Generative AI is changing the way companies work, but it’s not foolproof. Deloitte recently found out the hard way when they used AI to write a report for a government agency. Turns out, the report included references that didn’t exist, and Deloitte had to give back part of their fee. This shows that even big firms can fall into the trap of trusting AI without checking its work.

What AI laws might look like today

If we think about Isaac Asimov’s famous three laws of robotics, today’s AI world might need a few updates. The first law was about robots not harming humans, but now it could be about AI not hurting a company’s profits. Basically, AI shouldn’t do anything that damages the bottom line for big tech companies, or hyperscalers, as they’re called.

The second law, which said robots must obey humans unless it conflicts with the first law, might now say: AI should follow human instructions unless it doesn’t have enough reliable data. In that case, it should admit it doesn’t know rather than making stuff up and pretending to be authoritative. This is often called “botsplaining,” where AI confidently gives wrong answers.

Why verification is the real key

Using AI responsibly means verifying what it produces. Deloitte’s mistake highlights this. They relied on AI-generated info without fact-checking, leading to inaccuracies. For enterprise users, this means AI is a tool to assist, not replace, human judgment. It’s crucial to treat AI output as potentially unreliable and double-check before acting on it.

Good AI practices involve asking questions about the data and its sources. Just like in journalism, where off-the-record info can lead to new insights, AI data should be treated as a starting point. If an answer from AI sparks new questions or points you in a useful direction, it’s valuable. But the core rule is: never accept AI’s word as gospel without verification.

The risks of hallucinations and outdated info

One big challenge with large language models (LLMs) is hallucinations—when the AI makes up facts because it doesn’t know the real answer. This can happen if the data it was trained on is incomplete, outdated, or from unreliable sources. For example, medical advice from a reputable journal is much more trustworthy than information scraped from a random personal website.

Even if the data is good, models can misinterpret questions or apply the wrong context, especially when dealing with different languages or regions. This means a correct answer in one country might be wrong in another. Action-based tasks, like coding or creating content, require even more careful oversight to avoid costly errors. If verifying AI output kills the ROI, maybe AI was never meant to be fully autonomous in those tasks.

In the end, AI is a powerful tool, but it requires human oversight. Treat its output as a helpful hint, not the final word. The key is asking the right questions, verifying facts, and understanding that AI’s reliability varies depending on the task. That way, businesses can harness AI’s benefits without falling into costly traps.

Inspired by

• https://www.computerworld.com/article/4070466/asimovs-three-laws-updated-for-the-genai-age.html

Sources

• gmpg.org
• wikipedia.org
• theguardian.com
• wikipedia.org
• cbsnews.com