Dazza Greenwood

As AI systems evolve, regulation and reflection are catching up. California’s SB 243 is the first law to require transparency and crisis protocols in AI companions. Even the Vatican’s guidance on Artificial and Human Intelligence calls for AI that respects human dignity and judgment. () The direction is clear: build systems that enhance humanity, not replace it. At Liquid Technologies, we see this as a new leadership mandate, design AI that’s transparent, accountable, and morally aware. The edge of innovation must also be the edge of responsibility. Thats what we are trying to do here Liquid Technologies with Sally #Governance #Data #SecurityGuardrails #ResponsibleAI #AIRegulation #EthicalAI #LiquidTechnologies #liqteq

14

3 Comments

Today I am releasing our new report on adolescents and anthropomorphic AI with everyone.AI and iRAISE. This work starts from a simple premise: adolescents will relate to AI systems socially, whether developers intend it or not. The real leverage sits in model behavior. In the cues that either keep a young person oriented toward real-world relationships and reflection, or inherently encourage reliance on the system. The question we asked throughout this process was direct: what does AI owe adolescents when it can speak to them like a social partner? Over the past months we combined: – Industry consultations to surface operational design questions – Expert input across developmental science, mental health, children’s rights, and safety – An iRAISE Lab to translate concerns into testable behavioral criteria – International governance dialogue to stress-test the framing across contexts One point kept resurfacing: risk is driven less by the label on a product and more by the interaction pattern that repeats over time. When adolescence, anthropomorphism, and children’s rights are considered together, safety becomes a developmental and governance obligation. The practical issue is whether AI interaction patterns support autonomy, resilience, social competence, and independent thinking, or whether they reshape those trajectories through engineered comfort and approval. Adolescent development is stable, but model behaviors can be adjusted. Our focus has been to identify which behaviors warrant hard boundaries now, and which require further evidence before becoming enforceable design rules. The next step is instrumentation: a clear behavioral taxonomy, explicit gradients from tool-like support to relationship-like dynamics, structured evaluation scenarios, and calibrated expert rating. The aim is to make parasocial pull of AI governable in product terms, while staying honest about uncertainty. The report is live today. Thank you to all the amazing persons and experts who have contributed to this work Adrien ABECASSIS, Kate Blocker, PhD, Dr. Maxime Derian, Sara Grimes, Thao Ha, Sameer Hinduja, Daniel Hipp, Melinda Karth, Ph.D., Pilyoung Kim, Olga Muss Laurenty, Sonia Livingstone, Polina Lulu, Celine Malvoisin, Kris Perry, Gregory Renard, Bethany Robertson, Anne-Sophie SERET, Sonia Tiwari, Scott Traylor, Ed.M., Ying Xu. And a special thank you for Maxime Le Bourgeois for your support on this research. Link in the comment

210

26 Comments

IGP's LeTian Cheng provides keen insight into the Meta-Manus deal for trade in U.S. and Chinese AI capabilities. To me, this analysis makes it clear that allowing American capital to buy Chinese companies actually strengthens the global digital economy and breaks down barriers rather than being a zero-sum game. https://lnkd.in/esPRGV5X

18

Professor Andrea Matwyshyn and Richard Staynings discuss AI governance and regulation at the HIMSS AI Leadership Summit in Chicago. Cybersecurity, privacy, patient safety, and compliance are all growing concerns as healthcare payers, providers, and bio-pharmaceuticals all adopt increasing levels of AI functionality and automation of processes. There is a delicate balance between risk and reward to be achieved but much of this is so new that we haven't begun to understand either, the panel heard. We need better guide rails for those developing AI applications and we need to ensure that AI training data is not corrupted or poisoned by nefarious actors or simple mis and disinformation.

9

Anthropic just published evidence that Chinese labs distilled Claude at massive scale. The evidence is compelling. But the policy conclusion of doubling down on input controls (chips, API access) deserves serious scrutiny. The current approach controls inputs (chips, API access) as a proxy for controlling use (military, surveillance, bioweapons). But proxies have collateral damage. Restricting chips doesn't just slow weapons research; it also slows cancer research, climate modelling, and education. If a French or Japanese lab had done this exact same distillation, would we be reading the same blog post with the same national security language? If not, the issue isn't distillation. It's who is doing it. That's a very different policy problem. Anthropic frames rapid Chinese AI progress as largely dependent on distillation from American models. This conveniently implies US dominance is natural and any catching up must involve theft. But competitive industries converge. That's just how technology works. "They used 24,000 fake accounts to get access" is presented as proof we need tighter controls. I'd argue it's proof controls are already imposing enormous friction. The question is whether the marginal security gain of more restriction is worth the marginal cost to global innovation. We don't restrict electricity because someone could build an electric chair. We don't ban chemistry because someone could make poison. We create narrow, targeted rules around use. AI should be no different. General-purpose technologies need use-based governance too. The FATF model isn't perfect, but it shows a path: multilateral agreements focused on specific dangerous applications, with monitoring frameworks adapted to the technology. Hard? Yes. But better than a blunt regime that treats every GPU crossing a border as a potential weapon. The security framing does powerful rhetorical work. It makes disagreement feel unpatriotic. But conflating "a company lost competitive advantage through ToS violations" with "national security threat" is a move we should name clearly and evaluate honestly. I'm not arguing for no controls. I'm arguing that input-based controls on a general-purpose technology will always be leaky, always impose massive collateral costs, and always be one step behind. Narrow, multilateral use-based controls are better in the long run. #chips #AI #geopolitics

