[ SUBJECTS ]

AI help for algebra that keeps students solving

In algebra, answer-giving AI can do quiet damage: the homework looks finished while the skill never forms. How Nexus's class modes keep students doing the steps.

Plan a pilot Watch the demo

Facts as of Oct 18, 2026Last reviewed Oct 19, 20264 min read

48%

higher practice scores with answer-giving AI while students used it

Source: Bastani et al., PNAS, 2025

17%

lower unaided exam scores after the AI was removed

Source: Bastani et al., PNAS, 2025

[ ALGEBRA ]

Where AI helps, and where it gets in the way

The clearest evidence on AI and learning comes from a math classroom. In a field experiment with about 1,000 high-school students, answer-giving AI raised practice scores 48% while students could use it, then cut unaided exam scores by 17%. A hint-first version of the same AI removed that harm, but it did not beat no AI. Students using the answer-giving version also overrated how much they had learned.

So the question for an algebra class is not whether AI is allowed. It is whether the student or the AI does the next step. Worked examples help when they are used well: across 55 studies in math, worked examples gave g = 0.48, and correct examples worked best. The lesson is a correct example, then a similar problem done alone.

How the three learning modes work in algebra

Modes describe how a class is set up to respond, and students cannot change a class's mode. AI can make mistakes; before any pilot we check Hints and Guided with real models.

Mode	What the student gets	Good for	Watch for
Hints	A question about the next step, such as "What happens if you distribute the 3?"	Practicing a new skill, such as solving multi-step equations	Students asking for hint after hint until the answer appears
Guided	One step at a time; the student performs each step before the next	Systems of equations, factoring, word problems	Students waiting for the next step instead of trying it
Full answer with critique	A complete worked solution, with a critique of the method	Checking a finished problem set; studying a model solution after an attempt	Answer-first use, which is closest to the harmful condition in the research

Learning modes in algebra

A hint-first exchange

Study assistantHints

StudentCan you just tell me what x is? 3(x - 2) = 2x + 5

Study assistantLet's get you there. What's the first thing you'd do with 3(x - 2) on the left side?

StudentMultiply it out? 3x - 2?

Study assistantClose. The 3 multiplies both terms inside the parentheses. What is 3 times -2?

Student-6. So 3x - 6 = 2x + 5.

Study assistantYes. Now there are x terms on both sides. What could you do to get them on one side?

StudentSubtract 2x from both sides, so x - 6 = 5. Then add 6. x = 11?

Study assistantCheck it yourself: put 11 into both sides of the original equation. Do they match?

Illustration: written by the Nexus team to show how a Hints class is designed to respond. It is not a transcript of Nexus output. Real answers vary by model, and AI can make mistakes.

Class instructions teachers can add

Each class can carry the teacher's own instructions on top of its mode. A few that work well here:

"Ask for the student's next step before giving another hint."
"If a student asks for the answer, ask them to show their last step first."
"Use the notation from our textbook, and never simplify for the student."
"When a student has an answer, have them check it by substituting."

Checking what students can do alone

Assisted work is not evidence of a skill. Give a short unaided check on new items a few days after practice, then again later: classroom quizzing improves learning (g = 0.50 across 222 studies), and spaced retrieval beats massed study (g = 0.74 across 29 studies). In a Nexus pilot, the primary measure is delayed unaided success: practiced concepts solved alone on new items, 7 days and 2-4 weeks later.

Students also see up to three next steps per class, such as "Try one on your own" after a run of hints on a concept, with the counts and dates behind each.

Accuracy in math

AI can make math mistakes, and hints can be wrong too. In one study of ChatGPT-3.5 hints for algebra and statistics, 32% of the raw hints failed quality checks; checked hints helped adult learners about as much as human-written ones. Ask students to verify answers by substitution, and review a sample of class chats each week.

[ FAQ ]

Questions and answers

Is it cheating to use AI for algebra homework?

That depends on the class rule. In a Hints class, AI asks about the next step instead of giving the answer, so the student still does the algebra. The research concern is answer-giving AI, which raised practice scores but lowered unaided exam scores.