Lecture 22 Ethics

How we ought to live — Socrates

• what is ethics
  • What is the right thing to do?
  • Why?
• Why Should We Care?
  • AI technology is increasingly being deployed to real-world applications
  • Have real and tangible impact to people
  • Whose responsibility is it when things go bad?
• Why Is Ethics Hard?
  • Often no objective truth, unlike sciences
  • A new philosophy student may ask whether fundamental ethical theories such as utilitarianism(实用主义，功利主义) is right
  • But unlikely a new physics student would question the laws of thermodynamics(热力学)
  • In examining a problem, we need to think from different perspectives to justify our reasons
• Learning Outcomes
  • Think more about the application you build
    • Not just its performance
    • Its social context
    • Its impact to other people
    • Unintended harms
  • Be a socially-responsible scientist or engineer

Arguments Against Ethical Checks in NLP

• Should We Censor(审查) Science?
  • A common argument when ethical checks or processes are introduced:
    • Should there be limits to scientific research? Is it right to censor research?
  • Ethical procedures are common in other fields: medicine, biology, psychology, anthropology, etc
  • In the past, this isn’t common in computer science
  • But this doesn’t mean this shouldn’t change
  • Technology are increasingly being integrated into society; the research we do nowadays are likely to be more deployed than 20 years ago
• H5N1
  • Ron Fouchier, a Dutch virologist, discovered how to make bird flu potentially more harmful in 2011
  • Dutch government objected to publishing the research
  • Raised a lot of discussions and concerns
  • National policies enacted
• Isn’t Transparency Always Better?
  • Is it always better to publish sensitive research publicly?
  • Argument: worse if they are done underground
  • If goal is to raise awareness, scientific publication isn’t the only way
    • Could work with media to raise awareness
    • Doesn’t require exposing the technique
• AI vs. Cybersecurity
  • Exposing vulnerability publicly is desirable in cyber-security applications
    • Easy for developer to fix the problem
  • But the same logic doesn’t always apply for AI
    • Not easy to fix, once the technology is out

Core NLP Ethics Concepts

• Bias
  • Two definitions:
    • Value-neutral meaning in ML
    • Normative meaning in socio-cultural studies
  • Ethics research in NLP: harmful prejudices in models
  • A biased model is one that performs unfavourably against certain groups of users
    • typically based on demographic features such as gender or ethnicity
  • Bias isn’t necessarily bad
    • Guide the model to make informed decisions in the absence of more information
    • Truly unbiased system = system that makes random decisions
    • Bad when overwhelms evidence, or perpetuates harmful stereotypes
  • Bias can arise from data, annotations, representations, models, or research design
• Bias in Word Embeddings
  • Word Analogy (lecture 10):
    • v(man) - v(woman) = v(king) - v(queen)
  • But!
    • v(man) - v(woman) = v(programmer) - v(homemaker)
    • v(father) - v(mother) = v(doctor) - v(nurse)
  • Word embeddings reflect and amplify gender stereotypes in society
  • Lots of work done to reduce bias in word embeddings
• Dual Use
  • Every technology has a primary use, and unintended secondary consequences
    • uclear power, knives, electricity
    • could be abused for things they are not originally designed to do.
  • Since we do not know how people will use it, we need to be aware of this duality
• OpenAI GPT-2
  • OpenAI developed GPT-2, a large language model trained on massive web data
  • Kickstarted the pretrained model paradigm in NLP
    • Fine-tune pretrained models on downstream tasks (BERT lecture 11)
  • GPT-2 also has amazing generation capability
    • Can be easily fine-tuned to generate fake news, create propaganda
  • Pretrained GPT-2 models released in stages over 9 months, starting with smaller models
  • Collaborated with various organisations to study social implications of very large language models over this time
  • OpenAI’s effort is commendable, but this is voluntary
  • Further raises questions about self-regulation
• Privacy
  • Often conflated with anonymity
  • Privacy means nobody know I am doing something
  • Anonymity means everyone know what I am doing, but not that it is me
• GDPR
  • Regulation on data privacy in EU
  • Also addresses transfer of personal data
  • Aim to give individuals control over their personal data
  • Organisations that process EU citizen’s personal data are subjected to it
  • Organisations need to anonymise data so that people cannot be identified
  • But we have technology to de-identify author attributes
• AOL Search Data Leak
  • In 2006, AOL released anonymised search logs of users
  • Log contained sufficient information to de-identify individuals
    • Through cross-referencing with phonebook listing an individual was identified
  • Lawsuit filed against AOL

Group Discussion

• Prompts
  • Primary use: does it promote harm or social good?
  • Bias?
  • Dual use concerns?
  • Privacy concerns? What sorts of data does it use?
  • Other questions to consider:
    • Can it be weaponised against populations (e.g. facial recognition, location tracking)?
    • Does it fit people into simple categories (e.g. gender and sexual orientation)?
    • Does it create alternate sets of reality (e.g. fake news)?
• Automatic Prison Term Prediction
  • A model that predicts the prison sentence of an individual based on court documents
    • bias: black people often give harsher sentence
    • non-explainable if use deep learning
    • is it reasonable to use AI to judge one’s freedom in the first place?
• Automatic CV Processing
  • A model that processes CV/resumes for a job to automatically filter candidates for interview
    • bias towards gender(stereotypes, does the model amplify the stereotype?)
    • the system can be cheated
    • how the data is sourced(privacy)
    • the reason for rejection? deep learning is black box
• Language Community Classification
  • A text classification tool that distinguishes LGBTQ from heterosexual language
  • Motivation: to understand how language used in the LGBTQ community differs from heterosexual community
    • dual use: potentially can be used to classify LGBTQ person, discriminate people.
• Take Away
  • Think about the applications you build
  • Be open-minded: ask questions, discuss with others
  • NLP tasks aren’t always just technical problems
  • Remember that the application we build could change someone else’s life
  • We should strive to be a socially responsible engineer/scientist