1 of 64

Generative AI and �Education Research

Roehampton doctoral students’ conference�Prof Miles Berry

These slides: bit.ly/genaier

23 March 2024

2 of 64

3 of 64

Foundations

Applications

Implications

How does this work

How can it be used

What this all means

4 of 64

Foundations

5 of 64

Foundations

Input

Model

Output

ML Algorithm

Training data

6 of 64

7 of 64

8 of 64

A new common language?

Google’s researchers think their system achieves this breakthrough by finding a common ground whereby sentences with the same meaning are represented in similar ways regardless of language – which they say is an example of an “interlingua”. In a sense, that means it has created a new common language, albeit one that’s specific to the task of translation and not readable or usable for humans.

9 of 64

10 of 64

11 of 64

Open AI?

GPT-4 is a Transformer-style model [39] pre-trained to predict the next token in a document, using both publicly available data (such as internet data) and data licensed from third-party providers. The model was then fine-tuned using Reinforcement Learning from Human Feedback (RLHF) [40]. Given both the competitive landscape and the safety implications of large-scale models like GPT-4, this report contains no further details about the architecture (including model size), hardware, training compute, dataset construction, training method, or similar.

Open AI, 2023

12 of 64

“It uses bits of what it’s heard and stitches them into something new … that’s exactly what we do.”

Ockleford, 2024

13 of 64

Applications

14 of 64

15 of 64

How can it help?

It’s very well read

It writes very well

It programs well too

It tries to be helpful

16 of 64

17 of 64

18 of 64

19 of 64

20 of 64

21 of 64

22 of 64

23 of 64

24 of 64

25 of 64

26 of 64

27 of 64

28 of 64

29 of 64

30 of 64

31 of 64

32 of 64

33 of 64

34 of 64

35 of 64

36 of 64

37 of 64

38 of 64

39 of 64

40 of 64

41 of 64

Other things that it can help with

Checking for readability / SPAG / argument

Reducing word count

What have I missed?

Transcribing interviews

Translations

Role play

42 of 64

Prompting well

Completion

Primary content

Examples

Cue

Supporting content

Have a conversation!

Be clear and precise

Break the task down

Chain of thought

Persona

System messages

Fine tuning

43 of 64

There are limits

It doesn’t really understand

It doesn’t really think - problem solving is a problem

It’s over-confident

It does make things up

It sometimes pays attention to the wrong thing

It’s kind rather than critical

GPTn are not up to date

Reliability costs

44 of 64

�Implications

45 of 64

Is it cheating if ChatGPT…

Explains something to you?

Gives you ideas for your paper?

Suggests how to improve your paper?

Writes your paper for you?

46 of 64

Academic integrity

Passing off the work of a generative AI tool as the student’s own is to be treated as any other form of plagiarism and colleagues will follow the university’s usual disciplinary processes. Similarly, using data from a generative AI tool in place of experiment, interview or survey would be considered falsification and treated as such.

47 of 64

Citing the AIs - persistent URL

In-text citation

The AI-generated flower (Shutterstock AI, 2023)

Reference list

Shutterstock AI (2023) Photo of pond with lotus flower[Digital art]. Available at: https://www.shutterstock.com/image-generated/photo-pond-lotus-flower-2252080005 (Accessed: 31 March 2023).

48 of 64

Citing the AIs - transient response

In-text citation

When prompted by the author, ChatGPT responded with a ‘definition of academic integrity’ (OpenAI ChatGPT, 2023). A copy of this response is in Appendix 1.

Reference list

OpenAI ChatGPT (2023) ChatGPT response to John Stephens, 2 April.

49 of 64

Publishers’ policies

AI use must be declared and clearly explained in publications such as research papers, just as we expect scholars to do with other software, tools and methodologies.

AI does not meet the Cambridge requirements for authorship, given the need for accountability. AI and LLM tools may not be listed as an author on any scholarly work published by Cambridge

Authors are accountable for the accuracy, integrity and originality of their research papers, including for any use of AI.

Any use of AI must not breach Cambridge’s plagiarism policy. Scholarly works must be the author’s own, and not present others’ ideas, data, words or other material without adequate citation and transparent referencing.

CUP, 2024

50 of 64

Data protection and IP

There are risks to privacy and intellectual property associated with the information that students and/or staff may enter. It is important to consider GDPR, data protection and whose intellectual property may be infringed when using generative AI.

51 of 64

52 of 64

Terms and �conditions apply

53 of 64

What stakeholders expect

Publishers, research funders and the public have a reasonable expectation that Roehampton's research is the original work of academics and research students and not the output of generative AI, unless the contribution of generative AI is clearly acknowledged. In any circumstance, academics must abide by all relevant ethical and legal frameworks and any contractual obligations for funded or published research.

54 of 64

55 of 64

56 of 64

57 of 64

Bias…

58 of 64

59 of 64

Mind and society

The child begins to perceive the world not only through his [or her] eyes but also through his [or her] speech

Vygotsky, 1978

60 of 64

Talk as the currency of learning

Talk is … the currency of learning — how we develop and shape our ideas, deepen our thinking, explore subject matter and share our thoughts and feelings.

61 of 64

Should learning be hard?

Learning is at its best, human beings are at their best, when they are challenged and overcome those challenges. AI will make life easy and strip away learning and teaching — unless we get ahead of it.

62 of 64

Are you thinking

As AI performance improves, human overseers face greater incentives to delegate. If the AI appears too high quality, workers are at risk of “falling asleep at the wheel" and mindlessly following its recommendations without deliberation. In such settings, maximizing combined human/AI performance requires trading off the quality of AI against the potential adverse impact on human effort.

63 of 64

And here's how you should think about memory: it's the residue of thought, meaning that the more you think about something, the more likely it is that you'll remember it later.

Willingham, 2008

64 of 64

Any questions?

These slides: bit.ly/genaier

m.berry@roehampton.ac.uk

0208 392 3241