Pinned Tweet
AI is redefining how the American economy works — how care gets delivered, how retailers serve customers, how businesses run.
But our ability to verify these systems work as designed hasn't kept pace.
Today, we're launching PACT AI to close that gap. 🧵
But our ability to verify these systems work as designed hasn't kept pace.
Today, we're launching PACT AI to close that gap. 🧵
9
9
34
12,287
"Embedded evaluators" is a new idea in AI governance, but embedded oversight has been around for decades. Graeme Finley, our VP of AI Assurance Provider Services, spent 20+ years with a badge and a desk inside large state government IT projects. Here's what that work teaches.
1
1
4
684
His recommendations: negotiate for access to where decisions get made, build independence in from day one with rotation and outside review, and set rules in advance for when to advise. Bank examiners, nuclear inspectors, and IV&V teams have done this for decades.
1
0
1
35
Read Graeme's full post on the PACT AI blog: pactai.org/resources-news…
0
0
0
34
“Buying a model from a frontier lab leaves a company responsible for how it uses that model.”
In @observer, Jeff McMillan makes the case for where enterprise deployers fit in the AI assurance ecosystem, including the work PACT AI members are doing.
In @observer, Jeff McMillan makes the case for where enterprise deployers fit in the AI assurance ecosystem, including the work PACT AI members are doing.
1
0
0
34
Read Jeff McMillan’s column in Observer: observer.com/2026/10/ai-saf…
0
0
0
5
Partnership for Assurance, Credibility, and Trust retweeted
I talked to @POLITICOPro about @pactai_org and the nascent AI assurance ecosystem. Right now, only a handful of organizations are doing independent evaluations of AI systems, and we’re going to need a lot more working on the frontier and the novel risks that are arising.
That’s trust infrastructure this field needs to scale, with assurance providers, deployers, and insurers at the same table.
That’s trust infrastructure this field needs to scale, with assurance providers, deployers, and insurers at the same table.
0
2
5
181
Four Layers of IVO Work:
AI assurance is the work of measuring, evaluating, and communicating the trustworthiness of AI systems. Dozens of organizations already do it, and most don’t call themselves independent verification organizations (IVOs) yet. That is changing fast. nia is building a system to designate assurance organizations to do this work for specific purposes.
So it matters what “independently checked” actually means. Last week, in An AI Evaluator Is Not an Auditor, we looked at the types of assurance work. This week, we look at where that work happens across the AI life cycle.
Internal deployment and development. Is the system trustworthy for use and testing inside the lab? Risks include alignment, loss of control, cybersecurity, and recursive self-improvement. This is the stage outsiders see least.
Pre-release. Is it trustworthy enough to release? Risks include CBRN weapons uplift, manipulation, and training flaws like sycophancy. Once a system is out, these are hard to contain.
Post-release societal. Is it trustworthy for external users? Risks include mental health effects, bias and fairness, and misuse. Many only appear at scale, so this work has to be ongoing.
Use-case level commercial. Is it trustworthy for your employees and customers? Risks include reliability, privacy, and confidentiality. The same system can fit one use case and fail another.
Across every stage: governance and the development process. Who decides how much risk is acceptable, whether risk policies are actually followed, and how training runs and releases get approved.
Each layer builds on the one before. If an early layer has gaps, later layers may miss the problem or be unable to fix it. So before relying on any IVO work, ask which layer it checked and what it leaves to someone else.
Deployers: Ask vendors which layers have been independently checked, and treat gaps as yours until shown otherwise.
Insurers: A claim of assurance means little until you know which layer it covers.
IVO work providers: Say plainly what question you answered, what you looked at, and what you can’t conclude.
Who decides what IVO work should check, who pays for it, and who confirms IVOs are qualified are still open questions. Answering them is the work of building a trustworthy assurance ecosystem, and it’s what PACT AI is currently focused on.
Full article ⬇️
AI assurance is the work of measuring, evaluating, and communicating the trustworthiness of AI systems. Dozens of organizations already do it, and most don’t call themselves independent verification organizations (IVOs) yet. That is changing fast. nia is building a system to designate assurance organizations to do this work for specific purposes.
