Somewhere in your organisation right now, somebody is about to send an email written by AI containing a made-up fact, and a client will notice the mistake.
This isn't a hypothetical. In 2023, two New York lawyers submitted a legal brief filled with case citations that sounded perfectly plausible, including correct format, correct tone, and confident legal language.
But there was a problem. None of the cases actually existed.
ChatGPT had invented them, and neither lawyer had checked. Not only did they have to live with the embarrassment that the judge spotted it, but the firm was fined and became an international punchline - LinkedIn, eat your heart out.
And it's not just ChatGPT. In a separate case, even Anthropic's own lawyers were caught out by their own AI: when Claude was asked to format a legal citation, it invented a different title and different authors for the source, authors who had never actually worked together. The underlying article was real, the link worked, but the details around it weren't. Anthropic's manual citation check missed it, and the company had to formally apologise to the court, calling it "an honest citation mistake".
This story doesn’t serve to shame the lawyers, but to highlight the very real pattern that is playing out in business decks, financial models, and client emails across every industry, every week.
The goblin problem
If you had been reading the news in May 2026, you’d have seen a rather odd headline about AI and goblins. Something like “OpenAI tells ChatGPT models to stop talking about goblins”.
ChatGPT, out of seemingly nowhere, started talking more and more about goblins in its query answers. You might be thinking, well, if you ask for “top 10 goblins in films”, that makes sense, but what if you’d asked “good conversation starters in a job interview”, you might have been a little more worried (depending on the interviewee, of course).
OpenAI, the team behind ChatGPT, claimed that the term “goblin” had risen by 175%, and other gobliney-terms like “gremlin” had also been generating 52% more than usual.
While they never quite got to the bottom of the reason for this, developers discovered it was triggered by the launch of GPT-5.1 in November 2025. The company even had to generate this line of code: "never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant to the user's query", to protect themselves from future creature-related queries.
It's not only humans who are capable of hallucinating, it seems. Hallucination is the term for when LLMs like Gemini or ChatGPT start producing false or incorrect content. This can include sourcing wrong statistics, mixing up dates, or even bizarre examples like Gemini being adamant that there are 8 days of the week instead of 9 (we’re just joking, not hallucinating!).
This tells us two things: AI-produced content should not be the end result and users need to be fluent in digital and AI usage.
Digital literacy is not just about the tool
Before AI came along, digital literacy was largely used to describe skills like Microsoft Excel or Python coding.
These are all useful skills, but it’s worth noting they’re static ones. When you do something wrong, you know almost immediately. A broken spreadsheet formula throws an error or notification that it can’t complete a formula.
But a broken AI answer looks identical to a correct one, and it tells you it’s correct with unshakeable confidence. AI has changed the entire nature of the problem, because AI systems don't fail the way old software failed.
And that’s what we all have to account for when we risk putting too much faith in technology and not enough in our own human judgement. This partial scepticism is one element of having greater digital fluency.
The trust paradox
The unfortunate truth of AI is that those who need to use it more are least equipped to spot the mistakes it might make.
For example, if you’re a professor of classicist literature, you might recognise if a tool like ChatGPT assigns the quote “A woman must have money and a room of her own if she is to write fiction” to Charles Dickens instead of Virginia Woolf. But many people wouldn’t be able to spot the error.
This is the trust paradox at the centre of workplace AI use: the less you know about a subject, the more convincing AI's mistakes become, and the more likely you are to be leaning on AI for exactly that subject.
It's not a flaw in any particular person's judgment; it's just a structural feature of how these tools get used.
Digital and AI literacy is, in large part, the set of habits that counteracts this paradox. Just knowing which situations demand a second look, and building the reflex to take it, can save you a lot of headache!
What AI literacy actually involves
The first step to gaining better digital literacy is by understanding exactly what it encompasses. The World Economic Forum released a really useful and comprehensive breakdown of AI literacyrecently; you can find the visual below.
The second step is recognising the imperfect nature of technology, and how to use critical thinking to combat it. You can do this in many ways, but we’ve highlighted four core considerations below.
Technical know-how
You don’t have to be an expert, but having a rough knowledge of how tools like ChatGPT operate is useful in context. For example, large language models predict plausible next words based on patterns in training data, rather than "knowing" facts the way a database does. This single piece of understanding explains almost every quirk professionals run into: why AI can be extremely fluent and confidently wrong at the same time, why it struggles with things that require precise recall (exact dates, exact numbers, niche facts) more than things that require synthesis and language, and why it can "hallucinate" plausible-sounding sources, quotes, or case law that were never real.
Prompting expertise
Also, you need to understand how to actually use the tool correctly. You all know that LLMs rely on prompts to produce an answer, so it makes sense that vague instructions produce vague, generic output. Instead, if you provide it with specific instructions, with context, constraints, and examples of what "good" looks like, it will produce dramatically better output. This isn't a technical skill so much as a management skill: you're delegating to a tireless, extremely literal collaborator that will confidently do the wrong thing if you're unclear.
