A random list of advice for OII MSc students

From Max and Alyona.

Note: this is extremely opinionated and drafted very quickly. Many of these points may be incorrect, capture only one perspective, or be relevant only to people with a particular background or set of goals.

MSc supervision

  • Reach out to potential supervisors (in an informed way!).
  • Do not use AI to draft emails or summarise research when reaching out. Students and professors can easily tell (especially after they meet you), and they chat amongst ourselves (so can detect patterns). It is not a good look. Reaching out to supervisors at Oxford is not like cold-emailing for marketing - it is not a volume game.
  • Suggestion: reach out to a supervisor’s PhD students first. PhD students have more time and are more likely to respond, and many supervisors use their PhD students as an informal “screen”.
  • Showing that you are aware of the relevant literature is good. (In my opinion this is best seen as a natural consequence of being genuinely interested in a topic area, rather than something you strive for directly.)
  • However, there is no guarantee that you’ll get your preferred supervisor.
  • You can reach out to any professor at Oxford; co-supervision with non-OII professors is possible.
  • Try to figure out what professors are interested in and excited about. If they are excited about your project, they’re more likely to pick you, and more likely to invest energy in it. Note that both (a) professors’ websites and (b) their most recent publications often won’t reflect their current interests (publications take time to complete, let alone publish). The best option is to chat to them directly, or to their PhD students.
  • Professors are often time-poor and make snap judgements. They are also used to meeting highly specialised researchers. So if you talk to them about one research idea, they may mistakenly assume it is the only thing you want to research. It’s useful to signal that you have a range of interests, if that is indeed the case.
  • Calibrate your level of confidence. For most (but not all) people, this means being more confident than you naturally feel. It is good to position yourself as a potential collaborator with supervisors, rather than as a completely deferential minion. This is typically easier for people with professional experience. On the other hand, OII professors are very knowledgeable and sharp, so bullshitting is not a good idea.

Learning

  • You can just ask to join courses across Oxford. One interesting option is the CDTs (Centres for Doctoral Training), which run focused training in different topics, e.g. health data science, StatML, AIMS, and many others. If you send a polite email to the relevant administrator expressing interest and asking to join, this is often possible.
  • It’s much harder to join or audit official courses, which is why I recommend the CDTs.
  • Check out talks on Oxford Events. There are so many cool people giving talks here all the time.
  • Learning often happens around the classroom rather than in it. Lecturers put in effort, but the curriculum design is all over the place. I found that I learned more (especially about how to do and think about research) by talking about interesting papers, figuring out what to do for your summatives, reading up on different areas to come up with thesis ideas, and of course doing the assignments, which are all focused on different parts of doing research.
  • Some courses can be boring, but because almost all final assignments are some form of research project that you choose, you can have a lot of fun with them.
  • You can get a lot out of a course even if you think you already know the content. For example, in stats, many people said “I know what a linear regression is”, but didn’t know more deeply why it works, when it might fail, or how to apply it in practice. I studied maths and stats in my undergrad, and then worked as a data scientist for quite a while before coming to the OII. Despite all this, I still got a lot out of the stats class, just from the chance to read and think deeply about stats.
  • Chat to PhD students, and feel free to reach out to different people in the department.

For people with English as a second language

Advice from Alyona, not Max.

  • Try not to crash out if you can’t speak English well.
  • You’re smart, and your ideas are good even if you can’t articulate them super quickly in class.
  • People don’t judge you.
  • It gets better after a month or so.
  • Chat to others who don’t have super-polished English so you can share your experiences and support one another.

For context: Alyona got a better average mark than Max on the essay-based assignments, despite roughly 26 years’ difference in how long we’ve spoken English fluently; so do not fear.

Goals

  • It’s good to understand your goal(s). Some people are here literally just for the CV line, and that’s OK. Others want to move to the UK, move into a certain industry, get into research or academia, or even just figure out what their goals are!
  • Knowing your goal makes it easier not to feel jealous or insecure when comparing yourself with others. (E.g. if your goal is to enjoy Oxford life, it’s easier not to feel insecure when classmates are applying to jobs all the time.)
  • A goal is also good motivation when things get tough, whether because of academics, the weather, homesickness or anything else.
  • Having a goal also helps you focus where you spend your time.

Coming up with research ideas

  • Write down fun research ideas in a notes document. Some people were super great at coming up with fun ideas for assignments.
  • Notably, if a topic is fun, it is often a good idea.
  • If you’re struggling to choose between ideas, chat to classmates about them and see what they like. Other perspectives are useful.
  • Listen to podcasts like Hard Fork or Machine Gods, or read Platformer. If you’re into AI, this newsletter is pretty good and more technical (although maybe not ideal as an introduction). Also check out Read Max (not me), which often has good takes on tech and culture.

