AI can be so convincing that it lulls us into a false sense of security. You read a beautifully written paragraph, and you think it must be accurate. Then right in the middle of it is something completely wrong. That is why we need to talk about AI hallucinations.
AI is being used every day to write real estate listings, legal documents, social posts, and news summaries. It saves time and can feel like magic. But when it gets things wrong the results can damage reputations, cause legal problems, and erode trust that took years to build.
In my latest webinar I walked through where AI goes wrong, the current hallucination rates across the major tools, and the research that is helping us keep AI outputs more reliable.
👉 Watch the full webinar here:
The Risks of Relying Too Much on AI
Here is what can happen when AI is not checked properly:
-
It presents false information with absolute confidence.
-
It invents citations that simply do not exist.
-
It exposes people and organisations to legal consequences.
-
It damages reputations in seconds.
I have seen examples everywhere. Lawyers using AI-generated cases that were made up. A real estate listing that described schools nearby when there were none. Even travellers missing flights because they asked ChatGPT for departure times instead of the airline website.
The Numbers Tell the Story
Across the mainstream tools, the data is confronting:
-
ChatGPT GPT-5 paid version is very accurate, but still around 1 in 100 answers contain errors.
-
ChatGPT free tier is riskier with about 1 in 20 responses wrong.
-
Claude 4 sits around 4 to 5 percent.
-
Microsoft Copilot is the most concerning with errors in roughly one in four outputs.
My own research found something even more worrying. In a survey of 400 strategic communicators across eight countries, 40 percent said fact-checking AI content was only “moderately important”.
Nearly half admitted they barely edited AI-generated outputs before publishing.
That is a major risk for anyone representing an organisation or client.
What the Research Shows
Studies from 2025 are giving us some clear ways forward:
-
Asking specific and clearer questions reduces mistakes.
-
Double-checking each step helps catch errors more effectively than relying on one final check.
-
Providing AI with structured, contextual documents improves accuracy by almost 30 percent.
-
Fact-checking agents can review outputs and flag false citations, although they still let half-truths through.
The evidence is clear. AI works best as an assistant, not as the final authority.
A Practical Toolkit for Quality Control
From my book AI for Strategic Communication and my own research I have developed a framework to keep AI outputs under control.
-
Preparation – Write an AI policy, train your team, and provide contextual documents.
-
Content Generation – Use prompts that highlight accuracy, ethics, and boundaries.
-
Post-Production – Check facts, sources, and citations the old-fashioned way.
-
Editing – Remove clichés, add human nuance, and align with brand voice.
Want to see how this works step by step?
👉 View the complete slide deck.
Why Humans Still Matter
AI is articulate. It is persuasive. But it is not always right. If you publish without fact-checking, you will be the one accountable, not the tool.
That is why every organisation needs policies, guardrails, and workflows. AI can speed up the process but humans must stay in the loop. You are the last line of defence and the first line of trust.
So before you hit publish on an AI draft, stop. Fact-check it. Edit it. Make sure it has your human sparkle.
The best way to use AI is to make it look like you did not use it at all.
Final takeaway:
AI can help you move faster, but accuracy is everything. Protect your reputation by staying vigilant.
How do you ensure AI generated content is correct before posting it?
👉 Join my free AI for Strategic Communication Community to access monthly webinars, the latest updates, and exclusive workshop discounts.