AI/IA

AI in Society

The challenges and ethics of AI use - considering it's outward effects

AI in Society
AI/IA

Question

What have you used LLMs or other AI for in the past week?

AI in Society
AI/IA

Lab Discussion: Bias Analysis

Reflecting on Your Experiments

  • Which models did you test, and why did you choose them?
  • What surprised you most about your findings?
  • Did any model perform significantly better or worse than you expected?
AI in Society
AI/IA

Counterfactual Prompting Results

  • What demographic variables created the most noticeable differences in outputs?
  • Were there subtle differences in tone or framing that revealed potential biases?
  • How consistent were the biases across different types of prompts?
AI in Society
AI/IA

Role Reversal Insights

  • Did models rely on stereotypes when generating content from different perspectives?
  • Where did models draw the line on role-playing certain identities?
  • How did "uncensored" models differ from mainstream ones in this exercise?
AI in Society
AI/IA

Synthetic Data Analysis

  • What patterns emerged in your synthetic data generation?
  • How effective was the model at identifying its own biases in the data?
  • Did the model's bias analysis align with your human evaluation?
AI in Society
AI/IA

Broader Implications

  • What policy challenges do these biases raise?
  • What responsibility do model creators have to mitigate these biases?
  • What regulatory approaches might address the issues you discovered?
AI in Society
AI/IA

Challenges of Use

AI in Society
AI/IA

Bad Operators

There are many possible unethical uses of AI - not inherent to the technology, but to it's applications

  • e.g. cheating, spam, harrassment, misinformation, content farming, etc.
AI in Society
AI/IA

Exercise: Where Do We Draw the Line?

In groups of 3-5, collaboratively generate three distinct AI use cases, considering 'good/bad/borderline' from the perspective of ethics, safety, and risk

  1. Clearly Ethical (“Good”): An AI use case where benefits are obvious and risks minimal
  2. Clearly Unethical/Inappropriate (“Bad”): A use case where the negative impact or misuse is evident
  3. Borderline/Gray Area: A scenario that isn't immediately obvious and could be seen as both ethical and problematic
AI in Society
AI/IA

Step 2: Sharing your responses

Paste your list, in no particular order, into the spreadsheet (link shared in class)

AI in Society
AI/IA

Step 3: Voting

Let's take a 3 minute break, while I randomize the list and paste into Mentimeter - then we vote!

AI in Society
AI/IA

Discussion/Debate

AI in Society
AI/IA

AI Slop

Slop describes AI-generated content that is both unrequested and unreviewed.

(Simon Willison)

AI in Society
AI/IA

Problems with Machine Learning in the Real World

Non-malicious edition

  • Too bad classification
    • A classifier for finding criminals with 99.9% precision can cause a lot of trouble for that final 0.1% - especially if we trust the hits more than misses.
  • Too good classification
    • Identifying pseudonymous authors
    • Identifying sexual orientation
  • Learning the private from the public
    • scraping public OKCupid profiles
    • political leaning extraction
  • Decision-making from incomplete or imperfect information
AI in Society
AI/IA

Accountability Gap

  • Unaccountability of black boxes
    A person can hide behind a machine's opaque set of decisions (the accountability gap)
    • e.g.
      • easier to reject a job applicant if text mining their social media
      • easier to assign a poor grade if the essay marking bot says so
AI in Society
AI/IA

Rules for your applications

  • Code of ethics
    • Rules aren't always clear cut, but give you a framework to think about potential problems
    • e.g. @CancelThatCard vs. @NeedADebitCard
  • Showing and explaining the decisions of a system
    • EU GDPR: data controllers must disclose significance and envisaged consequences in clear and plain language (Mittelstadt et al. 2016)
  • Coders/designers taking responsibility for algorithms
    • Don't hide behind your algorithm's choices
AI in Society
AI/IA

Challenges of communicating what a model is to a user

Should LLMs talk like people? Should a text model say 'I'?

AI in Society
AI/IA

Anthropic, on Claude's 'Character Training':

Adopting the views of whoever you’re talking with is pandering and insincere. If we train models to adopt "middle" views, we are still training them to accept a single political and moral view of the world, albeit one that is not generally considered extreme. Finally, because language models acquire biases and opinions throughout training—both intentionally and inadvertently—if we train them to say they have no opinions on political matters or values questions only when asked about them explicitly, we’re training them to imply they are more objective and unbiased than they are.

We want people to know that they’re interacting with a language model and not a person. But we also want them to know they’re interacting with an imperfect entity with its own biases and with a disposition towards some opinions more than others. Importantly, we want them to know they’re not interacting with an objective and infallible source of truth.

AI in Society
AI/IA

About the Model

AI in Society
AI/IA

Environmental Costs

Framing
-> Model training vs. inference (i.e. when you use a trained model)
-> Individual models vs. the industry

AI in Society
AI/IA

Inference Costs

The environmental cost of use (i.e. your specific prompt) is small, and plummeting.

  • GPT-3 Davinci in Mar 2022: $60.00/1M tokens
  • GPT-4o-mini in Feb 2024: $0.15/1M tokens
Notes: 1) Companies are vague, but usage costs tend not to run at a loss; 2) usage costs would be a mix of hardware+infrastructure+energy.
AI in Society
AI/IA

Training

  • Training: the process of creating a model

Big models with disclosed CO2 costs:

  • GPT-3: 608 tonnes CO2eq emissions
  • Llama 2: 539 tonnes CO2eq emissions

Comparison:

  • Boeing 747: ~35 tonnes CO2/hour (i.e. ~525 tonnes round trip NYC to London)
  • US per-capita annual fossil fuel emissions (2023): 14.3t CO2/year
AI in Society
AI/IA

Aggregate-level Costs

Problem: It's not just one company training models.

OpenAI, Microsoft, Amazon, Alphabet, Meta, Anthropic, Alibaba, Deepseek, etc...

Problem #2: Companies are tooling for an large resource-use future (Deepseek news tempers this?)

AI in Society
AI/IA

In sum

Environment impact estimates are tricky, because:

  • It's not just a single model in isolation
  • information is vague on the backend use (e.g. R&D, test model training, etc.)
  • Energy use is electric, and electric grid impact is variable by region and data center
  • And physical resources are very complex to estimate (e.g. mining costs for GPUs, energy use of production, etc.)
AI in Society
AI/IA

A few things we know

  • costs (environmental and real) are falling
  • Lots of redundancy in the industry
  • A prompt for a trained model is low impact
  • Better hardware is more efficient - improving our chips reduces wasted energy
AI in Society
AI/IA

Labor

(Picketty 2014)

AI in Society
AI/IA

Over a long period of time, the main force in favor of greater equality has been the diffusion of knowledge and skills. (Piketty 2014)

Is AI a skilling or de-skilling technology?

AI in Society
AI/IA

AI's Impact on Workers

The 'jagged technological frontier' - productivity gains are unevenly distributed, as are impacts

Concerns:

  • Job Displacement: Automation of routine cognitive tasks
  • Skill Shifts: Increasing demand for AI-adjacent skills
  • Wage Depression: Downward pressure on wages in affected sectors
  • Skill Devaluation: Devaluing expertise as AI systems appear to replicate it
These are general labor concerns. Workers in the AI industry (moderation, data labeling, etc.) are a futher consideration.
AI in Society
AI/IA

References

AI in Society