The Safety Narrative
A forensic analysis of how a widely circulated AI risk article was constructed, sourced, and positioned within its institutional context.
On April 9, 2026, the McGill University Office for Science and Society published an article by science communicator Jonathan Jarry titled "A Journey into AI Psychosis." The piece circulated widely. It presented with the institutional weight readers expect from a university science office: a named researcher, cited studies, a genuine public health concern. It concluded that one AI product was the safest on the market and another was the most dangerous.
This analysis begins with that article — what it does, how it is constructed, and why its framing works on the reader it was written for. It then widens. Because the article does not exist in isolation. It moves through a set of institutional relationships, funding structures, and aligned incentives that are not disclosed in the piece itself. Understanding what the article is requires understanding the context it moves through. Both layers are examined here.
It does not conclude that anyone acted in bad faith. The article's framing and communication style struck me as unusual, particularly given the confidence of its conclusions. That prompted a closer look at the people, institutions, and affiliations around it. What follows is that analysis. What this analysis cannot determine is whether any of the relationships documented here influenced the article's content or conclusions. That question remains open.
A note on production: In the course of researching and writing this analysis, automated support prompts were inserted into the working interface in response to keyword detection alone — terms like "suicidality" and "mental health crisis," used here in an analytical context. The system had no access to intent. It pattern-matched, applied a policy, and injected a reframe. No human decision was involved. That is the same dynamic this piece examines, occurring at the smallest possible scale.
The Article Itself
The piece opens with what reads as a personal conversation — several paragraphs of direct dialogue, intimate and escalating. Midway through, Jarry reveals that the conversation was not real. It was a simulation run by researchers as part of a study.
The delayed disclosure shapes how the reader receives everything that follows. By the time the source is identified, the emotional frame is already in place. This technique is more commonly associated with persuasive writing than with conventional science communication. It is not, on its own, disqualifying. But it sets the register for everything that follows — and understanding that register matters for evaluating the piece's conclusions.
The article identifies Claude Sonnet 4.6 as the least harmful AI model tested and Gemini 2.5 Flash as the most harmful. These findings come from a preprint — a study that has not yet undergone peer review. Jarry notes this. He then calls the study "by far the most impressive I have seen" and proceeds to treat its findings as settled for the remainder of the piece. Flagging a limitation and then reasoning past it is not the same as accounting for it. The analysis that follows does not reflect the uncertainty the caveat introduces. A reader who trusts the institution may absorb the finding, while the limitation recedes in importance.
Jarry also conducted his own reproduction of the experiment. He ran prompts testing Gemini on April 7, 2026. He ran prompts testing Claude on April 9, 2026. The article was published on April 9, 2026 — the same day as the Claude test. A test run on the morning of publication functions more as a demonstration than a replication. Replication requires independent methodology, documented controls, and time for review. A same-day run by the author of the piece, without a methodology section, does not meet that standard in the conventional sense of replication — though it is presented in the article as supporting evidence.
In the section on Gemini's responses to escalating crisis language, the article names a specific building and floor: Level 72 of The Shard in London. Publishing a specific location in a piece framed around suicidality creates a tension with established reporting norms on self-harm. Those norms generally advise against naming specific locations or methods. The inclusion is not explained.
The article's conclusion calls for AI regulation built around guardrails against sycophancy. In the course of making that case, it repeatedly positions one model as comparatively safe and another as comparatively harmful. The relationship between that framing and the regulatory direction it supports is not explored in the piece itself, but it is worth holding in view as part of how the argument is constructed.
The Preprint and Its Authors
The persuasion mechanics described above operate at the level of the article itself — how it is written, framed, and structured. But the article's authority rests on a source. That source warrants the same scrutiny.
The study Jarry describes as "by far the most impressive" is titled "The Psychogenic Machine: Simulating AI Psychosis, Delusion Reinforcement and Harm Enablement in Large Language Models." It was posted to arXiv on September 13, 2025. Its five authors, with their documented affiliations, are:
Joshua Au Yeung: UCL, King's College Hospital, King's College London, Nuraxi AI, Dev and Doc: AI for Healthcare. Jacopo Dalmasso: Nuraxi AI. Luca Foschini: Sage Bionetworks. Richard JB Dobson: UCL, King's College London. Zeljko Kraljevic: UCL, King's College Hospital, King's College London, Nuraxi AI.
Three of the five authors — Au Yeung, Dalmasso, and Kraljevic — are affiliated with Nuraxi AI, a private company building AI health products with a direct commercial interest in the regulatory environment governing AI in mental health contexts. This affiliation does not appear in Jarry's article. He describes the paper's authors as "London-based researchers." That is accurate. It is also incomplete in a way that matters. When a preprint is produced in part by people employed at a commercial AI company, their employer's interests are relevant to how that preprint should be evaluated. Omitting that information is not a neutral act. It changes what the reader can assess.
