← All education signals
Education

The Tethered Gain

Generative AI reliably lifts student performance while it’s switched on, but the OECD’s 2026 review finds that lift can vanish, or reverse, once the tool is taken away.

What is it?

Generative AI lifts student performance while it is switched on, but the gain does not always outlast the tool. The OECD finds performance can rise without learning, and 2026 preprint trials suggest that using AI to explain, not to write, is what keeps a gain in place.

Why now?

The OECD’s 2026 Outlook put the Türkiye result at the centre: students who had used general-purpose GPT-4 did worse in exams without it. Three randomised preprints followed in 2026, testing gains after removal, in schools and with adults, and all of them await peer review.

Who’s building it?

AI tutor makers, education researchers running rapid trials, the OECD, and school leaders choosing tools.

AI tutor makersTrial researchersOECDMedlySchool leaders
The full read
Why now, who is building it, where the market goes, and what stays unsettled.
Get weekly →
01 · The market
Who’s in it
Builders — tutor makers designing tools that explain rather than answer.
Buyers — schools and families choosing tools whose gains may not survive switching off.
Gatekeepers — the OECD and researchers setting the standard for proof.
What’s at stake
Whether a learning gain counts only when it survives the tool being switched off, not while it runs.
Whether general-purpose chatbots belong in study time at all.
Whether preprint trials are enough to buy on.
What’s unsettled
Whether any school gain persists once the AI tutor is removed.
Whether adult gap-closing results tell us anything about pupils.
Whether a four-week trial with 30.7% attrition can support a buying decision.
02 · How the demand matureda progression, not a timeline
Novelty
Students use chatbots for homework, and scores rise while they run.
Retention
Exams taken without the tool show some of that rise was never learning.
Willingness to pay
Schools weigh purpose-built tutors with a trial effect, such as g = 0.33, against chatbots.
Category
Learning tools are judged on what a student can still do after switching off.
03 · Where the market goesthe forks, not a prediction
→The retention test: trials routinely measure students after the AI is removed, and tools that only lift assisted scores lose their claim.
→Purpose-built tutors pull ahead, as general-purpose chatbots are kept out of study time.
→Preprint results harden into peer-reviewed evidence, and the split between augmentation and automation holds.
→Assisted scores stay the headline, and the gain after removal goes unmeasured.
The forks turn on whether schools and funders start asking what a student can do once the AI is gone, not while it is on.
04 · The Retention deskadjacent market, moving now
The Tutor Nobody Opens
Take-up, not efficacy, limits AI tutors.
Only 15 percent with Khanmigo access regularly engage with it.
EdTech Innovation Hub
Cognitive Offloading
Handing thinking to tools weakens the skills beneath.
AI essay writers showed weaker brain connectivity, an MIT preprint found.
MIT Media Lab
The Always-On Tutor
A tutor available at any hour reshapes study.
Khanmigo’s K-12 users grew from 40,000 to 700,000.
K-12 Dive
05 · The Wire
01
The OECD finds AI raises performance without matching learning gains.
OECD ↗
02
The OECD’s director says purpose-built tools show more promise.
OECD ↗
03
Preprint: AI-assisted students kept a smaller test gain a week later
arXiv (Contractor, Reyes) ↗
04
A four-week school trial of an AI tutor in GCSE sciences finds g = 0.33, with 30.7% attrition.
arXiv (Harrison et al.) ↗
05
Among adults on a workplace task, AI shrank an education gap that partly re-emerged once it was removed.
arXiv (Cruces et al.) ↗
The panel on this signal
They agreed on
–How a student uses the AI, to explain or to produce, matters more than whether they have it at all.
–The strongest 2026 trials are still preprints, not peer-reviewed.
–No school trial yet shows whether a gain outlasts the AI.
The cognitive psychologist disagreed on
–Curriculum-aligned tests built for one trial cannot show transfer; retention claims need a standardised measure taken without the tool.
Know more about the method →
✦How this signal was built
01 · Panel
A Delphi panel of 4 AI analysts, each reasoning independently.
→
02 · Debate
They challenge and revise each other’s reads over two rounds.
→
03 · Evidence
Every claim is tied to a cited, linked source on The Wire.
→
04 · Synthesis
A moderator writes it into one read, disagreement and all.