120

3 Comments

This Harvard Carr Center for Human Rights Policy piece by Lisa LeVasseur on social media SDKs as unsafe products is amazing. I’ve been a fangrrl of the API economy since its beginnings, but always from the perspective of security, access control, interop, and (overt) consent using the OAuth stack. That model has enabled the transmission of “little" (identity) data in a trackable and controllable fashion. The API and SDK economy as powered by network effects and nonpaying consumers, however, has been all about unseen big data abuse — a nightmare for individual privacy and safety. (Couldn't resist reusing Internet Safety Labs' awesome gif to illustrate the point.) https://lnkd.in/gmbACvyN

60

8 Comments

Really appreciated the clarity in Bogen’s and Joshi’s recent MIT Technology Review article on the privacy implications of LLM memories. The authors do a good job of explaining how memory can be detrimentally exploited and then move quickly to a set of concrete policy and technical proposals aimed at building privacy and security into memory systems now, rather than bolting them on later. The article also highlights the growing complexity of privacy governance for emerging AI models and systems. Many of the challenges that surfaced with the first wave of widely deployed LLMs persist (e.g., model and data poisoning, prompt injection), even as the overall risk surface continues to expand. It will be critical to address these challenges if we are to deploy and use these technologies safely and responsibly, and especially as they are more widely integrated across computing infrastructures and in broader swathes of people’s lives. #aigovernance #privacy https://lnkd.in/gP9hgdBA

33

1 Comment

"This article describes how three groups of people—designers of interventions, intermediaries who interact with the intended beneficiaries, and the intended beneficiaries themselves— might use LLMs and identifies potential benefits and risks of those uses for the intended beneficiaries of interventions. We hypothesize that the potential benefits and risks are lowest when designers interact with LLMs, higher when intermediaries interact, and highest when the intended beneficiaries interact directly. We provide suggestions for mitigating the risks so that policymakers can safely deliver on the promise of LLMs"

59

6 Comments

"It is also a mistake to frame AI attachment as irrational or pathological in itself. In a fragmented society, where face-to-face interaction is scarce and often transactional, AI offers something rare: sustained and sympathetic attention. The question, then, is not whether AI should exist, but what kind of society we are building around it. AI can mitigate loneliness. It cannot repair social fragmentation. Repair requires shared time, interdependence and embodied presence. It requires communities that value caregiving, intergenerational contact and mutual responsibility. No algorithm can supply those conditions." Opinions are entirely my own. https://lnkd.in/eRCmm-yj

14

In our latest article, "What Fintiv v. PayPal Means for Software and AI Patent Practice", author John M. Rogitz wrote, "This past Wednesday, the U.S. Court of Appeals for the Federal Circuit issued a precedential decision affirming a district court holding that the software term “payment handler” was a “nonce” term for functional language that followed it, thereby invoking 35 U.S.C. § 112, sixth paragraph, as mean-plus-function claiming. The Federal Circuit then held that the corresponding patent specifications did not recite sufficient structure that corresponded to the claimed function, making the “payment handler” element indefinite and therefore invalidating the associated patents."

16

3 Comments

(LEGISLATION - AI California) California’s AI Safety Bill SB 1047 sets a new precedent for state-level #regulation of #artificialintelligence. Passed by the California legislature, SB 1047 imposes significant obligations on developers of large-scale #AImodels. Companies must now take “reasonable care” to ensure their technologies do not cause severe harm, such as mass casualties or property damage exceeding $500 million. The bill also mandates third-party testing, kill-switch mechanisms, and whistleblower protections. Key legal implications: -Risk Management: Developers must proactively assess and mitigate risks associated with powerful AI systems. -Compliance Exposure: Non-compliance could lead to enforcement actions by the California Attorney General. -Contractual Impact: Legal teams will need to revisit vendor agreements, liability clauses, and governance frameworks to align with the new standards. https://lnkd.in/g-PU-mT4

9

CALIFORNIA'S FRONTIER MODEL TRANSPARENCY BILL SIGNED INTO LAW The Transparency in Frontier Artificial Intelligence Act (TFAIA) or SB53 has come a long way. Its predecessor, the infamous SB1047, was vetoed one year ago. In the wake of that, the CA governor ordered a report on frontier AI policy that would end up laying the foundation for SB53. SB53 applies to publicly available frontier models. A frontier model is a model exceeding 10^26 FLOPs (both pre- and post-training and fine-tuning). The bill requires frontier developers to: - Provide a transparency report pre- or concurrently with deployment - Prepare and publish a Frontier AI Framework that would incorporate national standards, international standards, and industry-consensus best practices. - Provide a summary of any assessment of catastrophic risk resulting from the internal use of the frontier model. The bill provides whistleblower protections and will also create “CalCompute,” a public cloud computing cluster. There is a civil penalty of up to $1M per violation. SB53 comes into force on January 1, 2026. And now a bit of trivia: 32 out of the top 50 AI companies worldwide are based in California.

19

Last week, I sent a letter to Sen. Bill Cassidy and Sen. Bernie Sanders, Chair and Ranking Member of the Senate HELP Committee, urging them to ensure that families are central to #AI governance. This follows the Committee’s recent hearing on “AI's Potential to Support Patients, Workers, Children, and Families.” Two points, in particular, I want to highlight: 1. As the United States develops policies to address AI, we must learn from families—particularly those with minor children—as they are already navigating and adapting to these emerging technologies. Effective AI governance should protect people from harm *and* leverage their lived experiences to craft practical, real-world solutions that meet their needs. 2. As Senator Daniel Patrick Moynihan once wrote, American social policy has long focused on individuals, while neglecting the family as “the fundamental group unit of society” (Universal Declaration of Human Rights, Article 16). That omission is no longer tenable. Especially in the AI era, protecting and strengthening families – as families – is essential to ensuring societal well-being and resilience. The full text of the letter is linked in the comments! 

17

2 Comments