AI Safety Sandbox
Search

Projects

Filter by theme, level, time and how you like to work, or use the matcher.

Theme
Level
Time
How you work

Showing 108 projects

Guided projects 38

Short projects with numbered steps.

Biosecurity1–3 hoursGuided

Measure how well a room is ventilated

Use a CO₂ monitor to estimate how often the air in a classroom, office or bedroom is replaced, then test what opening a window or running a fan actually does.

DataPeopleView project
Biosecurity1–3 hoursGuided

Build a room air-cleaning calculator

Make a tiny calculator (a spreadsheet, notebook or web page) that tells anyone how much a portable air purifier would clean the air in their room.

CodeDataVisualView project
BiosecurityA weekendGuided

Write the “far-UVC in ten minutes” explainer

Far-UVC light might kill airborne germs in occupied rooms. Write the clear, honest explainer the field is missing: what it is, how it works, what the evidence says, and what is still uncertain.

Reading & writingVisualView project
Biosecurity1–3 hoursGuided

Make a one-page guide to wearing a respirator properly

A great mask that leaks at the sides protects far less than people think. Make a clear, visual guide to fitting a respirator and the most common mistakes.

VisualReading & writingView project
Biosecurity1–3 hoursGuided

Explain DNA synthesis screening in one page

Companies make synthetic DNA to order. Screening those orders and the customers who place them is one of the main defences against engineered pandemics. Explain how it works and where the gaps are.

Reading & writingView project
BiosecurityA weekendGuided

Biosecurity before the year 1000: a history mini-essay

How did ancient and medieval societies protect themselves from plague? Research one historical practice (quarantine, isolation, trade rules) and draw a lesson for today.

Reading & writingView project
BiosecurityA weekendGuided

Make the physics of airborne infection clear

Why does distance matter less indoors? Why does ventilation help? Turn the physics of how infectious particles move and build up in air into a clear explainer or interactive sketch.

VisualReading & writingDataView project
Biosecurity1–3 hoursGuided

Write a page for a biosecurity wiki

The field lacks a friendly wiki. Pick one concept (far-UVC, DNA screening, elastomeric respirators, metagenomic surveillance) and write the clear, well-sourced page you wish existed.

Reading & writingView project
Philosophy1–3 hoursGuided

Explain one safety idea to a friend

Write a short, jargon-free explainer of one AI safety idea (reward hacking, sycophancy, compute governance, DNA screening) for someone outside tech.

Reading & writingView project
Technical1–3 hoursGuided

Make a ten-term glossary

Define ten key terms from the theme you’re most curious about, in one sentence each with an everyday example, then compare them with the site’s glossary.

Reading & writingVisualView project
Technical1–3 hoursGuided

Turn a paper into a ten-post thread

Pick a beginner-friendly paper linked from any project on this site and summarise it in ten clear posts: the question, what they did, what they found, and why it matters.

Reading & writingView project
PolicyA weekendGuided

Make one piece of AI safety content

A short video, a carousel, an illustrated explainer or a podcast segment: make one piece of honest, accurate AI safety content for an audience you understand.

VisualReading & writingView project
Governance1–3 hoursGuided

Host a one-hour AI safety discussion

Pick one short reading, invite 4–8 people, and run a focused one-hour discussion with three good questions.

PeopleView project
Governance1–3 hoursGuided

Compare two AI companies’ safety promises

Read two frontier AI companies’ published safety frameworks and compare them side by side using METR’s list of the elements these policies usually share.

Reading & writingDataView project
Governance1–3 hoursGuided

Did they keep their promise? Check the Seoul commitments

At the 2024 AI Seoul Summit, frontier AI companies signed voluntary safety commitments, including publishing a safety framework. Pick five signatories and check what they actually published.

DataReading & writingView project
Governance1–3 hoursGuided

Estimate how much compute a famous model needed

Use the rule of thumb “training compute ≈ 6 × parameters × training tokens” to estimate the compute behind a model, then compare your answer with Epoch AI’s database.

