It’s a question that feels like it should have a simple answer. Either AI helps students, or it hurts them. Pick a side.

But a 2025 study published by Anthropic – the company behind the AI tool, Claude  – found something more complicated. Whether AI helps or harms your learning depends almost entirely on how you use it.

And most students are using it in the way that hurts most.

What the research found

The study, “The Impact of AI Assistance on Learning and Performance,” recruited 276 participants across a range of tasks and tested whether AI assistance improved performance and what happened to their understanding of the material afterward.

The findings were striking. Participants who used AI to complete their tasks performed worse on follow-up assessments than those who didn’t use AI at all. And the gap wasn’t small: those who used AI scored about 17 percentage points lower on comprehension questions taken immediately after the task, which is the difference between a B and a D (Anthropic, 2026).

The knowledge gap was consistent across academic levels. Whether participants were newer to the subject or more experienced, the pattern held.

A separate analysis of junior software engineers – highlighted in a Two Minute Papers video by Dr. Károly Zsolnai-Féhér that drew attention to this research – found a similar result. Engineers who used AI for coding tasks finished slightly faster but scored about 17 percentage points lower on a comprehension quiz than those who worked without AI. The largest gap showed up in debugging: the engineers who relied on AI the most struggled most when something went wrong and they had to fix it themselves (Two Minute Papers, 2026).

Why it happens

The study points to a mechanism most students will recognize once it’s named: offloading. When AI handles the thinking – drafting the sentence, generating the argument, suggesting the structure – the student’s brain doesn’t do the work that builds understanding. The output is there, but the learning isn’t.

This isn’t unique to AI. It’s the same reason copying notes by hand tends to produce better retention than just typing them. The friction of doing the work is part of how the brain encodes it. Remove the friction, and you remove some of the learning along with it.

The problem is one that compounds over time. A student who uses AI to draft a paper this week is a little less practiced at drafting than they would have been otherwise. Do that across a semester, and the gap grows.

But it’s not that simple

This is where the research gets more nuanced.

The Anthropic study also found that not all AI use produced the same results. Participants who used AI as a tutor – asking it to explain concepts, checking their own reasoning against its accuracy, requesting feedback on their thinking – retained more than those who simply asked it to complete the task. The difference wasn’t the tool. It was how they engaged with it (Anthropic, 2026).

Dr. Zsolnai-Féhér put it plainly in his summary of the research: use AI to automate things you already understand, and it can be a time-saver. Use it to do things you’re still learning, and you’re trading short-term convenience for long-term competence (Two Minute Papers, 2026).

The researchers themselves were careful about scope. This shouldn’t be considered the final word on AI and learning, but it is a pointed early look at a real pattern that is worth taking seriously.

What this means for students

If you’re writing a paper, the cognitive work of that paper – deciding what to argue, figuring out how to support it, wrestling with the structure, finding the right words for a complicated idea – is not incidental to your education. It is your education. That’s the part that builds the skills your degree is supposed to certify.

Using AI to do that work doesn’t produce a better student. It produces a finished paper.

The research suggests a more useful question than “should I use AI?” is “what am I using AI for?” If you’re using it to understand something you don’t yet understand or to get feedback on thinking you’ve already done – that’s using it as a tool that supports learning. If you’re using it to skip the thinking altogether, the research is pretty clear about what you’re giving up.

Where formatting fits in

There’s one category of academic work where automation doesn’t carry this cost: the mechanical, rule-based tasks that don’t build intellectual skill no matter how carefully you do them by hand.

Formatting a reference list is a good example. Knowing that a journal article goes in sentence case and the journal name is italicized doesn’t make you a better thinker, a stronger writer, or a more capable researcher. It makes you someone who has memorized a rule. And memorizing APA (or MLA or Turabian) rules is not what your professor is assigning a research paper to teach you.

PERRLA handles that kind of work automatically – citations, references, page layout, heading styles, all of it – so you can put your attention where it actually belongs: the argument, the evidence, and the writing. That’s not offloading your education. That’s clearing away the noise so your education can happen.

Try PERRLA free for 7 days – no credit card required.

Frequently Asked Questions

Does using AI hurt your grades?

The research suggests it can hurt your learning, which may eventually catch up to you and negatively affect your grades – particularly on assessments that test understanding rather than production. Students who used AI to complete tasks scored about 17 percentage points lower on follow-up comprehension questions than those who didn’t use AI.

Is AI bad for students?

Not inherently. The research distinguishes between using AI as a tutor – to explain and to give feedback – and using it to complete tasks in your place. The first can support learning. The second tends to undermine it.

What does the research say about AI and learning?

A 2025 study by Anthropic found that participants who used AI assistance scored significantly lower on follow-up comprehension assessments than those who worked without AI. The gap persisted across experience levels. How participants used AI mattered: those who used it as a tutor retained more than those who used it to offload the work entirely.

How should students use AI?

The research points toward using AI to automate tasks you already understand and using it as a tutor for things you’re still learning. Using it to do your thinking for you is where the learning cost shows up most clearly.

Is using PERRLA like using AI to cheat?

No. PERRLA is not a generative AI tool and does not write papers or arguments for students. Built by humans and tested by humans against the APA, MLA, and Turabian formatting manuals, PERRLA automates the mechanical formatting work that doesn’t build intellectual skill regardless of how it’s done. The research on AI and learning applies to offloading cognitive work. Formatting is not cognitive work in that sense.

‍