Office Watch calls AI Almost Intelligent for good reason. ChatGPT, Copilot, Claude and Gemini can rewrite a paragraph in seconds, summarize a long report, or turn rough notes into a polished email before your coffee cools. AI can also invent quotes, cite court cases that never existed, get a date wrong by a year, slip wrong numbers into a beautiful table, all presented with total certainty. The technology is genuinely impressive and genuinely fallible at the same time. Here is what Microsoft, OpenAI and Google will not put on the box, plus three prompts that keep AI honest.
The honest acronym Microsoft won’t put on the box
AI officially stands for Artificial Intelligence. At Office Watch we use a different version: Almost Intelligent.
That’s not a joke. It’s the most accurate description we’ve found for what these tools actually do and warn that AI is a great tool but needs human oversight.
Almost, but not quite
Modern AI is genuinely impressive. It can rewrite a paragraph in three different tones in under a second. It can summarize a 40 page report into a few sensible paragraphs. It can turn rough notes into a polished email in less time than it takes to make a cup of tea.
It can also:
- Invent a quote and attribute it to a real person
- Confidently cite a court case that doesn’t exist
- Get a date wrong by a year and present it with total certainty
- Format a beautiful table with one of the numbers quietly wrong
- Insist that a feature exists in Word when it doesn’t
- Translate a phrase into the wrong language while claiming it’s right
While researching Write better & faster with any AI seen all of those happen, repeatedly, across ChatGPT, Copilot, Claude and Gemini.
Some are funny. Some are embarrassing. Some, if you publish without checking, can end a career.
This isn’t a bug, it’s how the technology works. AI doesn’t know anything in the human sense. It produces text that statistically resembles a correct answer. Most of the time that’s good enough. The rest of the time it isn’t, and the AI sounds equally confident in both cases.
That’s what makes it Almost Intelligent. The almost is the dangerous part.
What Microsoft, OpenAI and Google won’t tell you
Read any official AI marketing and you’ll find phrases like “your everyday AI companion” or “unlock new levels of productivity”. You will not find phrases like “will occasionally fabricate citations” or “may confidently lie about your own files”.
If you look carefully, there’s a faint text warning like “AI can make mistakes. Please double-check responses.” Or “AI-generated content may be incorrect”
That’s the gap we exist to fill.
Office Watch has never been on the Microsoft payroll. We’ve spent more than 25 years explaining what Microsoft Office actually does, including the parts Microsoft would rather you didn’t notice. AI is the same job, with higher stakes.
The hype machine wants you to trust AI. We want you to trust but verify AI. Those are very different things.
What this means for you
If you use AI for writing, three habits will save you from the worst of it:
- Treat every AI output as a first draft, not a final answer. Even if it sounds polished. Especially if it sounds polished.
- Check anything specific. Names, dates, numbers, quotes, citations, version numbers, prices. If the AI gave you a fact, verify it before it leaves your screen.
- Read for voice, not just accuracy. AI tends to flatten everything into the same smooth, slightly bland register. If you can’t hear yourself in the writing, rewrite the bits where your voice should be loudest.
None of this is about being scared or avoiding AI. It’s about knowing what tool you’re holding. A power drill is fast and useful. It’s also dangerous in the wrong hands.
Prompts that keep AI honest
Three small additions to the way you ask AI for things will dramatically reduce the trouble it causes you. All four major AI services (ChatGPT, Copilot, Claude and Gemini) respond to these.
Catch the hallucinations before you publish
‘Hallucinations’ is the techy understatement for when AI makes stuff up. After AI gives you a draft, paste this straight back into the same conversation:
Review your previous response. List every name, date, number, quote and citation you included, one per line. For each, tell me honestly whether you are confident in the source or not. Don’t defend, just list.
You’ll be surprised how often AI freely admits to “low confidence” on facts it had just presented with total certainty. The ones it flags are the ones you need to verify before publishing.
Keep your own voice
Generic AI writing happens because you didn’t tell the AI whose voice to use. The fix:
Below are three samples of my actual writing: [paste 200 to 300 words of your own writing].
Now rewrite the draft below to match the voice, rhythm, sentence length and word choices in those samples. Do not add phrases I wouldn’t use.
The output is usually unrecognizable from the bland default. It’s the single most useful AI writing trick we know, and it costs nothing but a minute of copy paste.
Strip the AI tells
Modern AI has verbal tics: “delve”, “tapestry”, “navigate the complexities”, “in today’s rapidly evolving landscape”, three item lists with parallel structure, and the habit of opening every reply with “Certainly!” Catch them with:
Read the draft below. List every phrase, word, or formatting habit that sounds like generic AI writing. Do not rewrite the draft. Just list the AI tells so I can fix them myself.
This works particularly well because you keep editorial control. The AI flags the problems, you decide what to do about them. Asking AI to “fix” its own AI voice usually just produces more AI voice in different clothes.
Bonus: the privacy gut check
Before you paste anything into a public AI service, ask yourself the question we use every day at Office Watch:
If a stranger printed this exact text and pinned it on a noticeboard, would that be a problem?
That’s not a prompt for the AI. It’s a prompt for you. If the answer is yes, don’t paste it. Use a private AI service, an offline tool, or rewrite the sensitive parts before you ask for help.
The Almost Intelligent way
We’ve built a whole book around this stance. Write better & faster with any AI shows you how to get the speed and clarity AI offers, without falling for the confident hallucinations or the bland AI voice.
It’s not opposed to AI. It’s pro skepticism. There’s a difference.
The book is US$9.95, instant PDF download, with 230+ ready to copy prompts in a bonus Word document. The four prompts above are a small taste. Independent, plain English, and written by people who use AI every day and still don’t quite trust it.
Which is, we’d argue, exactly the right amount of trust.
Write better & faster with any AI
Excel’s COPILOT() Function Gets Facts Wrong: Our Tests Prove You Cannot Trust It
What the !@^? Is Microsoft Copilot “Work IQ”?
Copilot Can Now Actually Do the Work in Word, Excel, and PowerPoint
Microsoft buries the Copilot button and floats a nag in its place
