A non-profit working to promote and research AI safety, train future specialists, and prevent the risks arising from imminent AIs whose capabilities exceed those of humans.
Last updated: March 5, 2025
The International AI Safety Report 2025 identifies numerous AI-related risks [1]. These include:
Leading AI experts such as Geoffrey Hinton and Yoshua Bengio have warned of a major risk in the absence of strict security measures. [2]
CeSIA funds three types of mission:
The Centre pour la Sécurité de l’IA is a French non-profit research organisation that raises awareness of the risks arising from AI systems and advocates technically informed regulation. Their mission is to reduce present and future risks, whether they are systemic, the result of misuse, or arise from the operation of generative AI, which is currently opaque. CeSIA benefits from tax relief in France.
"The acceleration of generative AI developments, fueled by a competitive race and investments that increased by eight in one year to reach 25 billion dollars in 2023, forces us to look at the technological breakthroughs of the coming months and years. [...] Our mission: to provide technical insight into AI development trends, identify current risks and challenges, and anticipate future ones."
Centre pour la Sécurité de l'IA
Tax deductions in France are currently only available for donations to four of our recommended charities: Against Malaria Foundation, Good Food Institute, Helen Keller International, and Centre pour la Sécurité de l’IA
If you are taxable in France, you will benefit from a tax reduction of 66% of the amount of your donation up to the limit of the tax due or 20% of your taxable income.
In practice, this means that if you were planning to donate €100, you can triple this amount to reach a donation of €300 that you will declare, with a real cost of only €102 after tax reduction. This allows you to multiply the impact of your donation without increasing your planned expenditure.
You can read our full article on tax relief in France for more information.
We are able to offer tax deductibility for this association and all the organisations we recommend through our partners, Effektiv Spenden. If your tax residence is in Switzerland, please make your donations via this link, so that you can receive your tax receipt at the end of the tax year.
We work with partners around the world and can offer tax deductibility in many countries. Please contact us if you would like to make a donation from another country and we will discuss the tax options available to you. If you would like to donate and do not require a tax receipt, you can make a donation via Mieux Donner regardless of your country!
The international report on the safety of advanced AI [1] distinguishes three main categories of risk:
In the case of AI, numerous studies document dangerous, unexpected or manipulative behaviours emerging from AI systems without their designers’ knowledge. The international report on AI safety shows that regulating this technology is essential if we want to protect ourselves from its risks.
Furthermore, the scientific approach is not limited to directly observing phenomena that have already occurred. It also relies on theoretical understanding and predictive modelling. In many scientific fields, from astrophysics to climatology, specialists accept robust models of events that have never been directly observed.
Scientific evidence of the risks associated with advanced AI is based on three pillars:
For CeSIA, the prevention and mitigation of AI-related risks must be based on :
Charities that prioritise the final outcome of their action and not indicators of impact. When talking about impact, it is essential to focus on the desired final outcome, such as improved opportunities and quality of life, rather than focusing on intermediate outcomes such as the number of books distributed.
It’s surprising how few charities measure their impact. Our top recommended charities use scientific methods to evaluate the results of their actions, ensuring that their interventions produce the desired effects.
They save or improve the most lives per euro donated to them. Eg) our top recommended global health charities can save a life for as little as 3,000 euros, our top climate charities can avoid one ton of carbon emissions for just 1 euro and our top animal welfare charities can spare thousands of animals from factory farming for the same price of saving one animal in a refuge.
All of our top charities are independently evaluated by world leading charity evaluators. See more information below.
A life is valuable whether it is based in France or the developing world.
Our top recommended charities are all tax deductible in Switzerland, but not in France. This is because French tax law dictates that a charity must operate from France in order to be tax deductible. We always prioritise the impact of a charity over its tax deductible status.
Some of our charities such as Against Malaria Foundation have very low overhead costs, and some such as Clean Air Task Force – have much higher overhead costs. To explain why we don’t take overhead costs into account in our recommendations, imagine two charities:
In the above examples Charity A saves 1 life for every 10,000 spent on it, whereas Charity B saves 1 life for every 3,300 euros spent on it. In this example – it’s clear to see that it’s not overhead costs that are important – but in fact the ratio of overhead costs to the desired impact. Research suggests that there is not much of a correlation at all between overhead and effectiveness.
