On your own
Answer for your team.
About five minutes. Every role gets its own questions, about the work you do.
- The findings, sent to you before they are published
- A 20-minute conversation about your answers, if you want one
The AI Oversight Study 2026
How organizations check, monitor and trust the automated systems they run.
It takes about five minutes. Every team is working this out on its own, and together the answers show what none of them can see alone. Everyone who takes part sees the findings first.
Why this study
Every team that runs an automated system has answers to a few plain questions.
- How did you check it before the last change?
- What do you check it against?
- Do your own experts agree with each other?
- How would you find out if it failed?
Each team answers them on its own, with no way to tell whether its answers are ordinary, careful or thin. One team’s answers are an anecdote. Answers from across industries and roles show a pattern: what teams check before a change goes live, what they compare it against, how they learn it failed, and what gets a system signed off.
By automated systems we mean any software that decides, scores, flags or generates something without a person doing each one: AI, machine learning or plain rules.
Two ways to take part
With your organization
See where your organization stands.
Run the study across your teams with a private link, and learn how your organization compares with others in its industry and of its size.
- One private link, which you share with your teams
- The people who build, review and sign off, each compared with their peers
- The findings first, and a conversation about your results
Answer it yourself, and ask for your organization’s link at the end. Within two working days, we send the link and a short brief to share with it.
Or write to
Who it’s for
The people who build, run, review or sign off on automated systems.
- Software engineering
- ML or data science
- Data or analytics
- Product management or design
- Marketing, content or communications
- Fraud, AML or investigations
- Risk, compliance or model governance
- Executive or general management
Answer about your team, not your whole organization. If you work across several teams, pick one and answer for it.
The survey
Take the survey
There are no right answers. The most useful ones are the honest ones, including “we don’t check that yet.”
Every free-text question is optional, and you can stop at any time. By answering, you accept the terms of participation. The Privacy Policy explains how we handle answers.
Who’s asking
One question keeps us up: how do you know it’s right?
At U22A8 we build models for AI judgment and evaluations. Before anyone relies on one of them, we have to answer that question. So does every team that runs an automated system.
We want to learn how teams answer it today, from the people who do the work. So we’re asking, and we publish what we find.
Taras Yanchynskyy Machine learning engineer, and payments before that
oversight
noun
- Watchful, responsible care over something.
- Something missed by mistake.
This study is about the first, so there are fewer of the second.
Know someone who should answer? Send them to oversight.study