Food (the most important aspect)

  • The food is better in Cowley.
  • Best options near the Schwarzman Centre: Cohen Quad, Blavatnik, Manos.

Balance

  • Don’t stop exercising.
  • Don’t overcommit. Generally it’s possible to do one or two things besides the degree. (For me, this was rowing, and applying to PhD programmes plus doing things to strengthen my application.)

AI

  • Don’t use AI to write at all, at any point. For a longer argument on this point, see below.
  • Please, please, please, please try not to use AI to answer “dumb questions” or to tutor yourself. You came to Oxford for a community of interesting and engaged classmates: ask them instead! Chatting through these questions with other students has so many side benefits. You signal that you’re interested in an area or method, it starts conversations, explaining things to each other is a great way to consolidate knowledge, and it’s just fun to learn with and from one another. And by role-modelling that you don’t know everything, you make everyone more willing to open up and learn from one another.
  • Use AI to find articles, but read everything before you cite it. Be aware that AI isn’t great at literature review; it frequently puts the emphasis in the wrong places.
  • Use AI to write code, but make sure you understand the choices it is making.
    • As you will see from my code below, I’m pretty comfortable with letting AI write documentation completely autonomously. There are of course exceptions (e.g. explaining a concept alongside code), but IMO it is no longer super important to document code in a particular careful or hand-crafted way, because everyone has agents who can explain parts of code to them. This is a bit of a controversial take.
  • I have a Claude Code skill file for using Brains (the OII GPU server).
  • I cannot overemphasise this: do not produce AI slop or use AI to think on your behalf. People can easily spot it and will judge you for it. They will not judge you for saying something wrong or “silly”; they will judge you for putting out slop. It is even fairly easy to spot when someone is speaking aloud but reciting points an AI came up with for them.
  • Read good writing (articles, blogs, books) regularly, so that you are often exposed to writing other than AI-generated output.

Writing

these are all tips from past feedback

  • Fewer arguments, explained more fully in simpler steps and with examples, are preferable to super-dense work. I got destroyed in a couple of early summatives for writing super-dense article reviews. Give your work to parents or friends from different backgrounds and see if they understand it. If a university-educated person with a professional background doesn’t understand your argument, that’s typically a bad sign.
  • Be constructive: no criticism without a proposed fix or alternative.
  • Pay attention to formatting. Avoid AI-style diagrams, clean up your bibliography formatting, use colours and chart styles consistently, use LaTeX, and write good captions. Look at Paul Röttger’s IssueBench paper for inspiration.
  • Don’t try to be stylish in your writing. Dry humour is not appreciated; just go for academic prose.

Specifically for Research Design & Data and Society 1:

  • (I find this difficult to phrase well.) Be super neutral. You can make a bold argument, but some markers will penalise work whose tone they see as outraged, righteous or ideologically motivated, or in my case, having “moral conviction”. This is based on direct feedback to multiple students in Data & Society 1 in 2025/6 cohort.
  • Relatedly, in Research Design for Social Data Science, include both positive and negative aspects of each paper. Here is a summary of feedback from last year’s cohort.

Backups and submitting early

I myself have three different horror stories from the MSc:

  1. An unsolvable LaTeX compilation issue that first appeared ten minutes before submission (I know, this is ridiculous).
  2. Word crashed and I had forgotten to turn on autosave, losing all my progress (I had to retype it from memory).
  3. A third-party server I had hired suffered a critical error in its data centre (i.e. completely out of my control), and I lost eight hours of progress four hours before a deadline.

So: back up (all code to GitHub, all experimental data to external databases plus local copies), and don’t submit super late.

Read these

If you are considering applying for a PhD in the upcoming round

  • Get ready for a bad winter.
  • Start looking for supervisors now.
  • The SDS scholarship deadline is in January, but Engineering and many other relevant deadlines are in November or December. Check their deadlines.
  • There are a million different ways to get a PhD at Oxford, and to get funding. Many of them are not obvious or easily Googleable (or Claudeable). Please reach out if you want to know more.
  • Being here for the MSc makes you a much more attractive option for RA positions (surprisingly, this is often easier outside the OII).
  • Lecturers and professors earlier in their careers can have more time than more famous professors, who typically have millions of people clamouring for their attention (and to RA with them).
  • Feel free to message me. The OII runs one info session about applying straight away, but if I remember correctly it happens quite late in term. Professors are typically helpful, but in general the OII doesn’t make it easy to apply immediately. I wish I’d had someone to talk to who had applied during the MSc when I was considering it, so I’m happy to be that person now!
  • If there is enough interest, I can give tactical/strategic advice on how to position a PhD application for the OII. I don’t know much at all about how the sausage gets made, but a number of people gave me tips that should probably be more widely available.

