AI Myths I Tested: 5 Claims, and 4 Were Flat-Out False

I gave an AI twenty word problems with hidden traps. It got three of the tricky ones wrong and sounded completely sure about all of them.

That was one of five tests. The internet is full of confident AI claims, and I got tired of taking them on trust, so I picked five of the most repeated ones and ran them myself on real tasks. Some held up. Most broke once the task got real.

How I ran the tests

I used the same model and account for each claim. I repeated every task three times to rule out luck, and I kept the outputs. I wasn't trying to mock the tools. I wanted my own reasons to trust them instead of borrowing someone else's hype.

Claim 1: "AI can write a whole app from one sentence"

I tried it with a real, moderately complex app idea. It produced a convincing skeleton that broke on the first edge case and had no real architecture. Verdict: false for anything you would actually ship. It is still great for a prototype you plan to rewrite.

Claim 2: "AI never makes math errors"

This is the word problem test from the top. It nailed the simple problems and fumbled three of the tricky ones with total confidence. Verdict: false. It sounds just as sure when it is wrong as when it is right, so always check the numbers.

Claim 3: "Bigger context means it remembers everything"

I dropped in a long document and asked about a detail buried in the middle. Sometimes it found it. Other times it quietly used the start and ignored the rest. Verdict: half true. More context helps, but the model doesn't pay equal attention to every part. Structure your docs.

Claim 4: "Agents can run your business unattended"

I let an agent handle a multi-step task with a real account. It completed the happy path, then nearly did something dumb on a surprise input. Verdict: false if you mean unattended. It works fine with a human watching, but "set and forget" isn't here yet.

Claim 5: "Local small models are useless"

I ran a tiny local model on a plain laptop for daily writing and notes. It was slower and less clever than the frontier API, but it did the job and kept my data private. Verdict: false. Small local models are underrated for the right tasks.

Something else I noticed: wording matters

Along the way I gave the same vague request two ways. The first version got a confident but wrong answer. The second, with one clarifying sentence added, got it right. The model only has your words to work with, so say clearly what you want.

What I took away

Four of five claims failed. Demos show the best case. Real work shows the average case, and the average case is what you actually get.

AI is useful, and it is also overclaimed. So I treat every big AI claim as a guess until I have tested it on my actual task. The claims that survive that test are the ones I build on.

FAQ

Does this mean AI is overrated?

No. Specific claims need specific tests, and the useful ones pass.

Should I test every claim myself?

Only the ones you plan to build on. For the rest, stay skeptical and move on.

This testing habit is the same one behind a year of shipping in public.