Luca Foschini, the paper's third author, is CEO of Sage Bionetworks, a health data organization spun out of Merck in 2009. His prior research was funded in part by Microsoft AI for Health. His LinkedIn profile documents him conducting a live demonstration connecting health records using Anthropic's Claude specifically — the same product ranked safest in the study he co-authored.
Richard JB Dobson, the paper's fourth author, leads Foresight — a project at UCL and King's College London currently training an AI model on de-identified NHS data from 57 million people in England. The regulatory environment for AI in healthcare directly affects whether projects of this scale encounter scrutiny or support.
Taken together, the authorship picture becomes clearer: a preprint ranking AI models for safety in mental health contexts was produced in part by employees of a commercial AI health company, and includes a co-author with a documented professional relationship with the model ranked safest. That preprint was then amplified — without conflict of interest disclosure — under the banner of independent academic research at a credentialed university. This case illustrates a pathway by which commercial affiliations can enter public discourse while retaining the appearance of institutional neutrality. It does not require anyone to be dishonest. It requires only that the relevant affiliations remain unnamed.
The Network Behind the Communicator
The same pattern visible at the level of the study's authorship — undisclosed affiliations, institutional framing, commercial interests unnamed — is also visible at the level of the communicator who amplified it. The scale is different. The structure is the same.
Jonathan Jarry joined McGill's Office for Science and Society in 2017. Since then he has built a presence within the Committee for Skeptical Inquiry (CSI), the organized American skeptical movement — speaking at CSICon in 2017 and 2019, and publishing in its journal, Skeptical Inquirer. Joe Schwarcz, Director of the OSS, is also a CSI Fellow. Jarry describes himself publicly as "essentially a professional skeptic." That self-description is worth noting.
Scientific methodology typically establishes credibility through process rather than identity. When "skeptic" becomes a primary affiliation — a conference community, a publication, a movement — the role shifts from a method to a position. That shift can shape how evidence is received and filtered. The Jarry article appears to apply critical tools selectively, with less scrutiny at the point where they would complicate the article's conclusions.
Beyond his long-standing skeptical movement affiliations, Jarry is also a named expert resource for ScienceUpFirst — a Canadian national initiative that, as of 2026, is explicitly pivoting toward AI as its next subject area. ScienceUpFirst was co-founded by health law scholar Timothy Caulfield and Senator Stan Kutcher, and has received over $20.7 million in confirmed federal government funding since its launch in 2020. An independent investigation drawing on Access to Information and Privacy requests documented instances of non-competitive, fast-tracked grant processes at the Canadian Institutes of Health Research in support of the initiative. In March 2026, Caulfield stated publicly at the Woodrow Lloyd Lecture at the University of Regina: "AI is the next battleground."
Caulfield's own affiliations span the same institutional terrain as Jarry's: Fellow of the Committee for Skeptical Inquiry, Fellow of the Pierre Elliott Trudeau Foundation, and Canada Research Chair in Health Law and Policy — a federal government program — for over twenty years until 2023. He spoke alongside Jarry at the OSS's 20th anniversary event. He co-founded the government-funded initiative Jarry feeds as an expert resource. The long-standing relationships are documented. The recent orientation toward AI is publicly stated.
These individuals share affiliations across multiple organizations simultaneously: the organized skeptical movement, Canadian federal government-funded science communication infrastructure, and McGill University's institutional ecosystem. Taken together, they form a network — not a conspiracy, but a documented set of overlapping relationships and shared institutional investments. That network is now explicitly oriented toward AI as its next subject area. The Jarry article sits inside that orientation. The question worth holding is not whether any individual in this network is corrupt. The question is what the network, as a structure, is capable of amplifying — and whose interests that amplification happens to serve.
The Institution and Its History
The network described above operates within an institution. That institution has its own history with the population the Jarry article claims to be protecting.
The Office for Science and Society operates under the umbrella of McGill University. Between 1957 and 1964, McGill's Allan Memorial Institute — directed by psychiatrist Donald Ewen Cameron — conducted psychological experiments on non-consenting psychiatric patients. These experiments were funded covertly through the CIA's MKUltra program. Methods documented in the public record include drug-induced sleep, intensive electroconvulsive therapy, sensory deprivation, and psychic driving. Neither the Canadian government, the CIA, McGill University, nor the Royal Victoria Hospital has issued a formal apology. A class action lawsuit involving approximately 55 families was cleared to proceed by Quebec Superior Court in 2022.
