# Behavior Modeling, Secondary AI Effects, Bias Reduction & Synthetic Data

Devansh Devansh, Stealth AI Startup | MLOps Podcast | Episode 311 | 1:01:36
Hosted by Demetrios Brinkmann

Source: https://www.youtube.com/watch?v=jJXee5rMtHI
Channel: MLOps Community, now AAIF Live (https://www.youtube.com/@AAIFLive-x1r). Summarised by MLOps Talks.
Page: https://mlopstalks.com/talks/behavior-modeling-secondary-ai-effects-bias-reduction-synthetic-data
Published: 2025-05-02
Tags: agents, multi-agent, synthetic-data, workflows

## TL;DR
- Devansh Devansh argues that AI systems should be designed by first deciding which outcomes and social behaviors are worth encouraging, then working backwards through the data and infrastructure needed to produce them.
- Synthetic data and automated decision systems can reduce some forms of bias when they are designed to remove sensitive signals and prevent models from relying on demographic proxies.
- Businesses usually get better results from deterministic workflows with limited LLM steps than from generalized autonomous agents, whose scope and accountability are difficult to control.

## Summary
Devansh Devansh discusses AI as a way to inspect and restructure data, rather than only as a model trained on data. He sees synthetic data as useful when it preserves privacy, adds diversity, or makes sensitive datasets easier to share, while warning that unrealistic distributions can damage model performance. He argues that teams should begin with the experience or social outcome they want, then work backwards to decide what behaviors to model, what data to collect, and how to structure it. The conversation also covers bias reduction in loan decisions, where automated systems reportedly produced more diverse approvals by avoiding demographic proxies. Devansh is skeptical of generalized autonomous agents in business. He prefers deterministic workflows with narrowly scoped model calls. The discussion ends with the hidden effects of automation, using hotel doormen, customer service, hiring, and AI-generated content as examples of benefits that disappear when a narrow task replaces a wider human role. He also urges people to challenge the power structures forming around AI rather than accepting the roles assigned to them.

## Key ideas
### Language models can inspect the biases and missing information in their own training data
[04:48](https://www.youtube.com/watch?v=jJXee5rMtHI&t=288s)
Devansh Devansh was excited by the possibility of using language models as a reverse transparency tool. A model could help people ask what a dataset is missing, which behaviors training on it would encourage, and how valuable it is for a particular task. This changes the usual direction of AI work. Instead of treating data only as material for training a model, teams can use AI to find themes, correlations, and patterns in their existing data. He compared this with people uploading Instagram feeds or LinkedIn profiles to ChatGPT and asking what the data reveals about them. The same pattern can support more serious analysis of business and social datasets.

### Synthetic data is useful when it protects privacy and adds the right kind of diversity
[11:30](https://www.youtube.com/watch?v=jJXee5rMtHI&t=690s)
Discussing Gretel AI, Devansh Devansh described synthetic data as more than a way to generate plausible records. He mentioned privacy-preserving substitutions, including redacting personal information, as part of the data pipeline. He warned that generated data must be balanced carefully. Data that is too close to an existing distribution may fail to add useful variation. In computer vision experiments, he said simple random augmentation sometimes performed best because it introduced more diversity. When datasets are small, realistic probability distributions may matter more. The hard part is deciding what kind of variation supports the intended behavior instead of simply producing more data.

### AI projects should start with desired outcomes and work backwards through the data
[14:33](https://www.youtube.com/watch?v=jJXee5rMtHI&t=873s)
Devansh Devansh pushed the conversation beyond choosing a model for a business task. He asked what problems are worth solving as a species and what society values from AI. From there, teams can decide which type of AI would support those goals, what behaviors should be encouraged, and whether existing data needs to be reshaped, augmented, or collected from scratch. He argued that failing to make these choices leaves the conversation to groups with vested interests. The result is likely to be another productivity application, sales tool, or email integration because those are the outcomes that influential groups are already building toward.

### Automated decisions can reduce bias when models are prevented from using demographic proxies
[22:20](https://www.youtube.com/watch?v=jJXee5rMtHI&t=1340s)
Demetrios Brinkmann asked about reports that synthetic data and automated systems had reduced bias in loan or fraud-related models. Devansh Devansh explained that some automated loan approval systems produced more diverse lending outcomes than institutions with low racial diversity in their historical lending. A model may infer race without receiving race as an explicit feature. Location, financial information, and other variables can act as proxies, especially in places shaped by racial segregation. Devansh described the positive example as a case where the system was designed to avoid those signals and focus on whether applicants deserved loans. He connected this work to synthetic data methods such as differential privacy, which can remove sensitive information while preserving useful structure.

### The data layer can carry intelligence through relationships, behavior, and structure
[27:04](https://www.youtube.com/watch?v=jJXee5rMtHI&t=1624s)
Devansh Devansh argued that the industry has concentrated too much intelligence in the model layer. Graphs offer one example of moving intelligence into the data layer because nodes, relationships, and weights make relational structure explicit. The same source data can produce different knowledge graphs depending on which relationships are prioritized, and that choice affects performance. He also described behavior modeling as a possible form of data-layer intelligence. Systems could observe how users act, build models of those actions, and infer evaluation functions from what a person values. At a larger scale, tracking behavior across GitHub users might reveal patterns that precede problems, such as feature-branch divergence and difficult merges.