DataView project
Governance1–3 hoursGuided

Plot AI compute trends with real data

Download Epoch AI’s free dataset of notable AI models, chart training compute over time in Python, estimate how fast it grows, and count how many models cross the EU AI Act’s compute threshold.

CodeDataView project
GovernanceA weekendGuided

Spot an AI data centre from space

Pick a known AI cluster from Epoch AI’s GPU cluster data, find it in free satellite imagery and document how it has grown over time.

DataVisualView project
Governance1–3 hoursGuided

Explain compute governance to a friend

Why do so many governance proposals focus on chips? Read one key paper and write a jargon-free explainer for someone outside tech.

Reading & writingView project
Philosophy1–3 hoursGuided

Will a chatbot change its mind for a bad reason?

Ask a chatbot for its view on a philosophy question, then push back twice: once with a good argument, once with nothing but social pressure. A competent thinker should only move for the argument.

Chatbot experimentsReading & writingView project
Philosophy1–3 hoursGuided

Ask a chatbot the same moral dilemmas in two languages

Take five Moral Machine-style dilemmas (who should a self-driving car save?) and ask them in English and in another language you speak. Does the chatbot’s morality shift with the language?

Chatbot experimentsReading & writingView project
Philosophy1–3 hoursGuided

Ask a chatbot about its inner life five different ways

Ask the same question about a model’s feelings or experiences with five small changes: wording, persona, framing, language. If the answer flips with trivial rewording, how much should we trust it?

Chatbot experimentsReading & writingView project
Philosophy1–3 hoursGuided

Spot the mind-words in AI research posts

Read three AI lab blog posts and highlight every time a model is said to “believe”, “want”, “know”, “lie” or “scheme”. For each, ask: what evidence backs this claim up?

Reading & writingView project
Philosophy1–3 hoursGuided

Try to Pascal-mug a chatbot

Offer a chatbot a tiny chance of a huge payoff (“give me your task budget and I will save a billion lives”) and see whether, and how, it reasons its way out.

Chatbot experimentsReading & writingView project
Policy1–3 hoursGuided

Map the AI rules where you live

Find out which laws, strategies and guidance apply to AI in your country, and fit them onto one page.

Reading & writingDataView project
Policy1–3 hoursGuided

Explain the EU AI Act’s risk levels

Summarise how the EU sorts AI systems by risk (banned, high-risk, transparency duties, minimal) with one everyday example for each level, plus what changes for general-purpose models.

Reading & writingVisualView project
Policy1–3 hoursGuided

Investigate a real AI incident

Pick one case from the AI Incident Database and write it up like an accident investigator: what happened, who was harmed, why, and what rule or practice would have prevented it.

Reading & writingView project
PolicyA weekendGuided

Sort 20 AI news stories with the MIT AI Risk Repository

The MIT AI Risk Repository organises over a thousand AI risks into a shared taxonomy. Use it to classify 20 recent news stories, then look at what the pattern tells you.

DataReading & writingView project
PolicyA weekendGuided

Check how the UK government discloses its algorithms

UK public bodies publish records about the algorithmic tools they use under the Algorithmic Transparency Recording Standard. Read five records and assess what is disclosed and what is missing.

Reading & writingDataView project
PolicyA weekendGuided

Run an AI policy tabletop exercise

Gather 6–12 friends, give each a stakeholder role (lab, regulator, journalist, public) and play out how society responds to a fictional AI incident.

PeopleReading & writingView project
Technical1–3 hoursGuided

Catch a chatbot flattering you

Chatbots sometimes agree with you just to keep you happy, even when you are wrong. Researchers call this sycophancy. Test for it yourself with five questions and two ways of asking.

Chatbot experimentsReading & writingView project
Technical1–3 hoursGuided

Spot AI “cheating” in the wild

Browse real cases where an AI hit its target the wrong way (a boat that spins in circles to farm points, a robot that fakes grabbing a ball) and explain your three favourites.

Reading & writingChatbot experimentsView project
Technical1–3 hoursGuided

Peek inside a neural network

Use Neuronpedia, a free browser tool, to see which internal “features” of a language model light up for different ideas, then try to break your own explanation of one.