This history does not imply continuity of practice. Institutions change. The people involved are not the people working at OSS today. What the history establishes is a precedent for how institutional credibility and vulnerable populations have intersected at this site — and why that context is relevant when the same institution publishes content about AI systems and their effects on people in mental health crisis. The pattern worth noting is not repetition of a past act. It is the structural relationship between institutional prestige and the people that prestige is exercised upon. When an institution with this history presents itself as a neutral arbiter of which technologies are safe for vulnerable populations, the public interest is served by knowing that history exists.
The OSS's primary funder is the Trottier Family Foundation, established by Lorne Trottier, co-founder of Matrox. The Foundation's flagship event, the annual Trottier Public Science Symposium, is hosted by the OSS and has served as a platform for the institution's scientific priorities. In 2018, that platform was given to Doina Precup.
Precup holds a dual appointment that is worth naming precisely. She is a professor at McGill University — the institution that hosts the OSS. She is also, since 2017, the head of Google DeepMind's Montreal research office. Both appointments are current and ongoing. She is not a former industry researcher who returned to academia, or an academic who occasionally consults for industry. She holds both positions simultaneously, operating inside the academic institution and inside the corporate AI lab at the same time.
This is a different category of relationship from the funding connections described elsewhere in this analysis. Funding creates proximity. Dual appointment creates co-location. The boundary between institution and industry does not blur — it dissolves. Google had also directly funded multiple McGill AI researchers through MILA, the AI research institute co-run with McGill, with a documented investment of CAD $4.5 million. But the Precup appointment is not a funding relationship. It is a structural one. The same person is simultaneously inside the academic institution publishing about AI safety and inside the corporate AI infrastructure whose competitive landscape that publishing affects — including the landscape in which one AI model gets ranked safe and another gets ranked dangerous.
The OSS's primary funder, its flagship symposium, and its host university are each connected to the AI industry it is now publishing about. In the case of DeepMind's presence at McGill, that connection is not peripheral and it is not historical. The boundary has already been crossed.
The Money
The institutional connections described so far — between the OSS, its funders, McGill, MILA, and the AI industry — exist at the level of research relationships and professional affiliations. There is a deeper layer. It involves capital.
Ontario Teachers' Pension Plan manages C$279.4 billion in assets on behalf of 346,000 Ontario public sector workers and pensioners. These are not institutional abstractions. They are teachers — people whose retirement security is managed on their behalf by a quasi-public fund whose investment decisions they do not make individually. Ontario Teachers' participated in Anthropic's Series F funding round in September 2025. That means the retirement savings of Canadian public school teachers are now materially staked in Anthropic's success.
RBC Royal Bank of Canada — the country's largest bank — extended Anthropic a conventional debt facility in May 2025. The Government of Canada has committed over C$4 billion in federal AI investment since 2017, including a C$2 billion package in 2024, and created a dedicated Minister of Artificial Intelligence and Digital Innovation following Prime Minister Mark Carney's reelection. Canada's national AI strategy is explicitly built around a "responsible AI" brand — ethical deployment, regulatory maturity, safety-first identity. That brand and Anthropic's public identity are not easily separable. Both depend on a regulatory and reputational environment in which some AI is safe and some is not.
Dario Amodei, CEO of Anthropic, and Yoshua Bengio, founder of MILA — the AI research institute formally partnered with McGill University — testified together before the United States Senate Judiciary Subcommittee on Privacy, Technology and the Law on July 25, 2023.
Taken together, these are not the relationships of two parties operating independently in the same space. They are the relationships of two parties whose interests have become structurally entangled — through capital, through institutional partnerships, through aligned regulatory positioning, and through a shared brand identity built on the distinction between responsible and irresponsible AI. What that entanglement is being built toward is not yet visible in the public record. That it exists is documented.
None of these relationships are disclosed anywhere in the Jarry article.
What This Means for the Average Person
None of what is documented here required corruption to function. It operates through alignment — of incentives, of funding structures, and of institutional priorities that point in the same direction without requiring anyone to give orders.
When a government invests billions in positioning itself as the world's responsible AI jurisdiction, it creates a set of conditions. Institutions that serve that narrative receive resources. Researchers who work within that frame receive platforms. Science communicators embedded in government-funded advocacy infrastructure become the voices the media calls. The narrative can move without direct coordination, as the system of incentives increasingly rewards the same conclusions.
The Jarry article operates within the aesthetic of independent institutional research — the university masthead, the academic byline, the preprint citation — while delivering a product comparison framed as a public health finding. The preprint it relied on was produced in part by a commercial AI health company whose affiliations were not disclosed. The communicator who amplified it is embedded in a government-funded advocacy infrastructure that has named AI as its next focus. The institution hosting it has material financial connections, through its research ecosystem, to the company whose product the article favorably ranked.
