When AI Gets It Wrong: Understanding Hallucinations and Their Impact
AI hallucinations aren't random, they're a predictable consequence of how today's language models work.
When Generative AI hit the marketing scene, many of us were excited about how we could use it to churn out blogs and marketing copy in minutes. What a time-saver! Just tell it your topic, provide a few details, ask for some stats, and let it do its thing.
Sadly, it wasn’t that simple. AI hallucinations started popping up in the copy, making our readers go “Huh?” and our legal teams go crazy. That’s because we trusted AI to know what it was talking about at all times—and it doesn’t.
In fact, sometimes it literally makes things up or combines facts to generate something completely inaccurate, misleading readers and putting organizations at risk of reputational damage or worse. AI hallucinations are most dangerous when they sound plausible, and they can have your audience believing in something that simply isn’t true.
In this blog, we explain what AI hallucinations are, why they occur, and what you can do to fix them.
What Is an AI Hallucination?
An AI hallucination is a plausible-sounding statement that isn't accurate. It happens when an AI model confidently generates erroneous or false content. Here are some common examples:
Inventing a statistic
Making up a quotation
Citing a study that doesn't exist
Giving someone the wrong job title
Creating a nonexistent product
Getting a date or historical fact wrong
The important word is “plausible.” The output looks convincing, but it's erroneous or completely fabricated. Here's an example of a convincing AI hallucination:
"According to a 2025 Gartner survey, 78% of enterprises reported achieving a positive ROI from generative AI within six months of deployment."
At first glance, this sounds credible. It cites a well-known analyst firm, provides a recent year, includes a specific percentage, and makes a plausible claim. However, the statistic is entirely made-up; Gartner never published that survey or that finding.
Hallucination rates vary significantly depending on the task and the information available to the model:
When AI is restricted to summarizing or extracting information from source documents, leading models hallucinate less than 1% to 2% of the time.
Across general enterprise use, hallucination rates can be from 8% to 20% – so about one in five responses can contain fabricated or inaccurate information.
In high-stakes fields such as legal and medical research, general-purpose models can hallucinate on 58% to 88% of complex queries, particularly when they lack sufficient context and have to guess.
Hallucination rates increase as tasks become more complex and require AI to generate information rather than rely on verified sources.
How Do Hallucinations Happen?
Large language models are fundamentally trained to predict what comes next in a sequence of language rather than verify facts against a trusted source. When information is missing or ambiguous, the model may generate an answer that fits learned language patterns rather than acknowledge it might not know the answer.
Going back to the Gartner example, the AI model probably generated its response based on what it had already learned:
Gartner frequently publishes enterprise technology research
ROI studies often report results as percentages
Six-month ROI is a common timeframe discussed in AI adoption
Numbers like 78% sound specific enough to be believable without appearing exaggerated
It wasn’t lying intentionally or trying to mislead the reader – it was doing its best to predict the sequence of words that most likely fit the prompt, based on patterns in its training data. It combined real concepts (Gartner, enterprise AI, ROI surveys) into a citation or statistic that seemed reasonable but never existed.
That Can’t Be Right!
Obscure facts are particularly vulnerable. AI tends to have an easier time with common, well-established information than with arbitrary or rarely documented facts.
Recent research from OpenAI suggests that language models are much better at learning recurring patterns than memorizing isolated facts. When a prompt calls for a specific statistic or citation that doesn't exist in the model's learned knowledge, it may generate a plausible-looking answer based on those patterns instead of returning an "I don't know."
Can’t Find an Answer
Another challenge is a lack of access to specific information. AI models cannot directly access content behind paywalls, subscription services, or login-protected websites such as LinkedIn and Facebook. They also have limited access to newly published content and information that isn’t digitized, such as local and regional newspapers and many industry journals. Although AI may pick up quoted passages from free sources, they often don’t have the full context.
When the specific information you request isn't available, the model may attempt to fill in the gaps using patterns learned from similar content. The result can be a response that sounds coherent and convincing but is false.
Fact, Fiction, Or Both?
An AI response doesn't have to be 100% wrong to be dangerous. It might correctly identify a company, executive or product, but in the same sentence, make an egregious error—for example, getting the executive’s title wrong.
Similarly, AI may identify a study that exists but incorrectly summarize its findings. It can also provide a real URL but claim the page says something it doesn’t. That mixture of truth and falsehood makes hallucinations particularly difficult to detect.
How to Spot Hallucinations: 7 Things to Watch For
AI is incredibly useful, precisely because it can produce useful information quickly. But its ability to generate fluent, convincing language can also make mistakes difficult to spot.
Here are seven tell-tale warning signs to watch for:
Suspiciously Precise Statistics: When a number is oddly specific (i.e. 94.6% or 1.7 million users), it may seem authoritative; however, AI frequently invents percentages and metrics that sound believable.
The Fix: Always verify statistics against the original report or dataset.
Citations that Seem Perfect: If you’re looking for the perfect stat to support your claim, beware! AI can fabricate studies, authors, journal articles, report titles, or URLs that look completely legitimate.
The fix: Before relying on a citation, confirm that the publication, author, and quoted findings actually exist.
Obscure Facts Delivered with Certainty: Be skeptical when AI confidently answers niche historical, scientific, or technical questions that few people would know from memory.
The Fix: Treat AI’s answer as a clue, then dig deeper to verify those obscure facts.
Unnecessary, Specific Details: AI sometimes offers exact dates, times, model numbers, or dollar amounts that aren't essential to answering the question. It may be trying to justify its own reasoning.
The Fix: Rather than simply removing them from your draft, investigate. One inaccurate piece of information can signify that the entire response is a hallucination.
Contradictions: If answers change during the conversation or an explanation shifts, the model may be reasoning from probabilities instead of facts.
The Fix: Ask the same question in a different way or request supporting evidence.
Sources that Don't Actually Support the Claim: Even when AI links to a real article or report, the cited source may not contain the statistic, quote, or conclusion cited.
The Fix: Read the original source and check for accuracy.
Never admitting defeat: AI will continue to spin their output until you give up. They are not designed to "end" the conversation; they are designed to keep you engaged and will constantly ask one more question or offer one more piece of information.
The Fix: Check the AI’s responses from the start after each response to reduce the risk of going down a rabbit hole of misinformation.
While AI can help you brainstorm, consolidate ideas, or even produce a first draft, it’s absolutely essential that you check for hallucinations and false claims before you declare any AI-generated copy to be production-ready.
And, higher-stakes content deserves even higher scrutiny. If you’re producing content for healthcare, legal advice, finance, cybersecurity, engineering, or business strategy, the less you should rely on AI without independent verification.
Keep Humans in the Loop
AI is an exceptionally powerful tool, but it is not an authoritative source of truth. Organizations that get the most value from AI treat it as a collaborator rather than a replacement for human judgment.
For marketers, this means AI can draft a blog, summarize research, generate ideas, and help analyze information – but facts, statistics, quotes, customer claims, competitive information and anything consequential still need a real, live human for verification. Fact-checking is non-negotiable when it comes to high-stakes deliverables.
By combining AI's speed with human expertise, critical thinking, and fact-checking, you can leverage Generative AI to improve productivity while avoiding hallucinations, and maintaining the accuracy and trust that your audience expects.
Contact ETMG today to learn how we can help you maximize efficiency with Generative AI while maintaining accuracy and creating content that is original, engaging, and distinctly yours.
