About us
About us
We are AI optimists. We work on the positive uses of technology and on the legislation that decides where it goes.
Who we are
We are AI optimists.
We think the point of AI is to give people more of what makes them human: health, time, judgment, and the chance to care for one another. Our work is to make sure it is pointed there.
We take the risks seriously and want them mitigated. Our answer is not to slow technology down but to point it at good uses, and to help write the legislation that keeps it pointed there.
We publish the research and lobby for support where the evidence is good. Where it helps, we build a working version of what we argue for, so its value can be measured instead of guessed.
How we work
The research, and a build of what it describes.
We publish research on what AI can do for people, and on the legislation that guides it.
When a committee asks what a rule would cost, or whether a developer can meet it, we answer with figures from a system we run rather than an estimate. When the system turns out to be something people need on its own, we finish it and ship it as a product.
- 01
Understand the problem
We start from how the technology works, who it reaches and how it fails, not from the rules that already exist.
- 02
Build and test
Where it helps, we build a working version of the system the proposal describes, so its costs and limits can be measured.
- 03
Measure it in practice
We run it under real conditions and make it available to the people who would use it, which produces real figures on cost and performance.
- 04
Publish and brief
We publish the research with the evidence behind it and brief legislators and officials who ask for it.
Where we work
Eight areas of research.
Frontier AI
How do we make frontier AI dependable enough to trust with real work?
Evaluations and evidence that let good systems be deployed with confidence, and clear rules on liability so the people building them can move.
Biosecurity
How do we keep the gains from AI in biology without making a pandemic cheap to start?
Screening at synthesis, safeguards that keep pace with the tools, and rules written so legitimate research can move.
Capital and Access
How can a person reach capital, credit and payments without asking a gatekeeper's permission?
Digital identity, credit and payments for the people incumbent institutions do not serve. Inherited from Fidutam's field work, and still running.
Autonomy and Force
How should AI be governed where it is used to decide on the use of force?
Human control over the use of force, testing and procurement standards for autonomous systems, and export rules that hold without handing the field to the least careful.
Cyber and Critical Systems
How does the law reach attacks on power, water and hospitals when the attacker is in another country?
Ransomware, fraud and extortion at scale, attribution and cross-border enforcement, and who is liable when insecure software makes the attack possible.
Automated Decisions
How can a person contest a decision a machine made about them, and win?
Insurance denials, benefits, credit and hiring. The right to human review exists on paper in several places; the systems that would make it real mostly do not. Arvo is our first build here.
Privacy and Data
How can a system prove it holds only what it needs?
Data minimisation and limits on surveillance, by companies and by the state, built so that the proof is a property of the system rather than a promise.
Priorities and Measurement
Where does public money for AI do the most good?
Comparing what AI investments return in health, time and opportunity, so that funding goes where it helps people most.
On the record
…your initiative sets a powerful example of how passion and purpose can drive meaningful change…
Kashaf Alvi, former Policy Director at Perrin, on the BBC discussing technology governance.
Covered by, and worked with: BBC, MIT Solve, U.S. Senate, UNESCO, UNICEF, IBM, Meta, OpenAI, Google.
On stage
Keynote in Canada
Finn Järvi, Perrin's founder and Chairman, gave a keynote for Perrin at the World Talent Forum in Canada. At eighteen he was the youngest speaker there.
A student startup-pitch competition in Montreal. Its first event drew 230 teams from more than 15 countries to five finalist places; the second ran on a stage at Startupfest, Canada's largest startup conference, with a $10,000 grand prize.
- 2025Downtown Montreal, 7 June: five finalists, over 150 investors, founders and students in the room.
- 2026Startupfest, Montreal's Old Port, 10 July: Sielo Robotics won the $10,000 grand prize, backed by TD Bank and Olymbec.
Judges and partners
- Anaiy Somalwar Founder, Stealth; formerly Foundation Capital and Pear VC
- Federico Arellano Founder, PitchJams
- Haley Bryant Partner, Hustle Fund
- Joel Hansen Analyst, League of Innovators VC
- Julian Weisser CEO and Co-Founder, ODF
- Lucille Mazé Associate Director, Accenture
- Michael Gaul COO, Beyond Air; formerly COO, SynCardia Systems
- Mishel Wong Co-Founder, National Wash Network
- Myles Shedden CFO, Hyperlume; formerly VP, SkipTheDishes
- Silvana Huaman CRO, FTEX
- Zack Ahmed Strategic Director, Superpower Health
Partners included Harvard Ventures, Front Row Ventures, League of Innovators (LOI), Sun Life, CIBC, Berkeley AI, FTEX, Elantech, Propolys, PitchJams, Merge Club.
