10
Loading video...
By Jackie Peters
0
Share
Newsweek is a Trust Project member
See more of our trusted coverage when you search.Prefer Newsweek on Googleto see more of our trusted coverage when you search.
People should not have to understand an AI system’s design to expect reasonable protection for their financial information.
When a bank uses customer records to improve a service or detect fraud, the responsibility for limiting exposure belongs with the institution. I believe that responsibility should extend to the moment information is processed.
As organizations expand AI use, privacy commitments need more than written policies. Governance and employee judgment matter, but they should be supported by technical enforcement at the data layer.
Read More on Leadership Insights
Leaders should ask whether their systems actually restrict access to sensitive records, rather than simply instructing people to handle them carefully.
...
Imagine an AI assistant helping an employee resolve a payment query. If it can retrieve unrelated account details, even a routine request could expose information beyond its purpose.
This hypothetical scenario shows why a useful tool should receive only the access its task requires.
I would want that boundary built into the system, with approval requirements for sensitive actions and a defined point when access ends.
Employees should exercise judgment within those limits. They should not have to compensate for a design that gives an assistant unnecessarily broad permissions.
That scrutiny must include processing. Encryption at rest protects stored records, and encryption in transit protects them as they travel.
Neither alone establishes what happens when software works with the information. A protected database can still feed a process that makes sensitive details readable.
Conventional Processing
Historically, conventional processing has generally required encrypted information to be decrypted before analysis. Research into encryption in use is creating alternatives for certain tasks.
NIST describes cryptographic methods that allow computation on encrypted information. In simple terms, some systems can calculate a result without revealing the underlying records to the party performing the calculation.
That does not mean every application can operate this way at an acceptable cost or speed. It means executives should investigate what is feasible before treating exposure as unavoidable.
I see that investigation as part of responsible stewardship, alongside controls over who can request an analysis and receive its results.
A financial-services experiment illustrates the potential of protected collaboration. In September 2025, SWIFT reported trials involving 13 banks and 10 million artificial transactions.
Participants combined privacy-enhancing technologies with federated learning, which trains models locally at each institution rather than gathering customer records into one shared dataset.
According to SWIFT, the collaborative model was twice as effective at identifying known fraudulent transactions as a model trained on one institution’s dataset.
This was an experiment using synthetic data, not proof of equivalent results in live banking. It also does not establish that encryption in use alone produced the improvement.
Protection During Processing
For me, the useful lesson is that privacy protection deserves consideration as an enabler of analysis.
Institutions should explore whether safeguards can support appropriate collaboration before assuming their only choices are unrestricted sharing or abandoning useful work.
Regulatory standards also recognize protection during processing. Article 6 of the regulatory technical standards supporting the EU’s DORA framework addresses encryption of data in use where necessary, based on data classification and risk assessment.
Where that encryption is not possible, protected processing environments or equivalent measures are required. I read this qualified provision as a reason to examine processing risks carefully.
Technology cannot settle every question. Organizations still need to consider whether information should be collected and whether they have appropriate permission to use it, including when it has been de-identified.
Computing demands, integration, access permissions, and the handling of results also belong in the assessment.
The Need for Responsible AI
Leaders can fund these protections within already established budgets. Deferring them may transfer avoidable risks to customers.
Rejecting every sensitive-data application without evaluating safeguards may also mean missing opportunities to improve fraud detection or other worthwhile services.
Before approving such a project, executives should require a documented review identifying what becomes readable, who can access it, and which protections fit the task.
That review should include testing, the authority for using the information, a named accountable executive, and an explanation of remaining risks.
No approach eliminates every possibility of failure. But organizations can be expected to justify exposure instead of accepting it by default.
Responsible AI should make protecting information part of using it.
Jackie Peters is the co-founder and CEO of Blind Insight, a technology company that develops encrypted-data search and analytics tools for organizations handling sensitive information. The views expressed are the author's own.
View 0 Comments
Submit Correction
View Editorial & AI Guidelines