Finding the best aid organisations isn’t that easy. That’s why we work closely with experts who have been conducting extensive and in-depth research on the subject for many years. These experts are completely independent and work at the cutting edge of charity evaluation in their respective cause areas. They conduct rigorous tests on hundreds of charities to find out as precisely as possible how much good their programs achieve per euro spent. By comparing the cost effectiveness of so many charities, they are able to short list the most effective charities to donate to in order for your donations to have the biggest impact.
GiveWell is the world’s leading research organisation that studies global health and development charities.
‘We search for the charities that save or improve lives the most per dollar. Our goal is to produce the world’s top research on where to give. Free, for everyone. We recommend a small number of charities that do an incredible amount of good’ (GiveWell)
GiveWell was founded in 2007 by Holden Karnofsky and Elie Hassenfeld, two former hedge fund employees. They wanted to do as much good as possible with their donations and found that there was little solid information available on how to do this. GiveWell invests more than 40,000 hours of research each year and has raised more than €1 billion for high-impact charities, saving an estimated 150,000 lives.
The EA Animal Welfare Fund was rated as the top animal charity evaluator by GWWC’s ‘evaluating the evaluators project’ in 2024. They are therefore the primary charity evaluator we defer to in order to choose our top animal charities. We meet with them every few months to get up to date recommendations. They conduct thorough research into various animal welfare organisations to find those that help the most animals per dollar.
Giving Green is a nonprofit organisation that spends thousands of hours each year reviewing studies and climate charities to find the most cost effective interventions to combat climate change. From this research, they publish their top recommendations each year. Their team is made up of climate scientists, economists and impact evaluation experts with decades of experience working at the intersection of evidence-based policy and the environment. We meet regularly with Giving Green to discuss our climate recommendations and run climate workshops in collaboration with eachother.
Although we primarily defer to GiveWell, Giving Green and EA Animal Welfare Fund, we also read the research of other top charity evaluators including LongView Philanthropy, Founders Pledge and Animal Charity Evaluators to ensure we are not overly-reliant on any one source of research. When choosing our top charities we prefer organisations whose effectiveness is corroborated by multiple independent evaluators.
We process donations free of charge, which means we do not keep any of your donation for ourselves. However, the payment methods credit card, PayPal and SEPA direct debit incur fees from the payment services we work with. These fees are deducted from your donation. The amount is calculated as follows for donations within France:
Transfer : Free of charge within the SEPA payment area
Transfer via Donorbox: 2.55% + 0.25 €
Credit card : 2.95% + 0.25 € (for European cards)
PayPal : 3,15 % + 0,25 €
For donations of > 1,000 euros, we recommend a free bank transfer. PayPal and credit cards are not recommended for large amounts due to the costs.
Fees would also be incurred for a direct donation to the respective organisations. We have already negotiated with all payment providers and received particularly favourable conditions for non-profit organisations. We have specially advantageous conversion fees compared to traditional bank.
We also finance ourselves through donations, but completely independently.
If you would like to discuss the best way to donate through us, please get in touch.
On average, €200 donated to CeSIA raises awareness among 720 people [3].
[1] International AI Safety Report, Bengio et al. 2025
Synthesizing toxic molecules: Dual use of artificial-intelligence-powered drug discovery, Urbina et al. 2022, in Nature Machine Intelligence
Assistance to create pathogens : Can large language models democratize access to dual-use biotechnology?, Soice et al. 2023, MIT
Interfere on the results of their own evaluation: Apollo Research, Frontier models are capable of in-context scheming, Meinke et al. 2024, Paragraphe 3.4, Models sometimes scheme without any goal nudging, Anthropic, Alignment Faking in Large Language Model, Greenblat, Denison et al. 2024
Autonomous Self-Replication : Frontier AI systems have surpassed the self-replicating red line, Xudong Pan, et al. 2024, Université de Fudan
Unexpected emergence of negative values and biases : Utility Engineering: Analyzing and Controlling Emergent Value Systems in AIs, Mazeika et al. 2025, Center for AI Safety. Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs, Betley et al. 2025, Truthful AI