Governance
- Operating name
- Perrin Institution
- Registered as
- Perrin Institution, Inc.
- Status
- 501(c)(3) think tank
- Registered office
- 303 Banyan Blvd, Suite 400, West Palm Beach, FL 33401
- Founder and Chairman of the Board
- Finn Järvi
- Founder, President and Member of the Board of Directors
- Cash Hilinski
- Founder and President Emeritus
- Okezue Bell
A note on affiliation
Perrin & Fidutam counted more than 1,600 members in its community over the years, and Perrin Institution inherits that record. Membership does not constitute employment. The organization has only ever employed a small staff, fewer than 25 people, and any other individual claiming to be on the team of, employed by, or otherwise representing the organization is making a false claim. To verify someone's affiliation, contact us directly.
The evidence
The risks we work on, as others measure them.
- Considered relevant
- Actively mitigated
Source: Stanford HAI, AI Index 2026, fig. 3.3.5 (McKinsey survey), April 2026 (9th edition). Redrawn from the AI Index public data, CC BY-ND 4.0. The eight risks most often named.
View the data
| Considered relevant | Actively mitigated | |
|---|---|---|
| Inaccuracy | 74 | 70.6 |
| Cybersecurity | 71.8 | 61.5 |
| Regulatory compliance | 62.9 | 53.4 |
| Personal privacy | 54.4 | 44 |
| IP infringement | 51.4 | 39.7 |
| Autonomous system actions | 43.8 | 28.8 |
| Reputation | 37.3 | 22 |
| Explainability | 36.3 | 29.2 |
- Operational
- Under construction
- Planned
Source: Global BioLabs (King's College London), Global BioLabs Report 2023, February 2023.
View the data
| Operational | Under construction | Planned | |
|---|---|---|---|
| Europe | 24 | 1 | 1 |
| Asia | 9 | 0 | 11 |
| North America | 12 | 1 | 2 |
| Oceania | 4 | 0 | 0 |
| Africa | 2 | 1 | 0 |
| South America | 0 | 0 | 1 |
Source: UNODA BWC Implementation Support Unit, Confidence-Building Measures briefing, slide 7, August 2025. In 2024 that was 60% of States Parties.
View the data
| Reports per year | |
|---|---|
| 2010 | 71 |
| 2011 | 69 |
| 2012 | 69 |
| 2013 | 65 |
| 2014 | 72 |
| 2015 | 73 |
| 2016 | 83 |
| 2017 | 78 |
| 2018 | 80 |
| 2019 | 83 |
| 2020 | 86 |
| 2021 | 94 |
| 2022 | 99 |
| 2023 | 106 |
| 2024 | 113 |
- In favour
- Against
- Abstained
Source: UN General Assembly First Committee, as reported by Stop Killer Robots, recorded votes 2023 to 2025, 2023-11-01 / 2024-11-05 / 2025-11-06.
View the data
| In favour | Against | Abstained | |
|---|---|---|---|
| 2023 | 164 | 5 | 8 |
| 2024 | 161 | 3 | 13 |
| 2025 | 156 | 5 | 8 |
Source: SIPRI via Our World in Data, Military Expenditure Database, April 2026, April 2026 (OWID processed 2026-04-27). CC BY 4.0.
View the data
| US$ billion, constant 2024 prices | |
|---|---|
| 2016 | 1,961.8 |
| 2017 | 1,982.9 |
| 2018 | 2,044.6 |
| 2019 | 2,131.6 |
| 2020 | 2,209.7 |
| 2021 | 2,227.8 |
| 2022 | 2,299.3 |
| 2023 | 2,453.4 |
| 2024 | 2,692.3 |
| 2025 | 2,771.3 |
Source: METR, Task-Completion Time Horizons of Frontier AI Models, Last updated 8 May 2026. Each point is the state-of-the-art model at its release; confidence intervals are wide.
View the data
| Length of task, in the time an expert takes (log scale) | |
|---|---|
| GPT-2 | 3 sec |
| davinci-002 | 8 sec |
| GPT-3.5 Turbo Instruct | 36 sec |
| GPT-4 | 4 min |
| GPT-4o | 7 min |
| Claude 3.5 Sonnet | 11 min |
| o1-preview | 20 min |
| o1 | 39 min |
| Claude 3.7 Sonnet | 1 hr |
| o3 | 2 hr |
| GPT-5 | 3.4 hr |
| Claude Opus 4.5 | 4.9 hr |
| GPT-5.2 | 5.9 hr |
| Claude Opus 4.6 | 12 hr |
| Claude Mythos Preview | 17 hr |