### Workflow-based systems limit model responsibility and are easier to improve
[35:00](https://www.youtube.com/watch?v=jJXee5rMtHI&t=2100s)
Devansh Devansh described work in legal AI where improvements come from mapping the full process instead of repeatedly labeling inputs as good or bad. Contract review can be broken into individual components, with models used at specific points in a larger workflow. He said this approach propagates useful signals through complex knowledge work better than simply feeding examples into a model. Demetrios Brinkmann summarized the design principle as giving the model the least responsibility possible and surrounding it with systems and workflows. Devansh agreed that this is a better basis for many production systems than an open-ended agent.

### Generalized autonomous agents are usually a poor fit for business processes
[38:16](https://www.youtube.com/watch?v=jJXee5rMtHI&t=2296s)
Devansh Devansh called autonomous agents largely a shame in business settings because their sci-fi framing can defer accountability. He said that every business example he had seen eventually became more deterministic. He sees value in simple workflows where a repeated task has one or more LLM calls, such as scraping and summarizing a site or creating a task. Generalized HR or finance agents create trouble because users cannot know whether a failure came from the prompt, the task itself, or the system's limits. A master agent that selects among bounded workflows may be useful, but an agent that can be asked to do almost anything is difficult to run with predictable accuracy.

### Automation removes hidden benefits when it focuses only on the stated task
[42:52](https://www.youtube.com/watch?v=jJXee5rMtHI&t=2572s)
Devansh Devansh used the hotel doorman example to explain secondary effects. An automated door can replace the doorman's stated task of opening and closing the door, while removing greeting guests, helping with luggage, discouraging non-guests from staying in the lobby, and helping people find transport. He applied the same reasoning to AI customer service, hiring, and generated content. A system may improve a visible metric such as cost, traffic, or output volume while damaging trust, brand value, candidate experience, or customer loyalty. He said teams should examine what users value in the old process before replacing it with an automated one.

### People should choose consciously whether they want to shape AI or be shaped by it
[58:47](https://www.youtube.com/watch?v=jJXee5rMtHI&t=3527s)
Devansh Devansh argued that AI as a social institution is still relatively new, which leaves room for people outside mathematics and software engineering to influence it. He encouraged listeners to ask whether they are trying to change existing power structures or simply hoping to be selected by them. He did not condemn people who prefer stability or have other priorities. His concern is that many people accept the categories of expert and non-expert without examining them, becoming viewers of society rather than participants. The practical choice, in his view, is to decide deliberately whether to accept those roles or challenge them.

## Notable quotes
- Devansh Devansh: "I think there is so much potential in this space when we start turning this back on ourselves, start trying to think about what kind of behaviors do we want to encourage in a model?" (13:33)
- Devansh Devansh: "If you can move intelligence into the data layer, you can do a lot of work in so many domains in so many ways." (29:38)
- Devansh Devansh: "I think autonomous agents are largely a shame." (38:16)
- Devansh Devansh: "The more simplistically you look at a problem, a situation, an outcome, the more you're likely to miss critical pieces that end up unraveling it." (52:11)
- Devansh Devansh: "If you're going to make that decision, make it consciously as opposed to just kind of making it because you've never thought about the alternative." (1:01:16)

## Tools & references mentioned
- DeepSeek
- Gretel AI
- Nvidia
- ChatGPT
- GPT-3.5
- o1
- Arcade.dev
- 11x
- Claude
- Harvey
- LinkedIn
- GitHub
- Twitter
- Threads

## Who should watch
- You are designing an AI product and need to decide whether a model belongs inside a bounded workflow or inside a generalized agent.
- Your team is generating synthetic data, building knowledge graphs, or assessing whether existing data contains demographic proxies and unwanted behaviors.
- You are measuring an automation project by one business metric and want to examine the customer, employee, or social effects that metric may miss.

## Related talks

- [Unbearable Lightness of Data: The Weight of Information](https://mlopstalks.com/talks/unbearable-lightness-of-data-the-weight-of-information) (Rohit Krishnan, Bodo.AI, 54:17)
- [AI Operations Without Fundamental Engineering Discipline](https://mlopstalks.com/talks/ai-operations-without-fundamental-engineering-discipline) (Nikhil Suresh, Hermit Tech, 49:28)
- [AI-Powered Product Ideation with Synthetic Consumer Testing](https://mlopstalks.com/talks/ai-powered-product-ideation-with-synthetic-consumer-testing) (Luca Fiaschi, PyMC Labs, 1:00:44)
- [How Agentic Workflows Will Change Everything](https://mlopstalks.com/talks/how-agentic-workflows-will-change-everything) (Raj Rikhy, Microsoft, 49:13)
- [Beyond the Matrix: AI and the Future of Human Creativity](https://mlopstalks.com/talks/beyond-the-matrix-ai-and-the-future-of-human-creativity) (Fausto Albers, AI Builders Club, 55:09)