MIT's AI picture changed in 2026. The Institute now gives everyone in the community its own multi-model platform, Parley, and its IT office has a clear set of rules for what data can go where. Course rules still come from each subject's instructor.
Here's what the official pages say, how to use AI on psets and research without getting into trouble, and where an outside tool makes sense.
What MIT actually says
We checked these pages in October 2026.
The March 2026 letter on Parley. In An important AI update, dated March 16, 2026, the provost, chancellor and chair of the faculty announced Parley, an MIT-hosted platform for working with several AI models in one place. They gave two reasons: keeping MIT data out of commercial AI systems, and fairness, since some students can afford paid AI subscriptions and others can't. The letter also says a faculty committee on AI use in teaching and learning will propose an Institute AI policy.
IS&T's AI guidance. The IS&T guidance page, last updated June 23, 2026, sets the data rules. Use MIT-licensed tools whenever Institute data is involved. Public AI tools aren't recommended even for low-risk data and aren't allowed for medium-risk data. And "high-risk data should never be used with generative AI tools," including MIT's own. For academic integrity, it points to Institute policy on academic misconduct and to Academic Integrity at MIT.
The Teaching + Learning Lab. TLL advises instructors to "clearly state your policy on the use of generative AI" in their syllabus (TLL guidance). So the rule for any given pset is in that subject's syllabus.
What that means in practice
- Each subject sets its own AI rules. A Course 6 problem set and a HASS essay can have completely different policies.
- Parley is the default for MIT work. It's free to you and covered by MIT's data agreements.
- Research data has hard limits. If your UROP involves anything sensitive, check its risk level before it goes near any AI tool, MIT-hosted or not.
- Collaboration rules still apply. If a subject restricts help from other people, assume it restricts help from AI unless the syllabus says otherwise.
Which models suit MIT work
Psets: understand, don't outsource
Most MIT subjects grade problem sets heavily, and many restrict AI on them. Where AI is allowed for learning, use it on the concept, not the problem.
Try: "Explain why the Fourier transform of a Gaussian is another Gaussian. Show the key step and tell me which property of integrals makes it work. Don't solve any specific problem for me."
Claude is good at careful step-by-step explanation. ChatGPT is strong on math and code reasoning. Gemini handles very long documents, like a full lecture-notes PDF, well.
Code: debug your own
For programming-heavy subjects, paste what you wrote and ask what's wrong, not for a new version.
Try: "This function should return the shortest path but returns None on disconnected graphs. Explain why, and what test case I should add."
Research and UROPs
For literature searches, use a model that links to real sources, then read them. Our Gemini vs ChatGPT for research students post compares two good options for reading and summarizing papers. Remember the IS&T rule: unpublished results and anything covered by a data agreement stay inside approved tools, if they can use AI at all.
Writing and communication requirements
MIT's communication-intensive subjects grade your writing. Use AI as a reader that tells you where an argument is unclear, then revise it yourself.
Gray areas worth thinking through
- Office hours versus AI. If a subject lets you discuss a pset with classmates but not copy, apply the same test to AI: talking through an idea is different from taking an answer.
- Personal accounts. A free consumer chatbot isn't the place for lab data, even data that feels harmless. IS&T doesn't recommend public tools even for low-risk data.
- Old psets and solutions. Pasting a past solution into a model to rework it can break a subject's rules on prior materials as well as its AI rules.
How to use AI at MIT without breaking the rules
- Read the AI policy in each syllabus. It's the rule for that subject, and TLL encourages instructors to state it.
- Use Parley for MIT work. It's the Institute's own platform for a reason.
- Know your data risk level. High-risk data never goes into generative AI, full stop.
- Disclose and cite when it's allowed. MIT Libraries has guidance on citing AI-generated content.
- Verify everything. Derivations, numbers and especially references.
- Ask early. If a pset's rules are unclear, ask the TA before you start, not after you submit.
Where LazySusan fits
Be clear-eyed about this. For coursework and research at MIT, Parley is free, covers several leading models and keeps Institute data under MIT's agreements. Start there.
LazySusan is for what sits outside Institute work: a personal project, a startup idea, a hackathon, or when you want video and voice models in the same place as chat. It puts ChatGPT, Claude, Gemini, Grok, Perplexity and DeepSeek next to models like Veo, Runway and ElevenLabs, in one workspace with one bill. Multi-Chat sends one prompt to several models so you can compare answers side by side. Never put MIT data into it.
The Student plan is $99 a year with a .edu email, with 5,000 tokens a month that roll over. If you want to compare specific models first, see GPT-6 vs Claude Opus 5.5. You can sign up here.
LazySusan isn't affiliated with or endorsed by MIT. Your subject's syllabus and IS&T's data rules come first.
The short version
At MIT, each subject decides how AI may be used, and IS&T decides where data may go. Use Parley for Institute work, keep high-risk data away from every AI tool, read each syllabus and verify what any model tells you.