So it matters what “independently checked” actually means. Last week, in An AI Evaluator Is Not an Auditor, we looked at the types of assurance work. This week, we look at where that work happens across the AI life cycle.
Internal deployment and development. Is the system trustworthy for use and testing inside the lab? Risks include alignment, loss of control, cybersecurity, and recursive self-improvement. This is the stage outsiders see least.
Pre-release. Is it trustworthy enough to release? Risks include CBRN weapons uplift, manipulation, and training flaws like sycophancy. Once a system is out, these are hard to contain.
Post-release societal. Is it trustworthy for external users? Risks include mental health effects, bias and fairness, and misuse. Many only appear at scale, so this work has to be ongoing.
Use-case level commercial. Is it trustworthy for your employees and customers? Risks include reliability, privacy, and confidentiality. The same system can fit one use case and fail another.
Across every stage: governance and the development process. Who decides how much risk is acceptable, whether risk policies are actually followed, and how training runs and releases get approved.
Each layer builds on the one before. If an early layer has gaps, later layers may miss the problem or be unable to fix it. So before relying on any IVO work, ask which layer it checked and what it leaves to someone else.
Deployers: Ask vendors which layers have been independently checked, and treat gaps as yours until shown otherwise.
Insurers: A claim of assurance means little until you know which layer it covers.
IVO work providers: Say plainly what question you answered, what you looked at, and what you can’t conclude.
Who decides what IVO work should check, who pays for it, and who confirms IVOs are qualified are still open questions. Answering them is the work of building a trustworthy assurance ecosystem, and it’s what PACT AI is currently focused on.
Full article ⬇️
1
1
1
103
On the floor at #IAPPpsr26 in Seattle, where AI assurance is a big part of the conversation this week.
At 11:15 AM PT today, @pactai_org's Graeme Finley joins Ashley Casovan, Miranda Bogen and Craig Shank in Room 4C-2 for: "What is AI Assurance? Navigating a Rapidly Evolving here, come see Graeme speak or connect with him after the session.
At 11:15 AM PT today, @pactai_org's Graeme Finley joins Ashley Casovan, Miranda Bogen and Craig Shank in Room 4C-2 for: "What is AI Assurance? Navigating a Rapidly Evolving here, come see Graeme speak or connect with him after the session.
0
1
1
73
A thriving AI assurance ecosystem is key to realizing the many benefits of AI. Tomorrow at #IAPPpsr26, PACT AI's Graeme Finley joins Ashley Casovan, Miranda Bogen and Craig Shank to discuss how we’re bringing together the companies deploying AI with the evaluators, auditors, and are already doing great work to verify that these models are working as intended.
0
1
1
98
Partnership for Assurance, Credibility, and Trust retweeted
Confused about terminology? You're not alone! The field itself doesn't have universal definitions of evaluator v auditor v assessor v investigator v IVO...this is one example, of many, underscoring the need for work and collaboration *today* on definitions, best practices, and "standards."
Below is a great proposal by @pactai_org. I'd now also throw in "IVOs: independent experts, licensed by the government, who test AI models and systems against public standards."
x.com/prpaskov/statu…
Below is a great proposal by @pactai_org. I'd now also throw in "IVOs: independent experts, licensed by the government, who test AI models and systems against public standards."
x.com/prpaskov/statu…
more takes + terms from a recent (and under-appreciated, imo!) blog by @pactai_org:
- An investigator establishes what happened after an incident.
- An evaluator measures what the system does.
- An assessor appraises risk.
- An auditor concludes.
- A decision-maker separate ly a government (or a body it designates), should then convert assessments into criteria.
pactai.org/resources-news…
0
1
6
325
Congratulations to @pactai_org member GLACIS Technologies, and to @joebraidwood and team on their new agreement with Utah's Office of Artificial Intelligence Policy.
Utah will now require AI systems in its AI Learning Laboratory to produce a signed record for each covered AI decision or safeguard check. Those records follow OVERT, an open standard Glacis created and published for anyone to implement.
For people in Utah, that means regulators and independent evaluators will have better evidence of how covered safeguards are operating, rather than relying only on what a vendor says should happen.
