Hello

there!

Max Kmet

Senior Data Scientist at MacPaw

MacPaw — is a software development company that develops and distributes software for macOS and iOS.

​

MacPaw is the maker behind CleanMyMac X, Setapp, ClearVPN, and other products.�Today, MacPaw products have more than 30 million users worldwide. Every fifth Mac on the planet has�at least one MacPaw app.�MacPaw comes from Ukraine and is proud of it.

Why people are afraid of AI and how to stop it

​

Risks, regulations

and how to get prepared

Why are people scared of AI?

Will I lose my job to AI?

Will I lose my job to AI?

NY Times Jobs Protections AI

Reuters Jobs at high risk from AI

Will I lose my job to AI?

WEForum Automation Changes Future of Work

WEForum Jobs AI can’t replace

• By 2025, automation will displace�83 million jobs globally�
• The robot revolution will create�69 million new jobs

Jobs AI can’t replace

WEForum Jobs AI can’t replace

Jobs AI will create

WEForum Jobs AI Will Create

• Trainers
• Engineers developing AI tools
• Electrical engineers developing microchips�
• Explainers
• design the interfaces that enable people to interact with AI�
• Sustainers
• content creators (e.g. prompt engineers)
• data curators
• and ethics and governance specialists

Can software developers

be replaced by AI already?

Arxiv: Is Stack Overflow Obsolete?

Deepfakes are getting too real

Deepfakes are getting too real

CNN Deepfake CFO scam

Deepfakes are getting too real

WEForum AI deepfakes legalisation

How can we protect?

• Deep fake detection systems
• Watermarks for AI generated content
• Require that AI models refuse generate harmful content
• Zero-trust mindset

AI may treat people unfairly

AI may treat people unfairly

Reuters: Amazon AI recruiting tool

Racial Bias Credit Scores

Hospitals Racial Bias

If AI makes a mistake, who do I blame?

If AI makes a mistake,

who do I blame?

Judge Sanctions Lawyers for brief written with AI

AI is a black box

How can I trust it?

How can I trust a black box?

Twitter ylecun

How can I trust a black box?

EU AI Act

Parliament’s priority is to make sure that AI systems used in the EU are:

• safe
• transparent
• traceable
• non-discriminatory
• environmentally friendly

Who will save us from the evils of AI?

Regulations

EU AI Act

Enforcement could include fines of up to €30 million or 6 percent of global revenue

AI regulation tracker

MCKinsey AI EU Act

US AI Regulations

State by state AI regulations

US AI Regulations

Bloomberg Law: NY AI use in hiring law

• NYC employers using AI to hire must undergo bias audit�
• Employers must notify job candidates when tools used

How to prepare�for regulations?

Foundation Models Transparency Index Scores

​

by Stanford Center for Research on Foundation Models

Arxiv: Foundation Models Transparency Index Scores

Total 100 indicators. The top-scoring model scores only 54 out of 100

Foundation Models Transparency Index Scores

Arxiv: Foundation Models Transparency Index Scores

Model Cards

Arxiv: Model Cards

Model Card LLama-2-7b

Huggingface: Model Card Llama-2-7b

Model Card LLama-2-7b

Arxiv: Foundation Models Transparency Index Scores

What AI developers should remember?

Where could it

go wrong?

An Experimental Framework for Measuring Model Fairness and Stability in the Machine Learning Lifecycle

Areas of responsible AI

Explainability

Privacy

Security

Robustness

Stability

Fairness

Fairness

Fairness – absence of systematic discrimination against individuals or groups

Fairness metrics�

• False Positive Rate (FPR), False Negative Rate (FNR) for different population groups
• Disparate Impact – Inspired by the 4/5th’s rule in legal doctrines
• Statistical Parity Difference – asks if comparable proportions�of samples from each protected group receive the positive outcome

Disparate Impact:

Equalized Odds Violation�(False Positive):

Model Stability

Model Stability – property when small changes in the input of the model�lead to proportionally small changes in the output.

Model Stability metrics�

• Label Stability – metric used to measure the level of disagreement between the estimators in the ensemble for a particular test sample
• Standard deviation (STD) / Interquartile Range (IQR) of predicted probabilities�of the ensemble on an individual sample from a test set

Approaches to explainable AI

Interpreting Image Classification Model with LIME

Interpretation of image classification model

with LIME

• LIME
• SHAP values
• Partial Dependence Plot (PDP)
• Individual Conditional Expectation (ICE)
• Analyze forking in tree-based models
• Coefficients of linear model
• Permutation-based methods

Copyright issues

Copyright issues

Copyright issues

Bloomberg: Reddit signs AI Content Licensing deal

Copyright issues

Reuters: Artists take new shot at Stability and Midjourney

Stable Diffusion Copyright Issues

• Stable Diffusion trained on images generated by Midjourney�
• According to Midjourney’s terms of service, paying users hold all rights to the images they generate

Copyright issues

OpenAI Terms of Use

OpenAI terms of Use:

​

What You Cannot Do:�You may not use our Services for any illegal, harmful, or abusive activity.��For example, you may not:

Use Output to develop models�that compete with OpenAI.

Summary

🤔Why people are scared of AI�

• Loss of jobs to AI
• Deep fakes
• Unfair treatment
• AI makes mistakes

🔧Regulations�

• AI EU act
• US AI Regulations: anti-bias law in NY

🔍Preparation�

• Foundation Model
• Transparency Index
• Model Cards
• Approaches to responsible AI
• Methods for explainable AI

⚠️Copyright issues�

• OpenAI/Stable Diffusion/Midjourney

Q&A

Contact me 👋🏼

LinkedIn: maxkmet