1 of 9

��Dr Sonal Srivastava��Truth and Trust in the Age of AI�Theory-based hybrid models for reliable decision making

Sonal Srivastava

CJBS

2 of 9

Sonal Srivastava

CJBS

Machine Learning and Predictions

Data

​

{purchase history, browsing data, demographics, etc.}

Machine Learning Algorithm

Predictions

​

{will buy/won’t buy, clickthrough rate, etc.}

True because mathematically robust, accurate predictions based on inputs

3 of 9

Sonal Srivastava

CJBS

Structural models

Data

​

Goals

DECISION-MAKING

PROCESS

Preferences

Beliefs

Optimal Choice

Economic Theory

Insights

​

​

  • Explainable models grounded in economic theory (e.g., consumers maximize utility, firms maximize profit)

​

  • Interpretable parameters (e.g., preferences, price sensitivity, brand loyalty, etc.)

​

  • Enable counterfactual analysis (e.g., how consumers respond if a new product is introduced)

4 of 9

Sonal Srivastava

CJBS

Truth and trust

ML Truth

Structural Truth

  • ML fast and scalable, fits data, but millions of hidden parameters

​

  • Structural modelling slow, challenging to scale, but transparent

​

  • Performance vs. trust

George Box (statistician)�1919-2013

“All models are wrong but some are useful”

5 of 9

Sonal Srivastava

CJBS

What I do?

Scalable, fast

Interpretable,

Integrable

Structural Models

Machine Learning

Hybrid Models

=

+

  • Use structural models and estimate them using Reinforcement Learning (RL)

​

  • Models that can be trusted and used for strategic decision-making but also scalable

Explainable

6 of 9

Sonal Srivastava

CJBS

Why theory is making a comeback in AI

  • Not enough data to train AI agents, need a way to generate new data

​

  • Theory-based models can be used to generate reliable simulations

​

Sora (OpenAI)

Structure in Video Generation

Physics + Deep Learning + NLP + RL

AlphaFold (Google DeepMind)

Structure in Biology

Biochemistry + Deep Learning

What next?

​

  • Theory-based economic agents can simulate markets, pricing, and consumer decisions

​

  • Enables safe, data-efficient experimentation before real-world rollout��

7 of 9

Sonal Srivastava

CJBS

When truth starts to decay: AI Slop

  • Marketing and research insights are increasingly generated by AI tools (e.g., ChatGPT, Copilot)

​

  • Sources often blurred (AI tools summarize, remix, and sometimes invent data)

​

  • AI-generated content on web (“AI Slop”) is now feeding back into new training dataset

​

Raw Data on Web

AI

AI Slop added to Web

8 of 9

Sonal Srivastava

CJBS

Rebuilding Truth

Curated, real human generated data

Prioritize

Behavioural (revealed) and

Survey (stated) data

Combine/Validate

Data origin and

Modelling assumptions

Document

9 of 9

Sonal Srivastava

CJBS

Aligning Data, Truth, and Trust

Data

Insights

Model

  • Data quality underpins truth

​

  • Model transparency and robustness builds trust

​

  • Together, they create reliable insights