This is very impressive. I ran it through ChatGPT and got the following response. The only mistake it made is "really" instead of "rarely".

------

> Correct the spelling:

Eye halve a spelling chequer It came with my pea sea It plainly marques four my revue Miss steaks eye kin knot sea.

Eye strike a quay and type a word And weight four it two say Weather eye am wrong oar write It shows me strait a weigh.

As soon as a mist ache is maid It nose bee fore two long And eye can put the error rite It's rare lea ever wrong.

Eye have run this poem threw it I am shore your pleased two no It's letter perfect awl the weigh My chequer tolled me sew.

> I have corrected the spelling errors in the poem:

"I have a spelling checker

It came with my PC

It plainly marks for my review

Mistakes I cannot see.

I strike a key and type a word

And wait for it to say

Whether I am wrong or right

It shows me straight away.

As soon as a mistake is made

It knows before too long

And I can put the error right

It's really ever wrong.

I have run this poem through it

I am sure you're pleased to know

It's letter perfect all the way

My checker told me so."

I added "When you are spell-checking, please take extreme care when picking the corrected versions of each word." to my instructions and it got it completely right.

There's dozens of websites with the corrected poem out there, hard to know if ChatGPT just 'remembered' it.

Well it got it wrong, so no?

Language models like ChatGPT seem to "remember" paragraphs of text in a highly compressed form which is not always fully accurate when reproduced - so a mistake does not tell us whether the text is from memory.

Lossy reconstruction of training data is still memorisation.