Synthetic Data Has a Place, but It Is Not the Bias Panacea

Episode Summary

Speaking in February 2025, Matt Maccaux, head of customer engineering at Google Cloud, is asked what you do when the data you would train on carries the bias you are trying to remove. His answer starts with a warning rather than a method, because generating your way out of it may create problems of its own. Synthetic data has a real use, he says, for conditions someone has already thought of, and it stops working at the edge of what anyone thought to imagine, which he illustrates with a self-driving car and a horse on a highway. What he trusts more is people. Bringing humans into the loop while you fine-tune is, he thinks, the better route, though he is candid that it introduces a bias of its own. His practical advice is to start from the good data you already have and ask open-ended questions of it.

Key takeaways

  • Ethical AI is not a philosophical topic for his clients, it is an early practical one. He says it is in the first few questions he gets asked about a new AI use case, and what they want is guidance on setting up an ethical AI practice, reducing hallucinations and keeping outcomes from discriminating

  • His worked example of bias is a hiring one, and the mechanism is the point. An organization that has hired software engineers looks at the workforce it already has and uses that as the base for filtering incoming resumes, so the demographics of the existing team become the filter

  • The uncomfortable part is that the reasoning sounds fine. We have good software engineers, so should we not hire more engineers who look like that, and his answer is the yes-but, which is where he says the ethics actually live. What organizations want is to catch it early enough to avoid bad PR or bad business outcomes

  • Asked whether you fix biased data by generating new data, he opens with a warning rather than a method, and says the approach could create additional problems of its own

  • He walks through the options and finds each one costly. Eliminate a category such as an education background, filter it somehow, or build an equivalence that a degree from one university is worth a degree from another, which the college rankings may not agree with and which he calls a lot of work

  • What he trusts more than generated data is people, and he says so with a hedge rather than a claim. If we cannot trust the automation or the creation of synthetic data, then perhaps we just have to introduce lots of humans to nudge these models in the direction we want them to go

  • He immediately turns that on itself. Have we now created and introduced our own bias unintentionally, and will that narrowly drive the outcomes we are going for. Probably, he says, and only the organization can know whether that is good or bad

  • His rule for when synthetic data works is about what you already know. It has a use when you are trying to create known knowns or known unknowns, meaning situations someone has thought of and can describe

  • His rule for when it fails is the mirror of that, and the illustration is a self-driving one. A truck on the highway pulling a trailer with a horse in it: the car has been trained to brake for a horse on the highway, but should it brake here, and nobody knew that was even a thing

  • The conclusion he draws is about who does the work rather than which technique wins. It is going to take a human to define the conditions under which the problem exists, and there is a time and a place for synthetic data, but he would not say it is the panacea for removing all the concerns around bias

  • On what enterprises are actually doing, he reports no pattern yet. It was early enough in February 2025, he says, that he did not detect patterns across his clients, and most of them are working with the good data sets they already have, where good means high quality with metadata and policies about who may use it

  • His argument against reaching for public data is commercial rather than technical. If you can use a publicly available data set to nail your business problem, chances are someone else has already nailed it, which is why it is publicly available

  • The first reason AI stalls before production, in his account, is that the decision may have nowhere to sit. Maybe they do not have executive sign-off, he says, because perhaps they do not have an AI governance council in place, so they do the process, procedure and cultural legwork first, which he concedes sounds odd in an industry that usually fires first and aims later

  • The second reason is money, and it has two parts. The budget may never have been allocated for compute and specialized GPUs, which he called constrained and outrageously expensive in early 2025, and some organizations are not exactly sure how much AI will make or save them, so the ROI calculation stays open

  • That is why he sees so many companies starting inside their own walls. Productivity use cases point the model at your own data, keep it internal and make employees more productive at writing code, documenting it or searching a knowledge base, and the business case gets built on that

  • His worked ROI example is deliberately small and built on invented round numbers. Ten software developers producing a hundred features a year, made twenty percent more productive, gives you either twenty percent more features or the same features with fewer people, and either way the number is measurable

  • Asked whether companies with no technical debt reach production faster, his answer is unambiguous but qualified. The few digital natives he works with move significantly faster than the traditional enterprises, and what holds them back is not capability but budget, since computational cost is real and funding is a trade-off of a dollar here against a dollar there

  • His explanation for why traditional enterprises move slowly is about reputation rather than technology. They have reputations that are based on trust, so a feature that does not work, introduces bias or hallucinates is a real problem and can erode the customer base, while a digital native audience, with exceptions he flags himself, tolerates frequent releases that sometimes break

  • His first instruction to an executive early in the journey is to look before buying. Go and look at your existing data assets, peel back the onion to see what you have, and ask open-ended questions of it, because if your data scientists can answer them you probably have good data and the internal expertise to unlock it

  • His second instruction comes with a warning most vendors leave out. Start with productivity use cases in your development teams, and expect a dip in productivity and velocity as the tools go in, with the increase arriving a software release or two later

