Trust in Data: Earned in Drops, But Lost in Buckets

At a time when algorithms influence everything from healthcare to benefit payments to policing, public trust in data systems is more than a technical concern – it’s a matter of life and death.

“Ten years ago, I used to live in the Netherlands,” explained Fawad Qureshi, global field CTO at Snowflake. “There was a famous scandal at the Dutch tax authority. A black box AI model accused 35,000 families of childcare fraud and sent them notices that they needed to immediately repay €100,000.

The model was not only wrong – it was catastrophically so. Thousands were falsely accused.

“Tragically, people ended up committing suicide, because somebody in that agency thought, ‘We don’t need a human in the loop. It’s a black box model. Just send it out. Should be right,” said Qureshi.

To that end, Qureshi has created a five-point framework, aimed at rebuilding and preserving public trust in data systems.

1. Transparency Leads to Trust

The foundation of trust, said Qureshi, begins with transparency. “Your model needs to be explainable. You can’t use ignorance as an excuse. Black box models are not a defence in court.”

This isn’t just a matter of technical preference – it’s a legal requirement under regulations like GDPR and the UK’s ICO standards. Citizens have a right to explanation, and institutions can no longer hide behind algorithms they don’t fully understand.

“Gone are the days you can say, ‘I did not know how this model worked. I gave it XYZ, and I got ABC.’ That doesn’t work anymore,” said Qureshi.

Qureshi also shared a checklist from a commercial client for evaluating ethical transparency:

1. Would I be okay doing this with my own children’s data?
2. Would I feel comfortable presenting this use case in public?
3. Would I pass the ‘Daily Mail test’? — i.e., would I be fine with a tabloid headline about this tomorrow?

“If the answer to all three is yes, you’re probably doing it right,” he said.

2. Clarity Beats Complexity

“We have a lot of complex economic policy decisions that we need to make, and we need to make them understandable,” said Qureshi.

Complex economic policy must be distilled into digestible, intuitive insights. “We’ve got the shape of the data,” he said. “Most of the data that is coming out from the government is in raw wheat grains. But we need to transform it, from raw dough to frozen pizza, to something someone can actually eat.”

3. Design for the Public, Not Just Policymakers

Data systems must serve the broader public – not just civil servants or elected officials. That means creating tools that work for the average citizen.

“The average person on the street doesn’t know what a credit card with 19% APR means,” said Qureshi. “We need to simplify those data products so they’re understandable, not just for policymakers, but for the people making real decisions based on them.”

Designing inclusive data products means using plain language, offering context, and anticipating varying levels of data literacy, he said.

4. View the Solution in Totality – and Admit the Gaps

No agency operates in a vacuum. Public policy is a web of interconnected systems, and decisions made in one department can have consequences across society.

“We need to think about the whole solution,” said Qureshi. “Are all the agencies connected? Are we considering the negative externalities?”

Qureshi referenced Freakonomics, where researchers linked Roe v. Wade – which established a woman’s constitutional right to an abortion – to a lowering in long-term crime rate trends in the U.S., as a reminder of the far-reaching consequences of social policy.

Building robust national data libraries that connect across departments and regions could be a game-changer, he added. “We need to create systems where data is sharable, connected, and capable of supporting broad, societal-scale insight.”

5. Consistency Creates Credibility

Trust is not a one-off achievement – it’s a continuous process.

“You can’t produce a data product once and then go quiet for two years. It needs to be a constant stream,” said Qureshi. “The public needs to know they can rely on that data being there tomorrow, next week, next year.”

He likened data reliability to an API – something people can plug into and expect steady, predictable updates.

“Trust is earned in drops,” he concluded. “But lost in buckets.”