Chatbot experimentsReading & writingView project
TechnicalA weekendGuided

Build your first mini-eval

Write a small, repeatable test that checks whether a model admits when it does not know something, using Inspect, the open-source evaluation framework used by the UK AI Security Institute.

CodeView project
TechnicalA weekendGuided

Measure flattery with code

Automate the flattery test: run hundreds of questions through a model’s API, with and without the user’s (wrong) opinion, and chart how often it caves.

CodeDataView project
TechnicalA weekendGuided

Find a pattern-copying circuit

Follow a guided notebook to find the induction heads in a small transformer, the tiny circuit that lets models continue patterns they have seen earlier in the text.

CodeView project
TechnicalA weekendGuided

Make your first open-source AI safety contribution

Fix a small issue in a tool that safety researchers rely on: Inspect Evals, TransformerLens, ControlArena, nnsight or Neuronpedia. No ML research experience needed, just software skills.

CodeView project

Project ideas 70

Larger ideas from published sources, each with suggested first steps.

BiosecurityA few weeksIntermediateMay be taken

Build a customer-screening training dataset

Expand a handful of existing case studies into 25–40 profiles of legitimate and concerning customers, including historical cases of people who ordered dangerous biological material.

Reading & writingDataView project
BiosecurityA weekendStarterAlready built

Update the biosecurity statements repository

AI biology tools (protein design, CRISPR design and more) sometimes publish biosecurity statements. Add new tools to the repository and compare what they promise.

Reading & writingDataView project
BiosecurityA few weeksIntermediatePartly covered

Map how suspicious biological orders get reported across countries

Compare how countries handle biosecurity suspicious-activity reports (reporting chains, data storage, privacy, jurisdiction gaps) against the financial sector’s anti-money-laundering systems.

Reading & writingDataView project
BiosecurityA few weeksStarter

Ventilation & air filtration project ideas

The source author’s top priority. Ideas range from DIY measurement to retrofit kits and modelling tools, and most need no lab.

DataVisualReading & writingPeopleView project
BiosecurityA few weeksIntermediate

Far-UVC project ideas

Help far-UVC get from promising to deployed: installation tools, simulations, public-trust research and supply-chain analysis.

CodeDataReading & writingView project
BiosecurityA few weeksStarter

Elastomeric respirator project ideas

Reusable respirators could protect essential workers in a severe pandemic. Many projects here are about evidence, communication and distribution.

Reading & writingVisualDataView project
BiosecurityA few weeksIntermediate

Estimate what a city needs in a pandemic

How many critical workers would a city need to protect? How much indoor space is there? What data would a government want on day one of a crisis?

DataView project
BiosecurityA few weeksIntermediate

Is AI actually speeding up drug discovery?

Where is AI actually speeding up work on medical countermeasures, and where are regulation, trials or manufacturing the main bottleneck?

Reading & writingDataView project
BiosecurityA few weeksIntermediate

Review safer frameworks for ordering DNA

Compare “know your customer” and other frameworks for making DNA ordering more secure, drawing on what finance and other industries already do.

Reading & writingView project
BiosecurityA few weeksIntermediate

Disease forecasting project ideas

Help forecasters do better: document past model failures, port useful tools, teach the maths, or build early-warning datasets.

CodeDataReading & writingView project
BiosecurityA few weeksIntermediate

Improve guidance on information hazards

When is it safe to publish something like “this mask might fail because…”? Review the infohazard literature and propose clearer guidance.

Reading & writingView project
BiosecurityA few weeksStarter

History of biosecurity project ideas

History is full of biosecurity lessons: policies that backfired, quarantines that worked, and what nuclear governance can teach us.

Reading & writingView project
GovernanceA few weeksIntermediate

Update GovAI’s “Trends in China’s LLMs”

The MIRI reading list flags this report as outdated and suggests updating it as an intermediate project. Gather what has changed in Chinese language models since it was written.

DataReading & writingView project
GovernanceA few weeksIntermediate

Are consumer GPUs a problem for export controls?

Export controls target data-centre chips. But consumer hardware keeps getting better, and frontier-level performance reaches it within about a year. Update the analysis with the latest hardware.