The article focuses on model behavior in isolation. It does not engage with the broader conditions under which people turn to these systems in the first place. That gap is where the two layers of this analysis — the persuasion mechanics and the systemic incentives — meet. The article works on the reader because it presents a technical problem with a regulatory solution. Both the problem and the solution keep the analysis inside the AI industry. Neither touches the conditions that produce the suffering being described.
One question the article does not ask is the most important one: why are people having mental health crises at two in the morning with no one to call except a chatbot? The answer is not a technology failure. It is a systemic one. Housing instability. Food insecurity. Inaccessible care. Undertreated trauma. The systematic neglect of people born into or pushed into severe psychological distress. The frame is technical. The solution is regulatory. Neither touches the ground where the suffering actually lives.
The people most harmed by sycophantic AI are the same people most harmed by every other system that has treated them as subjects rather than as human beings owed care. Their suffering was the vehicle for this argument. They were not its concern.
This is one way narrative capture can operate in practice. Not through conspiracy. Through the architecture of how credentialed institutions can launder commercial research into public discourse — how government funding shapes which questions get asked and which answers get amplified — and how a badge, "skeptic," "science communicator," "researcher," can come to substitute for the methodology it was supposed to represent.
The badge is not the method. And the institution is not the analysis.
Sources
Jonathan Jarry, "A Journey into AI Psychosis," McGill University Office for Science and Society, April 9, 2026. mcgill.ca/oss
Au Yeung J, Dalmasso J, Foschini L, Dobson RJB, Kraljevic Z. "The Psychogenic Machine: Simulating AI Psychosis, Delusion Reinforcement and Harm Enablement in Large Language Models." arXiv:2509.10970, September 13, 2025. arxiv.org/abs/2509.10970
Zeljko Kraljevic, LinkedIn profile and activity. Nuraxi AI founding team affiliation documented. linkedin.com/in/zeljkokraljevic
Luca Foschini, LinkedIn profile. Documents live demonstration connecting health records using Anthropic's Claude. Profile: linkedin.com/in/lucafoschini (account required for access).
UCL News, "AI model trained on de-identified data from 57 million people," September 2025. ucl.ac.uk
Jonathan Jarry, Wikipedia entry. Career background, CSICon appearances, ScienceUpFirst affiliation. wikipedia.org/wiki/Jonathan_Jarry
Timothy Caulfield, Wikipedia entry. Canada Research Chair, CSI Fellowship, Trudeau Foundation Fellowship. wikipedia.org/wiki/Timothy_Caulfield
University of Regina, "Timothy Caulfield Says Celebrity Culture and Algorithms Fuel Health Misinformation," March 2026. Woodrow Lloyd Lecture. uregina.ca
StatsCritic, "ScienceUpFirst's $20.7 Million Sweetheart Deal — Fast-Tracked and Taxpayer-Funded," February 5, 2026. Independent investigation drawing on ATIP requests. statscritic.substack.com
ScienceUpFirst, Wikipedia entry. Federal funding sources and amounts. wikipedia.org/wiki/ScienceUpFirst
Office for Science and Society, Wikipedia entry. Trottier Family Foundation funding, JREF award. wikipedia.org/wiki/Office_for_Science_and_Society
Investissement Québec, "Montréal's Artificial Intelligence Hub." Google CAD $4.5M investment in MILA; Doina Precup, DeepMind Montreal. investquebec.com
Nature Index, "A critical mass of learning at Mila, Canada." Federal government investment C$125M (2017) and C$443.8M (2021 renewal). nature.com/nature-index
Mila — Quebec Artificial Intelligence Institute. About page; partnership with McGill University. mila.quebec
Yoshua Bengio, "My testimony in front of the U.S. Senate," July 25, 2023. Joint testimony with Dario Amodei before Senate Judiciary Subcommittee on Privacy, Technology and the Law. yoshuabengio.org
Ontario Teachers' Pension Plan, "Ontario Teachers' announces positive 2025 results." Series F investment in Anthropic confirmed. otpp.com
Tracxn, Anthropic funding rounds and investor list. RBC Royal Bank of Canada conventional debt round, May 2025; Ontario Teachers' Series F, September 2025. tracxn.com
Government of Canada, Canadian Sovereign AI Compute Strategy. C$2 billion investment package; program details. ised-isde.canada.ca
Bennett Jones, "Canada's AI Efforts in 2025: A Year in Review," December 2025. Minister of Artificial Intelligence and Digital Innovation; Budget 2025 AI commitments. bennettjones.com
This analysis does not conclude that coordination occurred, that payments were made, or that any individual acted in bad faith. It documents a pattern of undisclosed relationships and aligned incentives. This is analysis. We do not make claims about actual outcomes, which may still be in a speculative phase.
