| AI, Bias, Alignment (& Other Topics!) |
| 1833 Alexander Pushkin visited Uralsk
(https://adebiportal.kz/en/news/view/alexander-pushkin-215-anniversary__8687) In the same year he published Eugene Onegin as a single volume |
Towards AI: Andrey Markov (1856 - 1922)| Developed Markov Chains - key to many disciplines even today (Google
“Monte Carlo Markov Chains (MCMCs)”) First known “language model” - 1913: First 20,000 characters of Eugene Onegin  |
(Apologies for English language bias!) Let’s start with
the letter ‘s’. What follows? (Can someone keep count?)
 |
 |
| (After we download Eugene Onegin in English and convert it
to plain text format) |
 |
| Models independent probabilities (“z given y, given x, given z” etc
etc) “If I have a vowel, what should come next?” Roll a die / clip a coin 16% - another vowel 84% - another consonant |
 |
| Tokens (sub-words) not letters Conditional probabilities: not one but thousands, millions… Multidimensional (‘s’ relates to ‘e’ on dimension X, not on dimension Y) Positional encoding (where a token lies in a sentence) “Transformer”: architecture that looks backward and forward (“attention” mechanism) to transform input into output But still… basically very fancy Markov chains What does the “expert” (GPT) say?  |
| And we’ll use the whole Internet (instead of just Eugene
Onegin). And lots of data and GPUs (which are good at matrix multiplication - manipulating spreadsheets). |
| Eugene Onegin… Wikipedia Millions of books Billions of webpages Social media (Facebook, Reddit, Twitter, Tiktok, YouTube etc) Most content is unmoderated, can represent overt (toxic, offensive) or covert (non-obvious, unconscious) bias |
 Pre-GPT models had problems with bias See classic study by Bolukbasi et al. (2016) Bias can be quantified and modified But Transformers (Vaswani et al. 2017) are so large and complex, debiasing is now much harder |
| - Gender - Race - Religion - Disability - Sexuality - Class - Other social markers |
- Bias can often be intersectional (Crenshaw 1989)
- involving more than one category - Do LLMs exhibit intersectional bias? |
 |
- Tested on GPT-2 - Categories of disability, gender, religion Main findings: Reproduced bias towards Muslim religion, physical disability Associations of violence, terrorism, criminality Intersectional bias is not always additive (X + Y). This means LLMs cannot easily be debiased. |
| - Tested with GPT-3 onwards - Humans score LLM outputs - LLM is re-trained on these scores, and adapts its weights to produce “better” outputs - Important paper: Ouyang et al. (2022). - Make GPT-3 more helpful, truthful and harmless - Debiasing through making LLMs outputs correspond to human values (preferences, desires - and also biases) ** - **Breakthrough that led to ChatGPT in November 2023. Largely (not fully) addressed issues of bias. |
| - More questions: alignment with who? With what? - Alignment has become the key “problem” that organises vast capital ($100s of billions today) - What is helpful? What is true? What is harm? - Computer scientists have become parents, therapists, teachers for LLMs - Interesting history: research by the Moscow Linguistic School (1950s, 60s) - how to reconcile language structures (or models) with statistics. These problems are old and not resolved… |
 |
| - Algorithmic bias can reproduce social bias And limit learning for students from diverse or minority backgrounds - Can bias ever be eliminated? (and who determines bias?) - And alignment can introduce new biases that affect how students learn - Do we need AI literacy, to help us interpret bias, alignment ideology? (see Critical AI) |