The growing adoption of Gen AI, esp. LLMs, has re-ignited the discussion around AI Regulations — to ensure that AI/ML systems are responsibly trained and deployed. Unfortunately, this effort is complicated by multiple governmental organizations and regulatory bodies releasing their own guidelines and policies with little to no agreement on the definition of terms.
In this talk, we will provide an overview explaining the key Responsible AI aspects: Explainability, Bias, and Accountability. We will then outline the Gen AI usage patterns and show how the three aspects can be integrated at different stages of the LLMOps (MLOps for LLM) pipeline. We summarize the learnings in the form of Gen AI design patterns that can be readily applied to enterprise use-cases.
2. AGENDA
¡ Enterprise AI
¡ Ethical / Responsible AI
¡ Explainability
¡ Fairness & Bias
¡ Accountability
¡ Generative AI Usage Patterns
¡ Generative AI - Responsible
Design Patterns
3. ENTERPRISE AI
Enterprise
AI
Natural Language
Processing (NLP)
Computer
Vision/Image
Processing
Predictive
Analytics
Natural Language
Search
Demand Forecasting
(Churn prediction)
Text Classification
Object Detection
Recommendations
Chatbots
(Dialog Systems)
Image Classification
Summarization
Optical Character
Recognition (OCR)
Predictive Maintenance
of Machines
• Enterprise AI use-
cases are pervasive
4. RESPONSIBLE AI
“Ethical AI, also known as responsible AI, is the practice of using AI with good intention
to empower employees and businesses, and fairly impact customers and society. Ethical
AI enables companies to engender trust and scale AI with confidence.” [1]
Failing to operationalize Ethical AI can not only expose enterprises to reputational,
regulatory, and legal risks; but also lead to wasted resources, inefficiencies in product
development, and even an inability to use data to train AI models. [2]
[1] R. Porter. Beyond the promise: implementing Ethical AI, 2020 (link)
[2] R. Blackman.A Practical Guide to Building Ethical AI, 2020 (link)
5. REGULATIONS
¡ Good news: is that there has been a recent trend
towards ensuring that AI applications are
responsibly trained and deployed, in line with the
enterprise strategy and policies.
¡ Bad news: Efforts have been complicated by
different governmental organizations and regulatory
bodies releasing their own guidelines and policies;
with little to no standardization on the definition of
terms.
¡ For example, the EU AI Act mandates a different set
of dos & don’ts depending on the ‘risk level’ of an AI
application. However, quantifying the risk level of an
AI application is easier said than done as it basically
requires you to classify how the capabilities of a
non-deterministic system will impact users and
systems who might interact with it in the future.
6. KEY RESPONSIBLE AI ASPECTS
¡ Explainability
¡ Bias & Fairness
¡ Accountability
¡ Reproducibility
¡ Data Privacy
*D. Biswas. Ethical AI: its implications for Enterprise AI Use-cases
and Governance.Towards Data Science (link)
*D. Biswas. Privacy Preserving Chatbot Conversations.
3rd IEEE AIKE 2020: 179-182
7. EXPLAINABLE AI
¡ Explainable AI is an umbrella term for a
range of tools, algorithms and methods;
which accompany AI model predictions
with explanations.
¡ Explainability of AI models ranks high
among the list of ‘non-functional’ AI
features to be considered by enterprises.
¡ For example, this implies having to
explain why an ML model profiled a
user to be in a specific segment —
which led him/her to receiving an
advertisement.
(Labeled)
Data
Train ML
Model
Predictions
Explanation
Model
Explainable
Predictions
8. EXPLAINABLE AI FRAMEWORKS - LIME
¡ Local Interpretable Model-
Agnostic Explanations
(LIME*) provides easy to
understand explanations of
a prediction by training an
explainability model based
on samples around a
prediction.
¡ The approximate nature of
the explainability model
might limit its usage for
compliance needs.
*M.T. Ribeiro, S. Singh, C. Guestrin. “Why Should ITrustYou?” Explaining the
Predictions of Any Classifier, 2016 (link)
LIME output showing the important features, positively
and negatively impacting the model’s prediction.
9. EXPLAINABLE AI - FEASIBILITY
¡ Machine (Deep) Learning algorithms vary
in the level of accuracy and explainability
that they can provide- the two are often
inversely proportional.
¡ Explainability starts becoming more
difficult as as we move to Random
Forests, which are basically an ensemble
of DecisionTrees.At the end of the
spectrum are Neural Networks (Deep
Learning), which have shown human-level
accuracy.
Explainability
Accuracy
Logistic Regression
Decision Trees
Random Forest
(Ensemble of
Decision Trees)
Deep Learning
(Neural Networks)
10. EXPLAINABLE AI - ABSTRACTION
“important thing is to explain the right thing to the right person in the right way at the right time”*
Singapore AI Governance framework:“technical explainability may not always be enlightening, esp. to the
man in the street… providing an individual with counterfactuals (such as “you would have been approved if
your average debt was 15% lower” or “these are users with similar profiles to yours that received a
different decision”) can be a powerful type of explanation”
*N. Xie, et. al. Explainable Deep Learning:A Field
Guide for the Uninitiated, 2020 (link)
AI Developer
Goal: ensure/improve
performance
Regulatory Bodies
Goal: Ensure compliance with legislation,
protect interests of constituents
End Users
Goal: Understanding of
decision, trust model output
11. FAIRNESS & BIAS
¡ Bias is a phenomenon that occurs when an algorithm
produces results that are systemically prejudiced due to
erroneous assumptions in the machine learning process*.
