AI with/vs Human Writing

• Encourage critical reflection on the difficulty in distinguishing between human and AI-generated text; 

• Raise awareness of the capabilities and implications of AI in writing tasks; 

• Ethical considerations of AI assistance in academic writing.

What you need

Setting it up

• Participants need access to the Real or Fake Text game and/or Paul Fyfe's paper. 

How long does it take?

• A game of Real or Fake Text is expected to take no more than 5-10 minutes. 

• Reading and reflecting on Fyfe’s paper will likely take 1-2 hours. 

How it works

1. Play the Real or Fake Text game and/or read Paul Fyfe's paper.

Suggested follow-up

• For Real or Fake Text, reflect on your experience, highlighting any insights gained or observations made during gameplay.

• For Fyfe’s paper, reflect on one or more of the key ideas presented in the paper, such as the nature of writing as a cyborg technology, patchwriting, the hermeneutic contract between authors and readers, or the fetishization of 'originality' in academic writing. 

• Discuss the implications of AI in writing tasks, exploring topics such as authenticity, authorship, and academic integrity.

Where it works well

Real or Fake Text

• It's clear that distinguishing text written by a computer from those written by human is very difficult, although genre seems to matter (recipes in the game are easier) 

• The list of 'whys' for fake text from the game can be helpful in thinking about what distinguishes computer generated text. 

Fyfe paper

• Presents a strong case for reflecting and carefully considering generative AI’s impact by delineating how technologies we use to write affects how we write (and think).

What to watch out for

Real or Fake Text

• It has been pointed out that a game with unfamiliar text, author, and slow reveal isn't quite the same as reading the whole essay and having knowledge of the writing style of the authors (e.g. students).

• Note that the game uses computer generated text rather than generative AI generated text - though I think the point that in small doses or with clever editing they are difficult to tell apart is still valid. 

Fyfe paper

• The study was done with GPT-2, which is significantly less capable than the generative AI models we now have. This and the media hype of generative AI may cause students to feel differently about using generative AIs to assist their writing today. 

Authorship