For organizations deploying AI, it's an early example of how to earn trust through tangible verification.
glacis.io/blog-utah-sign…
Utah will now require AI systems in its AI Learning Laboratory to produce a signed record for each covered AI decision or safeguard check. Those records follow OVERT, an open standard Glacis created and published for anyone to implement.
For people in Utah, that means regulators and independent evaluators will have better evidence of how covered safeguards are operating, rather than relying only on what a vendor says should happen.
For organizations deploying AI, it's an early example of how to earn trust through tangible verification.
glacis.io/blog-utah-sign…
0
0
1
53
Partnership for Assurance, Credibility, and Trust retweeted
What a night! Thank you to everyone who joined us in DC on the Wild Days rooftop to celebrate the launch of @pactai_org, and a huge thank you to the PACT AI team who made it happen.
The conversation around AI assurance has moved fast. We're no longer asking whether we need it, That's why PACT AI exists: to bring together the deployers, assurance providers, insurers, and researchers building the infrastructure for a thriving AI assurance ecosystem.
It was great to see so many of you in one room.
If you want to learn more or connect with our team, check out pactai.org.
The conversation around AI assurance has moved fast. We're no longer asking whether we need it, That's why PACT AI exists: to bring together the deployers, assurance providers, insurers, and researchers building the infrastructure for a thriving AI assurance ecosystem.
It was great to see so many of you in one room.
If you want to learn more or connect with our team, check out pactai.org.
0
1
1
75
Congratulations to Rob Reich, @pactai_org Advisory Council member, who was named to Governor Newsom’s panel of AI experts that will help bring the AI assurance ecosystem to life in California.
This builds on news from last week that California is speeding up implementation of its new laws to build out the AI assurance ecosystem –– establishing a framework for independent verification organizations (IVOs) that assess AI models and systems for safety and risk, and create a state registry for AI auditors with standards for their independence, transparency, and integrity.
Thank you to the four experts named to this group: Rob Reich, Gillian Hadfield, Alondra Nelson, and Jason Goldman for this important work advising the establishment of California’s AI assurance ecosystem.
This builds on news from last week that California is speeding up implementation of its new laws to build out the AI assurance ecosystem –– establishing a framework for independent verification organizations (IVOs) that assess AI models and systems for safety and risk, and create a state registry for AI auditors with standards for their independence, transparency, and integrity.
Thank you to the four experts named to this group: Rob Reich, Gillian Hadfield, Alondra Nelson, and Jason Goldman for this important work advising the establishment of California’s AI assurance ecosystem.
1
0
3
309
Learn more: gov.ca.gov/2026/09/23/gov…
0
0
1
39
Congratulations to Rob Reich, @PACTAI_org Advisory Council member, who was named to Governor Newsom’s panel of AI experts that will help bring the AI assurance ecosystem to life in California.
This builds on news from last week that California is speeding up implementation of its new laws to build out the AI assurance ecosystem –– establishing a framework for independent verification organizations (IVOs) that assess AI models and systems for safety and risk, and create a state registry for AI auditors with standards for their independence, transparency, and integrity.
Thank you to the four experts named to this group: Rob Reich, Gillian Hadfield, Alondra Nelson, and Jason Goldman for this important work advising the establishment of California’s AI assurance ecosystem.
This builds on news from last week that California is speeding up implementation of its new laws to build out the AI assurance ecosystem –– establishing a framework for independent verification organizations (IVOs) that assess AI models and systems for safety and risk, and create a state registry for AI auditors with standards for their independence, transparency, and integrity.
Thank you to the four experts named to this group: Rob Reich, Gillian Hadfield, Alondra Nelson, and Jason Goldman for this important work advising the establishment of California’s AI assurance ecosystem.
0
0
0
45
Learn more: gov.ca.gov/2026/09/23/gov…
0
0
0
15
Partnership for Assurance, Credibility, and Trust retweeted
Big news for PACT AI out of California: Governor Newsom just signed an executive order to accelerate the implementation of SB 813: the creation of an independent AI verification ecosystem.
BREAKING: @CAgovernor Gavin Newsom just issued an EXECUTIVE ORDER directing actions to address the dangers of recent AI incidents.
The order accelerates California’s new law establishing first-in-the-nation independent oversight of AI companies and safety checks and advances kill switch.”
More to come — watch this space!
1
1
2
109