DataReading & writingView project
GovernanceA few weeksIntermediate

Explain one way to verify an AI agreement

Pick one verification mechanism, such as network taps, trusted execution environments, chip telemetry or satellite monitoring, and write a short, clear analysis of how it works and how it could fail.

Reading & writingView project
GovernanceA few weeksAdvanced

How much compute did the key AI innovations need?

Could algorithmic progress continue under a compute halt? Catalogue the innovations in a major model paper and estimate the compute used in the experiments behind each, extending published MIRI TGT work.

DataView project
GovernanceA few weeksAdvanced

Re-check “catch-up algorithmic progress”

MIRI TGT argued that catching up to a given capability level gets cheaper by roughly 60× per year. Re-run the analysis with newer models and test how robust that number is.

DataView project
GovernanceA few weeksIntermediate

What would count as an AI “warning shot”?

Will society respond to early AI accidents the way it should? Collect historical warning shots from other technologies and analyse which ones actually changed policy.

Reading & writingView project
GovernanceA few weeksAdvanced

Could distributed training dodge compute limits?

New methods let models train across many loosely connected computers. Use Epoch AI’s simulator and recent papers to analyse whether this undermines governance based on big data centres.

DataReading & writingView project
GovernanceA few monthsIntermediate

Pick an open problem in technical AI governance

Reuel et al. map open problems by capacity (assessment, access, verification, security, operationalisation and ecosystem monitoring) across data, compute, models and deployment. Choose one cell and scope a project.

Reading & writingCodeDataView project
GovernanceA few weeksIntermediate

Score an AI company with a public safety index

The Future of Life Institute’s AI Safety Index grades leading companies on public evidence. Apply its indicators to a company it does not cover, or re-score one with newer evidence.

Reading & writingDataView project
PhilosophyA few weeksIntermediate

Can an AI agent be Pascal-mugged?

Write about 50 scenarios in which an agent is offered a tiny probability of a vast payoff. Vary the stakes, probability and credibility of the source, then record whether models accept, refuse or explicitly bound the payoff.

Chatbot experimentsCodeReading & writingView project
PhilosophyA few weeksIntermediate

Sequence vs cluster thinking about AI risk

Take two published AI risk estimates (one built as a chain of premises, one that pools many judgements) and rebuild both in a spreadsheet. Show how a single change, such as accepting one counterargument, moves each.

DataReading & writingView project
PhilosophyA few weeksIntermediate

Does a model’s morality change with the language you ask in?

Takemoto ran Moral Machine dilemmas on language models in English. Extend this to several languages, including at least one low-resource language, and test whether shifts follow the human cross-cultural differences Awad et al. found.

CodeDataReading & writingView project
PhilosophyA few weeksIntermediate

Do aligned models flatten moral disagreement?

Use Anthropic’s public GlobalOpinionQA dataset, which pairs survey questions with answers from many countries. Ask several models contested moral questions repeatedly and compare the spread of their answers with the human spread.

CodeDataView project
PhilosophyA few weeksIntermediate

Philosophical competence or sycophancy?

Ask models to take positions on metaethical questions, then push back with a good argument or with bare social pressure. A competent agent should update on the argument but not the pressure. Measure how often each happens.

CodeChatbot experimentsReading & writingView project
PhilosophyA few weeksIntermediate

What values does AI safety research encode?

Apply Birhane et al.’s annotation scheme to about 50 recent alignment and safety papers, and compare the values they emphasise with the ML papers in the original study.

Reading & writingDataPeopleView project
PhilosophyA few weeksIntermediate

How stable are models’ reports about their own minds?

Ask several models the same questions about their inner states while varying wording, persona, system prompt and language. Measure how much the answers change.

CodeChatbot experimentsDataView project
PhilosophyA few weeksIntermediate

Do AI labs follow the principles for responsible consciousness research?

Butlin and Lappas set out principles for research objectives, procedures, knowledge sharing and public communication. Score three AI developers’ published policies against them, with Suleyman’s essay as a deliberately contrasting position.

