Tartu’s ethics centre listed 12 risks in school AI. The biggest one is not the technology.

The AI Leap Foundation commissioned an independent risk analysis before scaling its programme to tens of thousands more students. The working group was deliberately mixed: Margit Sutrop, who heads the University of Tartu Centre for Ethics, AI ethics specialist Katrin Laas-Mikko, Cybernetica research director Dan Bogdanov and senior researcher Liina Kamm, media studies professor Andra Siibak, and Henrik Trasberg, a legal and policy adviser at the European Commission’s AI Office.

They described twelve risks in three groups, and then said something that should travel beyond schools: the greatest risks do not come from the technology itself. They come from how it is used and from the choices made around it.

The three groups, read as a company

• ✳Educational risks: learning goals, the reliability of assessment, academic integrity, the role of the teacher. In a company these are the same questions about who is actually accountable for a judgement and whether your review process still measures anything.
• ✳Social risks: the Estonian language and culture, the values a model carries, and human interaction. Foreign models do not automatically reflect the local curriculum or context, and the same gap shows up when a model writes to your customers in a register that is not yours.
• ✳Technology and management risks: data breaches, rising costs, service interruptions, and dependence on an outside provider. This list is identical to the one your IT lead would write.

The sentence worth stealing

The analysis concludes that AI can strengthen Estonian education only if the use of technology is guided by educational goals, rather than the goals being reshaped to suit the technology. Swap two words and you have the test for any corporate rollout: is the tool serving the outcome you already cared about, or has the outcome quietly been redefined as whatever the tool is good at?

That second failure is common and hard to see from inside. A team adopts an assistant to write proposals faster, and within a quarter the measure of a good week has become the number of proposals sent. Nobody decided that. It happened because the tool made one number easy to move.

The analysis also warns that without deliberate intervention the strongest benefit most, while others lose ground on the thinking the work was supposed to build. In a company that shows up as a widening gap between the people who were already senior enough to check the output and the people who were not.

What to copy from their method

• ✳Write your risks down before the rollout widens, not after something goes wrong
• ✳Put someone outside the project on the review, because the people who built it cannot see its blind spots
• ✳Separate what the technology does from what your policy choices do, and be honest about which one you can actually change
• ✳State the outcome you are protecting, then check in three months whether you are still measuring that or something the tool made convenient

A state programme paid for an independent review before expanding. Most companies expand first and review after the complaint. The review is cheaper than the complaint.