Stage 4: tested with real people.
The rule that placed itA randomized trial or field study has tested it with real people doing the task.
The evidenceIn a randomized trial at a large US public university, access to a course-integrated AI tutor lowered final grades by 0.37 standard deviations in same-course comparisons. Liu et al., EdWorkingPaper 26-1598, Oct 2026.
A stage says how far the evidence has got. It is not a rating of the AI and not advice.
- 1ClaimedNo evidence yet
- 2LaunchedEvidence found
- 3Independently testedNo evidence yet
- 4Tested with real peoplePlaced here
- 5Regulator-clearedNo approval process exists
- 6Widely used with a verified recordNo evidence yet
Everything we found.
In ladder order, newest first within each kind. Makers' own claims are labeled as theirs.
What has launched
- ChatGPT study mode
ChatGPT study mode is available to the public.
OpenAI (company-reported) · Jul 29, 2025
Tested with real people
- Randomized trial, 2,379 undergraduates, 30 instructors
In a randomized trial at a large US public university, access to a course-integrated AI tutor lowered final grades by 0.37 standard deviations in same-course comparisons.
“tutor access reduced final grades by 0.37 standard deviations”
Liu et al., EdWorkingPaper 26-1598 · Oct 2026
- Randomized trial, 18 Tennessee middle schools, two school years
In a two-year randomized trial in 18 Tennessee middle schools, being assigned Khan Academy's Khanmigo tutor raised maths scores by about 0.06 to 0.08 standard deviations per term; students rarely used it for real maths dialogue.
“Assignment raises math achievement by 1.3 national percentile ranks per term, or about 0.06 to 0.08 standard deviations”
Oreopoulos and Low, EdWorkingPaper 26-1551 · Aug 2026
- Randomized trial, 194 Harvard students
In a randomized crossover trial in a Harvard physics course, students learned more in less time with a custom AI tutor than in an in-class active-learning lesson.
“students learn significantly more in less time when using the AI tutor”
Kestin et al., Scientific Reports · Jun 3, 2025
- Randomized trial, 900 tutors, 1,800 K-12 students
In a Stanford randomized trial, students whose human tutors had an AI helper (Tutor CoPilot) were 4 percentage points more likely to master maths topics.
“students working with tutors that have access to Tutor CoPilot are 4 percentage points (p.p.) more likely to master topics”
Wang et al., Stanford, arXiv 2410.03017 · Jan 26, 2025
- Randomized trial
In a six-week randomized program in Nigeria, students who used an AI tutor gained about 0.3 standard deviations in learning.
“The learning improvements were striking—about 0.3 standard deviations.”
World Bank · Jan 9, 2025
Not found in our sources yet: what the makers claim; what independent tests found; how many people use it. No regulator approves AI for this task, so the approval step doesn't apply.
Track record and tools.
Failures and court cases are shown beside the stage, never folded into it. Tools are listed because the evidence names them, in no order of merit.
- Khanmigo
- Tutor CoPilot
- ChatGPT study mode
Can AI do these yet?
Answer my health questions
Stage 4 of 6Tested with real people
Randomized trial: chatbot users made no better health decisions
Spot breast cancer on a mammogram
Stage 5 of 6Regulator-cleared
Transpara has been FDA-cleared since 2018, most recently in Aug 2026
Check a mole for skin cancer
Stage 5 of 6Regulator-cleared
EU Class III CE mark lets DERM clear benign cases alone (company)