Artificial Intelligence
26 papers and posts.
Cultural Bias in Language Models: The Top 10 Test
How can AI-generated "Top 10" lists of cultural influencers expose the cultural biases and defaults that are built into our most popular language models? I propose and report an initial test using lists of the Top 10 most influential musicians.
Golem.AI: The Experimenter's Regress and AI Decision Systems
How can the concept of the experimenter's regress help us understand the problems in AI decision systems? I explore two forms of circularity that can underpin AI decision tools through the lens of Collins and Pinch's 'The Golem'.
NotebookLM: Cast with Care
Google's NotebookLM "deep dive" feature is taking off in popularity. I subject three of my academic papers to the deep dive treatment, and reveal its tendency to subvert content for a happier ending.
Four Myths about Generative AI in Education
I outline four common misconceptions about the use of Generative AI which are widespread in Higher Education debates about the use of these tools: that it is possible and practical to detect the use of AI in writing, that text produced by GenAI is bland, repetitive or predictable, that GenAI tools struggle to cite sources accurately, and that more creative or reflective assessments are harder to complete using AI.
Cryptic AI
As language models are fine-tuned to acquire more capabilities, we continue to seek new tasks to push the limits of Generative AI. In setting cryptic crossword clues, so far GPT-4 fails the test quite spectacularly.
The AI Skills of Social Scientists
What skills do social scientists need to adapt to generative AI, and how should educators approach teaching them? A narrow focus on training everyone in technical skills is misguided - what's needed is authorial voice, leadership and management skills, and the critical force of the social sciences.
Leary Philosophers
Can a language model outperform old Edward Lear in describing philosophers in Limerick form? "There once was a Scotsman named Hume..."
'I did write that text': Ownership and Authorship Claims by Language Models
Will large language models acknowledge authorship of their own generated texts? Will a language model claim authorship or ownership of texts which it did not create? A mistaken comprehension of ChatGPT's abilities throws up a distinctive problem of intellectual property rights.
Perplexity, Creativity and the Zonkamoozle
What is the relationship between perplexity, creativity and novelty? Following on from 'Perplexing Perplexity', I set out to demonstrate that high perplexity texts are not always creative, and to showcase ChatGPT's ability to work with and even generate novel words - culminating in the Tale of Zonkamoozle.
Perplexing Perplexity
Detectors such as GPTZero use the property of 'perplexity' to try to detect AI authorship of texts. But I show that by writing in a specifically dull style, or engineering the prompt given to a language model, we can easily and systematically fool such detectors to label AI text as human and vice versa.
Machine Evidence II: The Abstract Setting
A recent study by Gao et al. (2022) validates the warning of 'Machine Evidence' (Blunt, 2019) that language models would soon become capable of beating detection attempts by human peer reviewers. This piece looks at the near-term steps that journal editors and conference organisers can take to prevent AI-generated abstracts bypassing their screening processes, along with a warning for the long-term viability of those strategies.
To the Tune of Pure Reason
Can a new iteration of GPT-3 write pop songs, raps and limericks...about Immanuel Kant's Categorical Imperative? It is a moral duty to find out.
'Modern' Philosophers by DALL·E 2
Can DALL·E 2 create images of ancient philosophers like Plato, Aristotle and Immanuel Kant as they'd look in modern day universities? Sort of. Should it? Definitely not.
Bias and the Myth of the Objective Average
Would you choose a black box AI surgeon with a 90% success rate over a human surgeon with 80% success? The answer exposes a fundamental and harmful assumption within dominant models of medical evidence.
Higher Orders of Evidence
DALLE 2 offers a far more powerful image generation AI than the popular open access 'Craiyon'/'DALLE Mini' model. How does DALLE 2 compare to DALLE Mini's visions of hierarchies and pyramids of evidence?
Visions of Evidence
How does a machine learning algorithm picture hierarchies of evidence and evidence-based medicine - and what do these visions of evidence remind us of the way we understand, order and assemble the information we use to guide clinical practice?
The Jurassic Critique of Micozzi on Evidence Hierarchies
AI21 Labs have just released a public demo of their giant language model, Jurassic-1. At 178bn parameters, it rivals GPT-3. Feeding it my own work, it generated some interesting and potentially novel views on evidence hierarchies... and then attributed them to CAM researcher Marc Micozzi! Is Jurassic Micozzi's critique of evidential pluralism in medicine sound?
Imitating Imitation: a response to Floridi & Chiriatti
In their 2020 paper, Floridi and Chiriatti subject giant language model GPT-3 to three tests: mathematical, semantic and ethical. I show that these tests are misconfigured to prove the points Floridi and Chiriatti are trying to make. We should attend to how such giant language models function to understand both their responses to questions and the ethical and societal impacts.
The Stochastic Masquerade and the Streisand Effect
What does Google have in common with Barbra Streisand? Since Google fired AI ethicists Margaret Mitchell and Timnit Gebru, our attention should turn to what they don't want us to read: "On Stochastic Parrots". Will the attempts to suppress this paper lead to it being overlooked, or will Google face Barbra Streisand's fate?
Socrates in the Dungeon
What happens when you ask a machine learning language model tuned to create a D&D style adventure to instead produce a Socratic dialogue?
Automatic Gadfly: Socrates by Machine
I was inspired to see if AI language model GPT-2 could create some thought-provoking - or just weird - new Socratic dialogues.
Dual Use Technology and GPT-3
Yesterday, AI researchers published a new paper entitled Language Models are Few-Shot Learners. This paper introduces GPT-3 (Generative Pretrained Transformer 3), the follow-up to last year's GPT-2, which at the time it was released was the largest language model out there....
Aphorisms from the Automatic Philosopher
GPT-2 is a large language model capable of generating some of the most convincingly human-like text we are yet to see from artificial intelligence. Previously, I've used GPT-2 to generate reports of clinical trials and several paragraphs of an essay on irrationality in which...
Automatic Philosophising
The following philosophical musings were generated by a machine learning language model called GPT-2. They were created by a very weak version of GPT-2 which was released to the public in May 2019. The full version has nearly 5 times as many nodes and produces much more...
All that Glitters is not... Evidence
Recently, I wrote about a new machine learning model called GPT-2 which was conspicuously not released by OpenAI. GPT-2 is a massive language model which can be used to generate often highly convincing text when given a prompt. Using the 'attention' framework, the language...
Machine Evidence: Trial by AI
Take a look at the following snippets from descriptions of clinical trials, thinking about how you'd rate the quality and strength of the evidence that comes from each: