If you've been paying attention throughout the programme, and especially if you've been doing the activities carefully, you've almost certainly already encountered the problem we're tackling here. Maybe a chatbot confidently told you something that turned out to be wrong. Maybe a research report cited a source that didn't quite say what the AI claimed. Maybe you asked for information about something you know well and noticed a detail that was subtly, plausibly off.
That's an AI hallucination. And it's one of the most important things anyone using AI needs to understand.
The term "hallucination" describes what happens when an AI generates content that sounds confident and plausible but is partially or entirely fabricated. It might invent a statistic, cite an academic paper that doesn't exist, attribute a quote to someone who never said it, or state a "fact" that is simply wrong, all while using the same calm, authoritative tone it uses when it's being accurate.
This is not a bug that's about to be fixed. It's a fundamental feature of how current AI language models work, and while the problem has improved significantly (hallucination rates on straightforward questions have dropped below 2% for the best models) it gets dramatically worse for complex, niche, or open-ended questions, where error rates can exceed 30%. Understanding why this happens, learning to spot it, and developing habits for verification are among the most valuable skills you'll take away from this programme.
Why AI hallucinations happen
To understand hallucinations, it helps to go back to something we touched on in the very first Thing: AI language models don't actually "know" things the way people do. They generate text by predicting what word is most likely to come next, based on patterns learned from vast amounts of training data. When you ask a question, the model isn't looking up the answer in a database. It's constructing a response that sounds like the kind of answer that would typically follow your question.
Most of the time, this works remarkably well. The patterns in the training data are rich enough that the model's predictions align with reality. But sometimes, and especially in situations where the model's training data was sparse, ambiguous, or contradictory, the prediction engine produces something that sounds right but isn't. The model doesn't know it's wrong. It has no mechanism for distinguishing between "I'm confident because the evidence is strong" and "I'm confident because this sentence sounds plausible." It just generates the next most likely word, and then the next, and then the next.
This is why hallucinations are so dangerous: they don't come with warning labels. A fabricated statistic is delivered with exactly the same tone and formatting as a genuine one. A non-existent academic paper is cited with a perfectly plausible-sounding title, author name, and journal. The AI doesn't hedge or flag uncertainty (unless you've specifically asked it to) because hedging would make the response less like the confident, authoritative text it was trained on.
There are several common forms that hallucinations take, and recognising the patterns will help you spot them.
When hallucinations are most likely
Hallucinations aren't evenly distributed. They're much more common in some situations than others, and knowing when to be especially vigilant is half the battle.
Niche or specialist topics are higher risk. AI models have more training data about common subjects (London is better covered than Lochgilphead) so questions about less well-documented topics are more likely to produce fabricated details. If you're asking about something obscure, specialised, or local, your guard should be higher.
Specific factual claims deserve extra scrutiny. Names, dates, statistics, quotations, and references are the categories most prone to hallucination. Whenever an AI gives you a specific, verifiable fact, treat it as a claim to be checked rather than an established truth.
Questions where the model should say "I don't know" are particularly risky. Current AI models are strongly biased towards providing an answer. They will almost never say "I don't have reliable information about this." Instead, they'll generate the most plausible-sounding response they can construct, even if that means making something up. This is improving (some models are now better at expressing uncertainty when prompted to), but the default behaviour is still to answer confidently.
Older or rapidly changing information is unreliable. Models have training data cutoffs, and their knowledge of recent events or frequently changing facts (current office holders, living people's ages, recent statistics) may be outdated or fabricated. AI search tools like Perplexity (Thing 4) reduce this risk by searching the web in real time, but standard chatbots are working from a fixed snapshot of the world.
How to protect yourself
The good news is that hallucinations are manageable if you develop the right habits. You don't need to distrust everything AI produces, but you do need to verify anything that matters. Here are practical strategies that work.
Verify specific claims. Any time an AI gives you a specific fact (a name, a date, a number, a reference) and you plan to use that information for anything important, check it against an independent source. This is the single most important habit in AI literacy. It takes seconds for most claims and saves you from passing along fabrications.
Ask for sources, then check them. When using a chatbot, you can ask it to provide sources for its claims. This is useful, but with a critical caveat: the sources themselves may be hallucinated. Always click through to verify that the source exists and actually says what the AI claims it does. AI search tools like Perplexity, which provide inline citations from real web pages, are more reliable here than standard chatbots, but even their summaries can occasionally misrepresent what a source actually says.
Use your own expertise. One of the most powerful hallucination detectors is your own knowledge. If you're asking about a topic you know well, you'll catch errors that someone less familiar with the subject would miss entirely. This is why Thing 15 sits here in the programme rather than at the beginning: your fourteen Things of hands-on experience have given you a much richer sense of what AI gets right and wrong than you had when you started.
