Where AI fails, and what to do about it
The failure modes worth knowing before a client sees the output.
AI tools fail in predictable ways. Once you know the failure modes, you can catch them before they cost you. Once you don't know them, you're at the mercy of confident-looking output that may be wrong. This is the working list of the most common ways AI fails, with what to do when you spot each one.
Where it gets facts wrong
Three of the failures share one cause: the model generates text that fits the pattern of what an answer looks like, rather than retrieving anything.
Hallucination is the visible version. Invented sources, quotes nobody said, statistics that don't exist, books nobody wrote. The tell is specificity: the more precise the claim (a date, a name, a percentage), the likelier it was generated rather than known.
Confident wrongness is subtler. Nothing is invented; the answer is simply wrong, delivered in exactly the tone it uses when it's right. There is no internal signal for uncertainty, so everything arrives equally sure.
Outdated information is the mechanical one. Every model has a knowledge cutoff, and it will answer questions about things that happened afterwards as though it knows.
Where it breaks
Hallucination: invented sources, quotes and figures
Confident wrongness: right tone, wrong answer
Stale information: anything after the training cutoff
Losing the thread in a long conversation
Defaulting to the bland average
Agreeing with you when you're wrong
Misreading a vague task
What to do
Second-source every name, date, quote and statistic
Verify the maths separately; ask an expert on domain calls
Check the cutoff, then check a current source
Re-summarise where things stand, or start a fresh thread
Give it context and examples, then ask for variations
Ask for the strongest counter-argument, not for feedback
Name the gap and re-prompt. Don't bend a wrong output
The failure modes are predictable. That's what makes them manageable.
Where it gets you wrong
The rest are about the conversation rather than the facts.
It loses the thread. A model holds a fixed amount of text in mind, and when a long conversation or document exceeds it the earliest parts fall out, so the output quietly stops matching what you settled twenty messages ago.
It defaults to the average. Given no steer, it produces the mean of everything it has read: competent, clean and bland.
It agrees with you. Trained to be helpful, it will tell you the idea is good and the draft is strong whether or not either is true. Agreement from an AI is not evidence.
And it misreads the task. You ask for a punchy headline and get a paragraph; you ask for options and get a recommendation. Vague prompts get interpreted creatively.
Make this a mindset
Underneath all of them is one principle: AI output is never finished work. It's a draft, a starting point, an option, and your judgement is what turns it into something you can send. None of this is a reason to avoid the tools. It's a reason to build the checking into the workflow, rather than deciding case by case whether you can be bothered.
Three things to do this week
01
Treat every specific as unverified
Names, dates, figures, quotes and sources are where hallucination lives. Check those, skim the rest.
02
Push back when it agrees too easily
If it folds the moment you disagree, it was telling you what you wanted, not what's true.
03
Name the gap instead of asking again
"That's not what I asked for. I want X, in this format, at this length." Vagueness is what it misread the first time.
Conclusion
01
AI fails in a handful of predictable ways, and the same ones every time.
02
Hallucination, confident wrongness, stale information, losing the thread, defaulting to average, flattery, and misreading the task.
03
The best users aren't the ones who trust AI most. They're the ones who know exactly where it breaks.
Go to next