  • His closing advice is to leave your own industry. Go and talk to someone with a similar job profile in a completely different one, and bring your problems to that conversation candidly and honestly; conferences and meetup groups are where he says it works, and he sees it work in CIO forums

About Matt Maccaux

Matt Maccaux is head of customer engineering at Google Cloud, where he works largely with legacy large enterprises on how they adopt AI. On this February 2025 episode he says ethical AI is among the first few questions his clients raise, and that what most of them mean by it is eliminating bias before it produces bad PR or bad business outcomes. His position on synthetic data is that it has a genuine use for known knowns and known unknowns and is not the panacea for bias, and that bringing humans into the loop during fine-tuning is perhaps the better route, with the caveat that it substitutes one kind of bias for another. Asked whether companies without technical debt reach production faster, he says the few digital natives he works with move significantly faster, that budget rather than capability is what holds them back, and that traditional enterprises move carefully because their reputations rest on trust.

 

In this episode

00:36 Welcome, and who Matt Maccaux is
01:19 The question: the ethical use of data in AI
01:28 The first questions clients ask about ethical AI
02:21 Eliminating bias, and the job interviewing case
02:46 Filtering resumes against the workforce you already have
03:05 The yes-but, and catching it before the bad PR
03:35 The question: what do you do about biased data
03:56 Writing the answer while you are taking the exam
04:10 Eliminate the variable, or build an equivalence
05:05 Humans in the loop, and the bias that introduces
05:28 Known knowns and known unknowns
05:51 The great unknown, and the horse on the highway
06:26 A human defines the conditions, and it is not the panacea
06:43 The question: is there a typical approach
06:54 No pattern yet across his clients
07:21 Starting from the good data you already have
07:42 Public data sets, and where synthetic data comes in
08:02 If a public data set nails it, someone already has
08:38 A chaos monkey, and where synthetic data belongs
09:03 The question: why AI stays in proof of concept
09:25 Maybe no governance council, so maybe no sign-off
09:58 Budget, GPUs, and an open ROI calculation
10:30 Productivity use cases, pointed at your own data
11:02 Ten developers, and twenty percent more productive
12:15 Legacy large enterprises are who he works with
12:34 The question: do digital natives move faster
13:05 They do, and budget is what holds them back
13:22 A dollar here against a dollar there
14:04 No technical debt means velocity, and reputations built on trust
14:23 What breaks that trust, and how a customer base erodes
15:12 The question: guidance for executives early on
15:22 Look at your data assets and ask good questions
15:56 Start with productivity, and expect a dip first
16:42 Look internally, draw the line to ROI, make the ask
16:49 The question: what resources to recommend
17:23 Talk to your peers, join groups, go to conferences
17:49 Talk to someone in a completely different industry
18:28 What each category can learn from the other

In Matt’s words

“are we writing the answer to the exam question while we’re taking the exam?”

Matt Maccaux   (03:56)

“then perhaps we just have to introduce lots of humans to help nudge these models in the direction we want them to go”

Matt Maccaux   (05:05)

“I wouldn’t say it’s gonna be the panacea to remove all of the concerns around bias in AI today”

Matt Maccaux   (06:26)

“if you can use a publicly available data set to nail your business problem, chances are someone else has already nailed it”

Matt Maccaux   (08:02)

“the lack of technical debt means the pace, the velocity is so much faster”

Matt Maccaux   (14:04)

“go talk to someone that has a similar job profile in a completely different industry”

Matt Maccaux   (17:49)

 

Resources

Named on air

  • AI Realized: The conference he names at 17:23 as an example of where to go and learn from people outside your own company

  • CIO forums and meetup groups: The two settings he says at 17:49 he has seen work, for practitioners bringing real problems to peers

  • The self-driving horse: His illustration at 05:51 of the limit of synthetic data, and of why a human has to define the conditions rather than generating data for every case

  • The chaos monkey: The industry term he reaches for at 08:38, saying you can throw almost like a chaos monkey at your data to see what breaks, as one of the ways synthetic data gets used

Ideas and terms discussed

  • Synthetic data: Generated rather than collected data. His position is that it has a real use for known knowns and known unknowns, a second use where an organization has little data of its own, and no claim to being the panacea for bias

  • Known knowns and known unknowns: His test at 05:28 for when generating data works. If someone has thought of the condition and can describe it, synthetic data can cover it

  • Humans in the loop: His preferred alternative at 05:05. People nudging a model during fine-tuning, with the acknowledged cost that they bring a bias of their own

  • Productivity use cases: Internal, pointed at your own data, aimed at making employees faster at code, documentation and search. His recommended starting point and the basis for the ROI example

  • Technical debt: Digital natives lack it and move faster; traditional enterprises carry it along with the data and the customer trust that come with age

Related AI Realized episodes and events

 

Frequently Asked Questions

 
 
 
 
 
 
 
 
 
 
 
Previous
Previous

Use the AI Tools Yourself, Then Show Your Colleagues

Next
Next

Use AI for the Questions Nobody in the Room Will Ask