Information Integrity: AI Amplifies Dangerous Data

(picture from pexels.com)

By Richard Geiger

Fundraisers have been instructed the identical factor about their information for 20 years: Keep it clear, preserve it present, preserve it full. It was good recommendation then, supplied principally within the context of avoiding a bounced enchantment letter or a red-faced apology to a serious donor whose identify bought misspelled for the third 12 months operating. 

It is important recommendation now for a a lot bigger purpose. Artificial Intelligence (AI) doesn’t repair dangerous information. It amplifies it.

A current nonprofit AI benchmark research from Virtuous discovered that 92% of organizations are already utilizing AI in some kind, but solely 7% report the form of main, strategy-level impression that adjustments what a group can really get performed. Most are caught in what researchers known as an “effectivity plateau” — a quicker first draft right here, a faster analysis pull there, nothing that provides as much as actual transformation. A separate 2026 survey of nonprofits with AI already in manufacturing by Coastal put a finer level on why: 72% or organizations are combating information accuracy or availability as soon as the instruments go reside. Ambition outran roads.

There will not be loads of love for the time period information hygiene. It suggests a once-a-year deep clear, one thing you schedule after which neglect about. What AI requires is nearer to information integrity as an ongoing self-discipline —  correct contact and giving data, constant formatting throughout each system that touches a constituent, de-duplicated households, present appended demographic and wealth indicators, and one reliable model of who somebody is and what your relationship with them appears to be like like. 

A 2026 survey on information integrity and AI readiness by the LeBow College of Business at Drexel University discovered organizations with actual information governance applications report excessive belief of their information at a fee greater than 20 factors larger than these with out governance in any respect, which tells you the hole will not be actually in regards to the AI in any respect.

A predictive mannequin constructed on duplicate data will double-count your greatest donors and dilute your worst ones. A generative software drafting outreach from an incomplete constituent historical past will write confidently to somebody who now not needs to listen to from you or tackle them by the flawed identify fully. AI doesn’t know it’s working from dangerous inputs. It simply produces dangerous outputs with glorious grammar and complete conviction, which is a extra harmful failure mode than the methods it changed, as a result of it appears to be like credible.

This work begins in an unglamorous place: Standardizing how data enter your methods: Consistent identify and tackle codecs; required fields on the level of entry; and, an outlined course of for merging duplicates as an alternative of letting them pile up. This course of accounts for a lot of the final result. Once that basis holds, the following step is that improvement, advertising and marketing, applications, and finance information often reside in methods that by no means discuss to one another. Resolving id throughout the entire constituent base earlier than layering AI onto any single one in all them means the mannequin you construct displays one donor, not 4 fragmented variations of the identical individual. 

Organizations extending current information governance to cowl AI, slightly than standing up a separate AI coverage, are these with notably differentiated performances. The piece usually missed is measuring the output and confirming its accuracy and authenticity afterward. Using AI will not be the purpose. Match and deliverability charges holding regular, fewer guide corrections flowing again from employees, mannequin suggestions that reward officers can act on as an alternative of quietly overriding — that’s what success appears to be like like. 

If your group spends extra time fixing AI outputs than trusting them, the information beneath must be adjusted earlier than the software does.

Many conversations land on the identical objection: We don’t have a knowledge group or the finances for one. That’s truthful. But information integrity work scales down higher than most expertise investments do. Start with the phase that issues most financially (recurring donors, main reward prospects, no matter drives the majority of proceeds), and get that inhabitants’s data right, up to date and arranged earlier than repeating throughout the entire file.

Lean on shared and cooperative information sources slightly than constructing id decision capabilities from scratch. The sector has spent a long time constructing that roads, so particular person organizations don’t have to resolve it alone. Resist the strain to purchase an AI software earlier than wanting onerous at what you’ll feed it, sensible use circumstances, desired outcomes, and so on. The most cost-effective AI venture is the one you by no means must redo a 12 months later as a result of it was constructed on a shaky basis.

Funders have began asking nonprofit leaders how they handle information as a part of due diligence on AI use, not as an afterthought. Sector analysis by AI and You this 12 months discovered that almost 40% of nonprofits report having no employees formally skilled in AI, and an analogous group report no formal AI governance coverage in any respect. 

Closing that hole within the subsequent 12 months or two will matter greater than any single AI pilot a group rolls out as a result of none of these pilots can be value a lot if the information feeding them can’t be trusted. That is the quieter competitors underway proper now, nicely under the headlines about AI.

*****

Richard Geiger is senior vice chairman at Data Axle Nonprofit. His electronic mail is [email protected]