Why you shouldn’t use AI to write

[from Max]

AI is improving massively across a wide range of what we would historically call cognitive abilities. This includes the ability to write fluently and well. For example, in a recent experiment, MFA literature students at a prestigious university preferred AI-generated writing. AI can no doubt be used to improve writing, for example by critiquing it. However, if you use AI to write, there are several important downsides to be aware of.

First, writing is about communication. Every sentence you write should serve a purpose. When people write with AI, they typically underspecify their prompt. This leaves a large number of decisions about the communicative intent of the writing up to the AI. The AI will go with its highest-likelihood judgement of the appropriate information to communicate, which is extremely unlikely to match what you intend to communicate, or what is most effective to communicate. In some circumstances this isn’t an issue. For example, I use AI extensively (with human editing) to write appendices. They are often quite procedural, so the AI’s guess about what to include is pretty accurate, and I can guide it very easily. However, I never use AI to write figure captions. Although they might seem as procedural as appendices, there is a careful art to deciding what information the caption should convey to the reader. The AI typically has nowhere near enough context (most of which is sitting in your head, and isn’t digitised anywhere) to make the right decisions about what to communicate. I have tried to automate this sort of writing before, and found that writing the prompt took more effort than writing the text I wanted directly.

Second, writing forces you to properly consider the argument you are making. It is easy to skip over logical steps, knowingly or otherwise, when writing dot points or a prompt to an AI. It is a widespread phenomenon that when you actually start writing an article or assignment, you realise some key part of your argument is wrong or missing a step. If you never go through this careful effort yourself and leave it to an AI, you will miss these problems. Notably, AI is optimised to produce text that is highly persuasive to its user, and is thus directly trained to paper over such gaps undetectably. You might say: “Can’t I just review the text after the AI writes it? I’ll make sure the argument is watertight without having to tap on a keyboard myself.” You won’t.

Third, writing with AI likely degrades your own critical thinking and creative ideation skills, or at the very least denies you the opportunity to improve them. I’m sure you’ve all heard this before. Consider, hard, why some of the most AI-pilled “automate everything” researchers are still intensely focused on careful hand-written communication.

Fourth, you do not want to be judged by others for using AI. It is almost certainly possible to auto-research every single assignment in the MSc and receive a distinction. The university and department are obviously struggling with how to respond to this fact. However, doing so is almost completely uncorrelated (and might even be anti-correlated) with achieving most of the goals people have for doing the MSc. It is intensely obvious when someone is delegating a large share of their thinking or writing to AI. A common response to seeing this is, to put it politely, to “update one’s beliefs about the quality of the researcher and/or their research”. This may or may not be a reasonable judgement, but it is one that many researchers make. I have been in many rooms and presentations where hints that something was AI-generated led people to joke about the quality of the research. It is sometimes obvious when someone’s thinking has been delegated to AI, even when they are talking aloud. This is harder to detect, but I suspect it does not serve your goals if other people believe your thinking is mostly regurgitated AI output. These judgements are incredibly easy to avoid: don’t delegate your thinking or writing to AI. People will judge you far, far, far less for asking a “silly” question, saying something incorrect (in fact, they will typically respect you more for it), or for not having Claude-polished slides than if they believe you are sharing AI-generated content.

One rational reason to judge people who produce AI-generated writing / presentations / slop is because it is, on average, much lower quality research than research which is presented using hand-crafted writing / presentations. This is not a claim about causality: it is perfectly possible to do excellent and ground-breaking research and present or write about it using an AI slop presentation or article. However, it is definitely the case that the average quality of research which is presented using AI is much lower than the average quality of research which is not presented using AI. Researchers are incredibly time-poor, and will frequently use quick heuristics to guess at the quality of a piece of research. Thus, if a paper, email or presentation has AI-generated writing, then it is both a rational and widely-used heuristic to guess that it is low quality.

Fifth, your value-add as a researcher (and likely also as an employee) is basically your creative ideation and critical thinking skills, along with your soft skills and network. If everyone, or many people, has access to the same AI you’re using, and you add little or no improvement over what the AI produces, then you are no different from a Python script. Some goals might be well-served by focusing entirely on soft skills and networking and largely giving up on critical or creative thinking (I have seen some people do this). But I suspect that is not why many of you are at Oxford doing a research-focused graduate degree in a research-focused department.