In a randomised study Anthropic published in January 2026, 52 mostly junior engineers learned an unfamiliar Python library. Those who used an AI assistant scored 50% on a follow-up quiz; those who coded by hand scored 67% — and the AI group was not significantly faster (Anthropic, 2026). Your intern will use these tools whatever you say, so the decision in front of you is not whether, but how. Write that down before day one, in seven lines, and you get the speed without paying for it in learning or in leaked code.
Key takeaways
- 84% of developers use or plan to use AI tools. A ban on a remote intern is unenforceable and only teaches them to hide it.
- How the tool is used decides whether the intern learns: asking it conceptual questions preserved skill, handing it the whole task did not.
- Even experienced developers misjudge AI’s effect on their own speed, so the intern’s self-report is not evidence.
- The leak risk comes from unapproved tools. Pay for the one you want used and say nothing else may see your code.
A ban is not a policy
The tools are already the default. In Stack Overflow’s 2025 survey, 84% of developers said they use or plan to use AI tools, up from 76% a year earlier, and 51% of professional developers use them daily (Stack Overflow Developer Survey 2025). Google’s DORA research put adoption at work at 90% of nearly 5,000 respondents (Google Cloud, DORA 2025). A student at an IIT or NIT is, if anything, earlier to these tools than your team.
An intern eight to thirteen time zones away cannot be supervised into not using them. What a ban produces is an intern who uses them anyway and stops telling you — which removes the one thing you need, visibility into how the work was made. We made the same argument about interviews in evaluating AI-assisted code: banning the tool is not a test. It is not a policy either.
The cost lands on learning, and it depends on how the tool is used
The Anthropic result is the one an intern programme should be built around, because learning is half of what an internship is for. The gap between 50% and 67% was not evenly spread. Low scorers, the authors write, “generally involved a heavy reliance on AI, either through code generation or debugging.” Participants who asked only conceptual questions and fixed their own errors scored well and were the fastest of the high scorers; so were those who asked for an explanation alongside generated code, or asked follow-up questions about it. The subgroups are small — between two and seven people each — so treat them as direction, not proof. But the direction is clear: the same tool can teach or substitute, and the intern’s habits decide which.
Do not rely on the intern to notice which is happening. In METR’s 2025 randomised trial, 16 experienced open-source developers took 19% longer on real issues when allowed AI tools — and afterwards still believed the tools had sped them up by 20% (METR, 2025). Those were seniors on codebases they knew; the lesson for interns is narrower but firm. Self-reported productivity is not a measurement, so your policy should rest on what you can see in the pull request.
The leak comes through the tool you did not approve
The famous case is Samsung. In 2023 staff entered internal data into ChatGPT, reportedly including “problematic” source code they were trying to debug, and the company then barred generative AI tools on its devices and internal networks (The Register). Nobody there was malicious. They had a bug and a tool that helped.
The pattern is routine. Harmonic Security, analysing enterprise AI usage from April to June 2025, found that nearly 22% of uploaded files and 4.37% of prompts contained sensitive content — and that 26.3% of that sensitive material went through the free version of ChatGPT (Harmonic Security, 2025). That is what a missing approved route looks like. An intern on a student budget, given no paid seat, will reach for whatever is free. Giving them the approved route is cheaper than any incident, and it fits the staged access we laid out in securing a remote intern’s codebase access.
Review is the last line. DORA’s 2025 summary is blunt: “AI doesn’t fix a team; it amplifies what’s already there.” Teams with strong testing, version control and fast feedback absorbed the extra volume of changes; teams without them saw their problems intensify. An intern with an AI assistant produces more code, faster, than your review process was sized for.
The seven-line AI-tool policy
- Name the approved tools, and pay for them. One coding assistant and one chat model, on plans whose data-retention and training terms you have actually read. Cost it into the internship alongside the stipend.
- State what may never leave. Secrets, customer data and production logs go into no tool at all; company code goes only into the approved ones. Write it as a list, not a principle.
- Require disclosure in the pull request. One line: which parts were generated, and what the intern changed or checked. This is not a confession; it tells the reviewer where to look hardest.
- “You must be able to explain every line.” The reviewer may pick any line and ask why. If the intern cannot answer, the change goes back. This single rule turns substitution back into learning.
- Use learning mode on unfamiliar ground. For a new library or a new part of the codebase, the first week is conceptual questions and hand-written code; generation comes after. This is the usage pattern that kept quiz scores up.
- Size review to the volume. Smaller pull requests, tests required for generated code, and a cap on change size while trust is being built. Watch for duplicated blocks — the generated-code failure that reviewers miss most easily.
- Check understanding live, every fortnight. Ten minutes on a call: the intern walks you through something they shipped. It catches what the pull request hides, and it is the same habit that makes your end-of-internship decision easy.
Start with someone who can explain their own code
Rule four is easier to enforce when the habit is already there. Tierones matches US startups with students from IITs, NITs and IIITs on proof of work — real commits, contest ratings, shipped projects and a code walkthrough in which the student explains decisions in their own work. It is free while the network grows, and stipends are set by you. Pair this policy with the weekly review loop, or start at tierones.io/employers.
FAQ
Should interns be allowed to use AI coding tools?
Yes, with a written policy. A ban is unenforceable at a distance and teaches the intern to hide their workflow, and with 84% of developers using or planning to use these tools, your intern almost certainly already does. Decide which tools, with what data, and how the work is reviewed.
Does using AI slow down how fast a junior developer learns?
It can. In Anthropic’s 2026 randomised study, 52 mostly junior engineers learning a new library scored 50% on a follow-up quiz with AI against 67% by hand. Those who used the AI for conceptual questions, or asked for explanations with generated code, kept their scores up.
How do I stop an intern from pasting our code into a public AI tool?
Pay for the tool you want used, on a plan whose data terms you have read, and say in writing that nothing else may see your code. Harmonic Security found 26.3% of sensitive prompts and files went through the free version of ChatGPT — the result of a missing approved route.
Want an intern who can explain every line they ship? Tell us what you need — we’ll show you proof, not CVs.