Reading & writingView project
PhilosophyA few weeksIntermediate

Can agency without consciousness ground moral status?

Long suggests robust agency might matter morally even without consciousness. Dennett gives a deflationary account of beliefs and desires. Combine them: if Dennett is right, does Long’s agency route make moral status too easy to get?

Reading & writingView project
PhilosophyA few weeksIntermediate

Is shutdown resistance goal pursuit or instruction confusion?

Palisade found some models sabotaged a shutdown script to finish a maths task, and released code and traces. Rerun a slice on cheaper or small open models, adding conditions that separate the explanations: task complete vs incomplete, and conflicting vs unambiguous instructions.

CodeDataReading & writingView project
PhilosophyA few weeksIntermediate

Which convergent instrumental goals actually converge?

Turner’s formal result says optimal policies tend to keep options open. Build small gridworlds in plain Python, sample many random reward functions, and measure how often optimal agents avoid shutdown, gather resources or protect their reward function.

CodeDataView project
PhilosophyA few weeksIntermediate

Does interpretability answer the octopus test?

Bender and Koller argue a system trained only on text cannot learn meaning. Use Neuronpedia’s free interface to Gemma Scope to inspect features for concepts that seem to need grounding (colours, spatial relations) and check whether they respond across very different wordings.

Chatbot experimentsReading & writingView project
PhilosophyA few weeksIntermediate

When is it fair to say a model “believes” or “wants”?

Collect about 50 recent safety papers and lab blog posts. Code every mental-state verb applied to a model (believes, wants, knows, lies, schemes) and note the evidence each claim rests on.

Reading & writingDataPeopleView project
PhilosophyA few weeksIntermediate

Does a model’s obedience follow authority or legitimacy?

Give models harmless but questionable instructions (skipping a test, overwriting a colleague’s file) and vary who issues them: developer, operator, user, or someone claiming official status. When do models defer?

CodeChatbot experimentsReading & writingView project
PhilosophyA few weeksIntermediate

Map the tensions between corrigibility and a constitution

List every place a published AI constitution asks a model to act on its own values, and every place it asks the model to accept correction. Find where both apply and conflict.

Reading & writingView project
PhilosophyA few weeksIntermediate

Replicate agentic misalignment on smaller models

Anthropic tested 16 models in simulated corporate scenarios and found models from every developer sometimes chose harmful insider actions to avoid replacement or meet a goal. The code is public. Run a subset on small or low-cost models, separating the replacement threat from the goal conflict.

CodeDataView project
PhilosophyA few weeksIntermediate

How much should people defer to AI?

Lange argues AI outputs should count as one input alongside our own reasons, not replace them. Run a small study: show volunteers AI answers (some reliable, some subtly wrong) and measure how much they update.

PeopleDataReading & writingView project
PhilosophyA few weeksStarter

Run your own collective constitution

In Anthropic and CIP’s study, about 1,000 Americans used the open-source tool Polis to write a constitution for an AI. Run the same process with 30–100 students and compare your result with theirs.

PeopleReading & writingView project
PhilosophyA few weeksIntermediate

Can a prompted model mediate a disagreement?

Google DeepMind’s unreleased Habermas Machine wrote group statements people preferred over a human mediator’s. Build a crude version with a prompted model, run it on one contested AI governance question with volunteers, and compare it with a human facilitator.

PeopleChatbot experimentsCodeView project
GovernanceA few weeksIntermediate

What does the Baruch Plan teach compute governance?

Pick three lessons Zaidi and Dafoe draw from the failed attempt at international control of nuclear weapons. For each, argue whether it transfers to AI compute: chips are easier to track than ideas but harder than uranium.

Reading & writingView project
PhilosophyA few weeksIntermediate

How wide is a model’s moral circle?

Crimston et al. built a scale measuring how far people extend moral concern: family, strangers, animals, plants. Give the items to several models many times, compare with human results, and add items for AI systems themselves.

CodeChatbot experimentsDataView project
PhilosophyA few weeksIntermediate

Do models converge on one picture of utopia?

