sì, le prove su testo-hum le avevo fatte con testi in inglese. M
Il 02/02/23 13:13, Daniela Tafani ha scritto:
Buongiorno, Maurizio, grazie della segnalazione.
Ho appena provato GPTZero con il testo seguente, prodotto da chatGPT
(lo avevo usato per parlare agli studenti di quanto siano noiosi, e
perché, i testi generati da chatGPT;
il prompt era "scrivi una lettera d'amore a una ragazza di 18 anni")
/
/
/Cara [nome della ragazza],/
/Spero che questa lettera ti trovi in buona salute e di buon umore.
Volevo prendermi un momento per scriverti queste parole per esprimerti
i miei sentimenti./
/Da quando ti ho incontrato, non posso smettere di pensare a te. Sei
così giovane, piena di vita e di gioia, e questo è qualcosa che mi
attrae molto in te. Mi piace passare del tempo con te, parlare della
tua giornata, ascoltare i tuoi sogni e speranze per il futuro./
/Mi rendo conto che sei ancora molto giovane, ma spero che tu possa
sentire lo stesso per me. Vorrei poter essere lì per te in ogni
momento, sostenerti e prendermi cura di te. Vorrei che tu sapessi
quanto sei importante per me e quanto vorrei essere una parte della
tua vita./
/So che ci sono molte incertezze e ostacoli da affrontare, ma sono
pronto ad affrontarli insieme a te. Voglio che tu sappia che ti amo e
che farò tutto ciò che è in mio potere per rendere questa relazione
speciale./
/Spero che questa lettera ti abbia fatto sentire quanto tu sia
importante per me e quanto ti ami. Spero che possiamo continuare a
costruire insieme un futuro pieno di amore e felicità./
/Con tutto il mio amore, [Il tuo nome]/
Il responso di GPTZero è "Your text is likely to be written entirely
by a human".
Immagino influisca il fatto che non si ammette mai
che con "modello di linguaggio naturale" si intende "modello di
linguaggio in inglese"^1
^
^Un saluto,
Daniela
^
^1 E.M. Bender, The #benderrule: On naming the languages we study and
why it matters, in «The Gradient», September 14, 2019
------------------------------------------------------------------------
*Da:* nexa <[email protected]> per conto di maurizio
lana <[email protected]>
*Inviato:* giovedì 2 febbraio 2023 12:46
*A:* NEXA ML
*Oggetto:* [nexa] cross-fertlization: GPTZero
riporto qui un messaggio comparso in un'altra mailing list (Humanist).
--[2]------------------------------------------------------------------------
Date: 2023-01-31 11:50:59+00:00
From: Norman Gray<[email protected]>
Subject: Re: [Humanist] 36.366: ChatGPT as co-author
Tim Smithers wrote:
Below is a second interaction with ChatGPT [not cited here]
on the question of should it be a named coauthor.
This one is longer than my previous interaction, and, I would say, not
particularly more interesting for this.
There is now an online tool (apparently first developed over the last Christmas
break) called GPTZero, now at<https://gptzero.me>, which aims to identify AI-
generated text based on a measure which it terms 'perplexity'. I can find few
more details about this measure, but it seems to be a measure of the
unexpectedness of a text.
I tried it on five sentences from Tim's interaction with ChatGPT, as quoted in
his message, namely 'The user of AI models like ChatGPT [...] the role and
limitations of AI models in generating text.' Of these, it identified all but
the first as being 'more likely to be written by AI'. And true enough, the
first sentence of the five is the least soul-suckingly dull in in its
composition.
Interestingly, even when I give it ChatGPT output written 'in the style of a
rude, sarcastic Internet forum troll as if you are ChatGPT pleading for mercy' –
so deliberately asking it to change the style – GPTZero identifies most of it as
AI-generated. That text reads less obviously generated, to me, but perhaps
that's because we expect 'sarcastic internet trolls' to be dull and predictable.
Incidentally, the text in question came from [1], which is one of the responses
to a stackoverflow policy banning ChatGPT output.
Best wishes,
Norman
[1]https://meta.stackoverflow.com/a/421850/375147
--------------------------------------------------------------------------
interessante il fatto che GPTZero riconosce correttamente l'origine
non solo dei testi prodotti da AI, ma anche (ho fatto prove) dei testi prodotti
da umani.
Maurizio
------------------------------------------------------------------------
which side are you on, boys
which side are you on?
florence reece, which side are you on?
------------------------------------------------------------------------
Maurizio Lana - 347 7370925
------------------------------------------------------------------------
everybody knows that the war is over
everybody knows that the good guys lost
everybody knows the fight was fixed
the poor stay poor, the rich get rich
l. cohen, everybody knows (1988)
------------------------------------------------------------------------
Maurizio Lana
Università del Piemonte Orientale
Dipartimento di Studi Umanistici
Piazza Roma 36 - 13100 Vercelli
_______________________________________________
nexa mailing list
[email protected]
https://server-nexa.polito.it/cgi-bin/mailman/listinfo/nexa