The Ugly, the Bad and the Good
The Good, the Bad and the Ugly
will AI upend science dissemination?
Kavli day
AI and machine learning
in modern-day science
Den Haag, 5 October 2023
Jean-Sébastien Caux
Plan of the talk
Brief (from Kobus, Gijsje):
To what extent are you, as an editor, worried or excited or both about how chatGPT (or similar program) could affect the dissemination of our scientific results?
We need to talk.
- The Ugly
- The Bad
- The Good
The Ugly →
Instrumental convergence
Generated with StableDiffusionTurbocharged paper mills
Generated with StableDiffusionDeepfakes, Obfuscation & Impersonation
"Write this report in the style of [your academic rival]"
Supercharged fraud...
← The Bad
Copyright
LLMs are guilty of gargantuan-scale copyright infringement...
...but AI-generated stuff can't be copyrighted
Théâtre d’Opéra Spatial, Jason Allen (using Midjourney)
Confusion on editorial guidelines
Lack of clarity/consistency of editorial policies
Should ChatGPT be a coauthor? (plagiarism?)
Difficulty of detecting AI content using traditional plagiarism tools
Lack of honesty (ouch!) on AI use acknowledgements
Laziness, slide towards mediocrity?
Laziness empowered: "just let the machine do it"
Safeguard (?!): "Humans should always do a check"
Have no illusions: this won't happen, humans like to cut it short and systems tend to reward bullshit
Long-term fear: truth drowning, bogus research much more voluminous than correct one
The ↑ Good
Smart harvesters and digesters
Huge opportunities for much-improved, flexible and highly-optimizable apps
Super simple example: IArxiv
Challenge for publishers: introduce next-generation digesters, search
Assessors
Machine being able to assess research works
- prose quality
- logical structure
- cohesion
- completeness
- correctness
- reproducibility
Accelerators / facilitators
- coding
- computing
- labwork
- data handling
- plotting
- writing (hum...)
New careers
harnesser / puppet master / detective
The Magic Sword of Muramasa
Setup of the joseki
The Magic Sword of Muramasa
The traditional joseki
The Magic Sword of Muramasa
AlphaGo's innovation
The question of original thought
Definition of "original":
- not secondary, derivative, or imitative
- independent and creative in thought or action
Go: machine is capable of "original" thought
Often read/heard:
AI can only rehash existing stuff, not invent new
I fundamentally disagree
The question of creativity
Nature, inanimate matter is creative: it invented us
AI is clearly a step above inanimate matter
All the current commotion is not really
about Machine Learning
In the end it's all about human learning,
what future it has,
and whether we give up, adapt or (need to) fight
Personal recommendations
Publishers:
- give clear guidelines
- don't assume they will be followed
- build AI into your search/discovery facilities
- discourage/penalize AI use in authoring
- resolutely fight the use of AI in writing reports
(writing reports is not the same as refereeing, it's only part of it) - innovate, with reason
- try to not f... things up (c.f. OA snafu)
Personal recommendations
Scientists:
- go easy on using AI for authoring (namely: don't)
- go wild on using AI for discovery/digestion
- using AI for writing referee reports?
Don't you dare - work hard to keep up with the tech
- be critical (now more than ever)
- cling to the driving seat
- understand instrumental convergence. Avoid it.