How to Reduce the Risk of Generative AI Mistakes

The biggest risk of GenAI isn’t hallucination - it’s overconfidence.

 Because GenAI outputs sound fluent and authoritative, we’re more likely to accept them at face value. This is called automation bias.

To counteract it, you have to stay alert and curious.

For instance, when I asked Claude about its document handling limits, it confidently claimed it could process 150,000 to 200,000 words. That struck me as suspicious - it's the length of a full book. So I pushed back. Only then did Claude admit that its real sweet spot is closer to 15,000 to 20,000 words - a 10 x difference.

This wasn’t just a factual correction. It was a reminder: AI doesn’t lie, it guesses—and it often guesses wrong with confidence.

To stay sharp, build your critical thinking muscle. As Dwyer et al. (2023) emphasize, deliberate critical thinking reduces the influence of mental shortcuts and common reasoning errors by:

• Reflect: Pause and ask, What assumptions am I making?
• Analyze: Study past AI mistakes to spot failure patterns
• Debate: Engage in discussions that challenge your thinking
• Solve: Practice with puzzles or cases that require sound reasoning

One simple way to reduce risk is to have multiple GenAI tools cross-check each other’s outputs.

Try this method:

1. Draft your content with Tool A (e.g., Claude)
2. Feed that draft into Tool B (e.g., ChatGPT or Gemini)
3. Ask: “What factual errors or assumptions do you notice?”

Different models have different blind spots. Where one may hallucinate, another may flag a red flag.

Research backs this up: According to Dhuliawala et al., 2023, “ensemble prompting” - using multiple LLMs - can significantly improve factual accuracy

Traditional quality checks - like “dotting the i’s and crossing the t’s” - aren’t enough when your draft might contain hallucinated facts or fabricated citations. The GenAI era demands a disciplined, structured check, especially when the output looks polished.

Here’s a workflow we’ve found to reduce risk and save time:

1. Start with your own outline or intent. Use GenAI to speed up, not substitute, your thinking.
2. Let the AI iterate and enhance. Ask it to rephrase, clarify, and offer alternatives - but final decisions stay with you.
3. Finish with a focused review. Don’t just skim. Read with purpose. Scrutinize the facts, names, dates, and claims.

That last step matters because even the most mundane details can be wrong. For example, when preparing emails for our GenAI course, one draft listed June 1 as a Saturday. It wasn’t. A simple error - one I didn’t catch until a second pass. Exactly the kind of mistake that's easy to miss when you're moving fast.

Yes, this kind of review adds a few extra minutes - but it saves you far more time (and credibility) in the long run.

The good news? It gets easier. The more you work with GenAI, the better you become at spotting these subtle flaws. And in my experience, I’m still far faster - and far sharper - co-creating with GenAI than working without it.

Search-augmented GenAI tools (like Bing Copilot or Perplexity or ChatGPT using the Deep Search mode) offer powerful insights when used strategically.

Use deep search when:

• You need a detailed background on a technical topic
• You’re writing for high-stakes audiences (clients, judges, executives)
• You’re unsure if your sources are complete or up to date

Skip it when:

• You need speed, not nuance
• The topic is familiar or routine
• You’re drafting a low-risk communication

Innovation tip: In early-stage discovery, more data isn't better - relevant data is. Don’t use a tank when a car gets you there faster, cheaper, and more nimbly.

Innovators are taught to challenge assumptions, and that mindset is critical for anyone using GenAI.

Don’t just read the output. Interrogate it:

• Where is this claim coming from?
• What’s being assumed?
• Does this match what I already know?

Build small verification checks into your routine:

• For citations: Did I confirm each source exists and is used accurately?
• For summaries: Did I skim the original to make sure nothing essential got lost or distorted?
• For timelines: Are the dates and sequences realistic?

Create personal “if-then” triggers:

• If I’m working on something unfamiliar, then I’ll do a second-level check
• If this is for a high-stakes audience, then I’ll verify every claim

Behavioral science backs this up: implementation intentions (e.g., “If X, then Y”) are proven to increase follow-through and reduce error rates, not just for GenAI (Gollwitzer, 1999).

These speed bumps take a few extra minutes, but help avoid bigger risks.

Microsoft trained over 23 million people in AI skills last year. But training hasn’t always translated into safe, effective use.

That’s because most programs focus on:

• Prompt engineering
• Tool feature walkthroughs
• General warnings about hallucinations or biases

What’s missing:

• Real-world error detection
• How to cross-check AI outputs
• Domain-specific red flags
• Structured reflection and feedback loops

Until we teach verification literacy, many GenAI mishaps will keep occurring instead of being rare.

GenAI can save time, amplify ideas, and improve outcomes - but only with sound human judgment behind it.

The professionals who thrive won’t be the ones who prompt the fastest.
They’ll be the ones who:

• Know when to trust
• Know when to verify
• Build habits that make smarter use of these tools

If you’ve led a team, challenged the status quo, or launched a new idea - you already have what it takes.

Treat GenAI like a talented but error-prone colleague. Collaborate with it. Question it. Stay sharp.

Mistakes will still happen. But with the right systems in place, they’ll happen a lot less often - and probably not on your watch.