Verification habits
Knowing which outputs need a fact-check and which don't will also save you lawyer embarrassment. A draft email summarising a meeting you attended is low-risk; you were there, you'll notice if it's wrong, and it’s information that mostly you’ll keep. A statistic, a legal citation, a technical spec, a number in a client-facing deck are much more high-risk. These need verification every time, because they're exactly the kind of specific, checkable detail that AI tools get wrong most often and where being wrong is most costly.
Human judgement
Know which tasks are AI-appropriate and which aren’t. Things like first drafts, brainstorming, summarising, formatting can be a useful way to integrate LMMs into your workflow. But things that involve genuine judgment about risk, ethics, people's careers, or legal exposure need human judgement. Knowing where that line sits, for your specific role, is arguably the most important and least-discussed layer.
Outside of AI-literacy
Of course it’s not all about AI; you can be proficient in many other digital skills, such as:
Information & media literacy
Communication & collaboration
Privacy & security
Data literacy
Technical/functional skills
Digital citizenship & wellbeing
But the same underlying pattern applies to all of them: understanding the tool, knowing its limitations, and judging when it's actually the right one to reach for. That's the transferable skill, and it matters more, not less, as new tools keep arriving.
Privacy and security deserve a closer look, because it's not just about hallucinated facts; it's about what happens to what you type in. In July 2026, hundreds of shared Claude conversations turned up publicly searchable on Google after a user discovered the pages were missing the tag that normally keeps them out of search results. It wasn't a breach or a hack, just a configuration gap, but conversations containing personal and business details were indexed and findable by anyone who knew where to look before the links were pulled from search results. It's a reminder that a chat feeling private and a chat actually being private aren't always the same thing.
If there's one thing that's clear, it's that we all need to get good at fast-learning, because the tool in front of us today won't be the last one.
Habits rather than rules
The instinct in a lot of workplaces is to solve this with policy, a list of "approved" AI uses, a compliance memo, a training video nobody watches twice. Policies have their place, but literacy isn't really a rulebook problem; it’s more of a habit problem.
A few habits worth building, regardless of role or industry:
Treat fluency as neutral, not as evidence. Confident, well-written output is not the same as correct output. Train yourself to notice when you're being persuaded by tone rather than substance.
Match your scrutiny to the stakes and the specificity. Vague, low-stakes drafts need light review. Anything with a specific number, name, date, citation, or claim that could embarrass you or harm someone if wrong deserves a deliberate check, every time, regardless of how confident it sounds.
Keep your own expertise in the loop. The safest way to use AI in your domain of expertise is as a first draft you'll correct. The riskiest way to use it is in a domain where you can't tell a good answer from a bad one, which is exactly when it's tempting to trust it most.
Talk about mistakes openly. The teams that build genuine AI literacy fastest are the ones where someone catches an AI error and says so out loud, instead of quietly fixing it and moving on.
The point of thinking about this
Digital and AI literacy isn't a technical credential, and it isn't about becoming suspicious of every tool you use. It's closer to a form of professional maturity, the same instinct that makes a good editor question a suspiciously tidy quote, or a good auditor double-check a number that looks too clean.
AI has just made that instinct newly essential, because for the first time, the tool answering your question can be wrong in a voice that sounds exactly like being right.
So, if you want to write an email written by AI containing a real fact that a client will notice for all the right reasons, make sure your AI literacy skills are up to scratch.
Want more AI insights? Read our other articles on the topic, including:
The People Who Will Not Be Replaced by AI
AI and the New Environment for Early Careers
Rewiring the C-Suite for AI: Roles, Strategy, and Chief AI Officer Responsibilities
How AI Recruitment Bots are Changing the Game for Senior Hires
Frequently Asked Questions
How do I know when an AI output needs a fact-check?
As a rule, anything with a specific number, date, name, or citation deserves a check every time, regardless of how confident it sounds. Low-stakes drafts (a meeting summary you were present for) need far less scrutiny than anything client-facing or high-stakes.
What AI skills should I be building right now?
Start with prompting: the clearer your context, constraints, and examples, the better your output. From there, build the habit of matching your review effort to the stakes of what you're producing.
Is there a framework I can use to benchmark my own AI literacy?
The World Economic Forum's breakdown (referred to earlier in this article) is a solid starting point if you want to see where your skills sit and what to prioritise next.
How is AI literacy showing up in what employers look for?
It's increasingly part of what gets screened for, even at entry level, particularly around prompting ability and knowing when to flag uncertainty rather than present it as fact.
How can I help my team build these habits, not just individual skills?
The teams that build AI literacy fastest are the ones where people flag mistakes out loud rather than quietly fixing them, so start there before reaching for a formal policy.
.png)