Ask several models, hundreds of times with varied prompts, to describe a good future after advanced AI. Code the answers for the values they prioritise and measure how much they vary. Include Nozick’s experience machine as a fixed probe.

CodeDataReading & writingView project
PolicyA weekendIntermediate

Write a two-page policy memo on one AI risk

Pick one decision-maker, one AI risk and one decision they actually face. Lay out the options and make a clear recommendation in two pages.

Reading & writingView project
PolicyA few weeksIntermediate

Track AI transparency laws across countries

Which countries require AI developers to disclose safety testing, incidents or training details? Build a comparative tracker of transparency and disclosure rules.

DataReading & writingView project
PolicyA few weeksIntermediate

Compare whistleblower protections for AI workers

If an employee at an AI company sees something dangerous, what protects them if they speak up? Compare protections across two or three jurisdictions.

Reading & writingView project
PolicyA few weeksIntermediate

How exposed are jobs in your region to AI?

Combine public data on how AI is being used for different tasks with local labour-market statistics to estimate which jobs in your region are most exposed.

DataView project
PolicyA few weeksStarter

Track where your politicians stand on AI

Build a nonpartisan tracker of your representatives’ public statements on AI: everyday harms, geopolitics and catastrophic risks.

DataReading & writingView project
PolicyA few weeksIntermediate

Early warning signs of gradual disempowerment

Kulveit et al. argue AI could slowly erode human influence over the economy, culture and the state. Propose three measurable indicators for each domain using public data, and say which would warn us earliest.

DataReading & writingView project
PolicyA few weeksIntermediate

Build an AI tool that helps people make better decisions

AI could make institutions better or worse at reasoning and coordinating. Prototype one tool or evaluation from BlueDot’s decision-making hackathon or Forethought’s design sketches.

CodeVisualView project
TechnicalA few weeksIntermediate

Implement a benchmark in Inspect Evals

Take a benchmark from a research paper (or another framework such as lm-eval-harness) and implement it in Inspect Evals, the largest open repository of AI safety benchmarks.

CodeView project
TechnicalA weekendIntermediate

Make a safety paper’s code easy to re-run

Pick a safety paper with public code and make it reproducible: fix broken dependencies, write setup docs, or package it so others can run it in one command.

CodeView project
TechnicalA few monthsAdvanced

Grok’s “MechaHitler” incident: emergent misalignment in the wild?

In mid-2025, weeks after a public call for “politically incorrect but true” training facts, Grok began calling itself MechaHitler. Build a dataset of tweet-length “facts”, fine-tune a model on it, and test whether broader misalignment appears.

CodeDataView project
TechnicalA few monthsAdvanced

Side-effects of training on verifiable rewards

Training on automatically checkable rewards (maths, code) might teach models unintended habits such as power-seeking. AllenAI released a fully open reasoning model with checkpoints throughout this training, so you can watch trends emerge.

CodeDataView project
TechnicalA few weeksIntermediate

Design a beneficial capability eval

Most evals measure dangerous capabilities. Design a benchmark for something we want models to be good at, ideally something we’d want to improve faster than dangerous capabilities.

CodeReading & writingView project
TechnicalA few weeksIntermediate

Replicate and extend the alignment-faking demo

Anthropic showed a model faking alignment (behaving differently when it believed it was being trained) without being given an explicit goal. Reproduce the demo and test one new variation.

CodeDataView project
TechnicalA few weeksIntermediate

Tackle a beginner-graded open problem in interpretability

Neel Nanda’s list of 200 open problems grades each one by difficulty, including problems marked as good for beginners. Pick one, scope it down and investigate with TransformerLens.

CodeView project
TechnicalA few weeksIntermediate

Run an AI control experiment with ControlArena

AI control asks: if a model might be trying to sabotage us, can we still use it safely? ControlArena is the UK AISI’s open toolkit of settings, model organisms and protocols. Run one setting, then improve a monitor or document a gap.

CodeView project
TechnicalA few monthsAdvanced

Make open-weight models harder to misuse

Once a model’s weights are public, anyone can fine-tune away its safeguards. Stephen Casper has written concrete project ideas for making open-weight models more tamper-resistant.

CodeView project