong answers or results, incorrect decisions that may sound too good but make absolute non-sense, or even getting stuck in loops (or infinite loops to drive you mad).

Below are common scenarios reflecting practical real-world examples:

1. Hallucinations: Making up non-sense

AI models can sometimes generate entirely false responses, results, or information. This is categorised as a hallucination. It's not a bug or glitch in the software, but a known limitation in the systems where words and patterns are not clearly defined in the "brain" of the AI model.

Example — Customer refund: A customer shops at Amazon and asks an AI customer service bot if they can receive a refund for a recent purchase. The bot says: "Yes, we can definitely perform a refund on all purchases in 100 days." The actual refund policy states 30 days for electronic devices and 100 days for other devices with proof of receipt. Conflicting or misinterpreted item types can create these hallucinated responses.

How to catch hallucinations:

2. Loops: Going round in circles

There are situations where AI can fall into a state of endless loops — especially in chats/conversations or automated tasks. This can happen for a number of reasons: unclear or undefined requirements, rules, or conditions, or simply poor prompt design rendering the conversation into an infinite loop.

Example: A customer support AI bot repeats "One moment, please wait while I check the details for you…" each time a customer restates their issue. This becomes repetitive and useless, frustrating the customer without a clear path forward.

How to prevent continuous loops:

3. Bad decisions: No logic and unethical

If an AI agent is asked to make a decision based on logic and reasoning but provides one that isn't logical or is unethical, that poses real safety concerns — often due to poor training data or gaps in human intervention and handoff systems.

Example: A bank fraud detection system rejects a refund request because the user's phone behaviour was deemed inappropriate or high-risk — when in fact the refund request was entirely valid. There was no human intervention to review the edge case.

Checklist to avoid bad decisions:

4. Over-generalisation: One size fits all

AI agents can over-generalise — applying general rules to specific cases where they can't appropriately reason or respond. This leads to decisions or messages that appear reasonable but are actually off-topic or don't make sense for the situation.

Example: A healthcare AI agent responds to a user reporting chest pain by suggesting more water and rest. The general response may sound plausible, but underlying symptoms may never be correctly identified — which is genuinely dangerous.

Checklist to reduce over-generalisation:

Conclusion

There are many scenarios in which AI needs guardrails — technical, ethical, logical, and procedural. Left unchecked without human intervention, these systems can make repetitive mistakes that cause reputational or personal damage.

For organisations, this means detailed checklists for developers, real-time monitoring tools, and human oversight for sensitive or high-risk cases. On a personal level, stay mindful of AI-driven responses, think critically, and validate outputs against authentic primary or secondary sources.