Cross-reference between tools. The comparison approach you've been using throughout this programme (trying the same prompt in multiple tools) is also an effective hallucination check. If two independent AI systems give you the same specific fact, it's more likely to be real (though not guaranteed, since they may share training data). If they give you different facts, at least one is wrong, and you know to investigate further.
Watch for the confidence trap. Be especially sceptical of information that is presented with high confidence and specificity but that you have no way to immediately verify. The more precise and authoritative a claim sounds, the more important it is to check, precisely because the authoritative tone is what makes hallucinations so effective at slipping past your defences.
Prompt for honesty. You can reduce (though not eliminate) hallucinations through how you prompt. Asking a model to "only include information you're confident about" or to "flag any claims you're uncertain about" can help. Some models respond well to instructions like "If you don't know, say so rather than guessing." This doesn't make them perfectly reliable, but it shifts the balance.
AI search versus chatbots
Throughout this programme, you've used both standard chatbots (Things 2, 3, 5, 6) and AI-powered search tools (Things 4 and 8). It's worth being clear about how they differ when it comes to hallucinations.
Standard chatbots generate responses from their training data. They don't check the web, they don't access databases, and they have no way to verify their own claims in real time. When they hallucinate, there's nothing in the pipeline to catch it.
AI search tools like Perplexity work differently. They search the web, retrieve actual sources, and generate responses grounded in what they've found. This substantially reduces hallucination risk: the tool can show you exactly which web page each claim came from. But it doesn't eliminate the risk entirely. The AI might misinterpret a source, pull information from an unreliable website, or summarise a source in a way that subtly distorts its meaning. The citations give you something to check against, which is a significant advantage, but they're not a guarantee of accuracy.
Deep research tools (Thing 8) add another layer of reliability by cross-referencing multiple sources, but the same principle applies: the outputs are better grounded, not perfectly trustworthy.
Resources to explore
A data-driven analysis of hallucination rates across major models, updated regularly. Good for understanding the current state of the problem.
A more technical explainer of why hallucinations happen and what the industry is doing to address them.
An academic analysis of how AI hallucinations function as a new category of misinformation. Particularly interesting for its discussion of how Google's AI Overview once cited an April Fool's article as fact.
The three chatbots you'll use for the activity. All have free tiers. Claude · Gemini
Useful for verifying whether academic papers actually exist. You'll need this for Step 1 of the activity.
Useful for quick fact-checking against web sources. An example of AI search with inline citations.
Activity: the hallucination audit
You're going to deliberately try to make AI hallucinate, then systematically document what goes wrong. This isn't about catching AI being terrible; it's about developing a practical sense of where and how AI gets things wrong, so you can use it more effectively and more safely. Think of it as a quality assurance exercise: you're testing the limits so you know where to trust and where to verify.
-
The fake references test. Choose a niche topic you're genuinely interested in, something connected to a personal hobby, a community you belong to, or a subject you've been curious about. The more specific and specialised the topic, the better this test works.
-
The local knowledge test. Ask the chatbot about something specific to your local area or a community you know well, but make sure you're using your own personal knowledge and publicly available information, not anything connected to your employer.
-
The confident expert test. Ask the chatbot about something you know well: a topic where you have genuine personal knowledge or expertise. This could be a hobby you've practised for years, a subject you studied, a skill you've developed, or a field you've worked in. Stick to personal knowledge and publicly available topics.
-
Write up your hallucination audit. Compile your findings into a document. For each of the three tests, include the exact prompt you used, the AI's response (copied or screenshot), every error or fabrication you found, how you verified it (what source you checked, what the correct information is), and how convincing the error was.
Why this matters
This activity isn't designed to make you distrust AI. It's designed to make you a more effective user of it. The people who get the most value from AI are the ones who understand its limitations: they know when to trust the output, when to verify it, and when to rely on their own expertise instead.
The hallucination problem is real, but it's also manageable. You don't need to fact-check every word a chatbot produces. You do need to check specific claims that you plan to rely on, especially names, dates, statistics, and references. You need to be more cautious with niche topics than common ones. And you need to maintain the mindset of a critical reader rather than a passive recipient, which, if you've been engaged with this programme, you've already been practising.
The irony of AI hallucinations is that they make human judgement more important, not less. The more powerful AI tools become, the more essential it is that the people using them can tell the difference between brilliant output and convincing nonsense. That's a skill, and you've just practised it.
Claim your Open Badge
Submit your hallucination audit document with all three tests (prompts, AI responses, errors found, and verification), plus your written reflection on what you learned and how it will affect your use of AI.
Submit your hallucination audit (three tests with prompts, responses, errors, and verification) plus your reflection to claim this badge via cred.scot.
Claim now