¡ AI models should behave in all fairness towards everyone,
without any bias. However, defining ‘fairness’ is easier said
than done.
¡ Does fairness mean, e.g., that the same proportion of
male and female applicants get high risk assessment
scores?
¡ Or that the same level of risk result in the same score
regardless of gender?
¡ (Impossible to fulfill both)
* SearchEnterprise AI. Machine
Learning bias (AI bias) (link)
Google Photo labeling pictures of a black Haitian-
American programmer as “gorilla”
“White Barack Obama” images
(link)
A computer program used for bail and
sentencing decisions was labeled biased against
blacks. (link)
12. TYPES OF BIAS
¡ Bias creeps into AI models, primarily due to
the inherent bias already present in the
training data. So the ‘data’ part of AI model
development is key to addressing bias.
¡ Historical Bias: arises due to historical
inequality of human decisions captured in
the training data
¡ Representation Bias: arises due to training
data that is not representative of the
actual population
¡ Ensure that training data is representative and
uniformly distributed over the target
population - with respect to the selected
features.
Source: H. Suresh, J.V. Guttag.A Framework for Understanding
Unintended Consequences of Machine Learning, 2020 (link)
13. ACCOUNTABILITY
¡ Similar to the debate on
self-driving cars with
respect to “who is
responsible” if an accident
happens?
¡ The same debate applies in
the case of AI models as
well — who is accountable
if something goes wrong?
Source:
https://www.theguardian.com/technology/2023/no
v/06/openai-chatgpt-customers-copyright-lawsuits
“If you are challenged on copyright grounds, we will
assume responsibility for the potential legal risks
involved,” the company said.
The move to protect customers from intellectual
property lawsuits comes after IBM Corp., Microsoft
Corp., and Adobe Inc. announced similar legal
protections for users of their AI products.
Source:
https://www.theverge.com/2023/10/12/2391499
8/google-copyright-indemnification-generative-ai
14. ACCOUNTABILITY CHECKLIST
• Data ownership: Data is critical to AI systems, as such negotiation of
ownership issues around not only training data, but input data, output
data, and other generated data is critical. For example, knowledge of the
prompts (user queries) and chatbot responses are very important to
improve the bot performance over time.
• Liability: Given that we are engaging with a 3rd party, to what extent are
they liable? This is tricky to negotiate and depends on the extent to which
the AI system can operate independently. For example, in the case of a
Chatbot, if the bot is allowed to provide only a limited output (e.g. respond
to the user with only limited number of pre-approved responses), then the
risk is likely to be a lot lower as compared to an open-ended bot like
ChatGPT that can generate new responses.
• Confidentiality clauses: In addition to (training) data confidentiality, do we
want to prevent the vendor from providing competitors with access to the
trained / fine-tuned model, or at least any improvements to it —
particularly if it is giving a competitive advantage?
15. GEN AI USAGE PATTERNS
*D. Biswas. MLOps for Compositional AI. NeurIPSWorkshop on Challenges in Deploying and
Monitoring Machine Learning Systems (DMML), 2022.
*D. Biswas. Generative AI – LLMOps Architecture Patterns. Data Driven Investor, 2023 (link)
¡ Black-box LLM APIs: This is the
classic ChatGPT example, where we
have black-box access to a LLM
API/UI. Prompts are the primary
interaction mechanism for such
scenarios.
¡ While Enterprise LLM Apps have the
potential to be a multi-billion dollar
marketplace and accelerate LLM
adoption by providing an enterprise
ready solution; the same caution
needs to be exercised as you would
do before using a 3rd party ML
model — validate LLM/training data
ownership, IP, liability clauses.
16. GEN AI USAGE PATTERNS – LLMOPS (MLOPS FOR LLMS)
*D. Biswas. Contextualizing Large Language Models (LLMs)
with Enterprise Data. Data Driven Investor, 2023 (link)
¡ LLMs are generic in nature.To
realize the full potential of
LLMs for Enterprises, they
need to be contextualized with
enterprise knowledge captured
in terms of documents, wikis,
business processes, etc.
¡ This is achieved by fine-tuning
a LLM with enterprise
knowledge / embeddings to
develop a context-specific
LLM.
Public data
Open Source Pre-
trained LLM
Data Processing
Pipelines
Knowledge Graphs /
Embeddings
(Vector Stores)
Enterprise
data
Supervised fine-tuning
/ Few-shot Learning
Context-specific LLM /
Small Language Model (SLM)
Mobile / Web UI
End user Apps
Prompts
Tasks /
Queries
Users
SLM API
Reinforcement Learning from
Human Feedback (RLHF)
Model
Monitoring
Model
Versioning
Model
Caching
Ethical AI Safeguards
17. GENERATIVE AI - RESPONSIBLE DESIGN PATTERNS
We take inspiration from the
“enterprise friendly”
Microsoft, “developer
friendly” Google and “user
friendly” Apple — to enable
this ‘transparent’ approach
to Gen AI system design.
• Guidelines for Human-AI
Interaction by Microsoft
• People + AI
Guidebook by Google
• Machine Learning:
Human Interface
Guidelines by Apple