Public Readiness for AI Risks: Capability-Risk Communication Asymmetry Across Frontier AI Labs’ X Accounts and Official Websites

The fields working at the frontier of AI development are not in consensus on what exactly the future of AI will look like — there are proponents of Artificial General Intelligence (AGI), proponents of Artificial Superintelligence (ASI), proponents of coordinated powerful agents as the future of AI, and there are critics and skeptics of these forms of AI, and of AI reaching any super or powerful form of intelligence. In fact, within each sub-community, for example, within the AGI sub-community, there is no consensus on what this form of super AI would look like. What must be attained for AGI to be AGI (McLean et al., 2023) ?

Nonetheless, there is at least a consensus that powerful AI is arriving soon. Different timelines have been proposed for the arrival of powerful, broadly diffused AI (Altman, 2025; Amodei, 2024; Angelo, 2026; Larsen et al., 2026), and with these predicted timelines are further predictions regarding how much broadly diffused, powerful AI could reset our world as we currently know it. Our world is not new to technology-based resets; the computer, personal computers specifically, mobile phones, the internet, and even social media have had global, multi-faceted influences on our world. Yet, it is said that the effects of broadly diffused, powerful AI — even the effects of AI as we see them currently — will likely outdo and supersede the effects of every other type of technology-based resets our world has seen (Gairola, 2026; Hassabis, 2026; Prakash, 2023).

It is established that powerful AI will arrive soon, be broadly diffused, and have global, deep-seated effects. Some researchers have modeled pathways through which advanced AI or artificial superintelligence could produce global catastrophe, including human extinction (Barrett & Baum, 2017; McLean et al., 2023). The expected risks of powerful, insanely capable, and broadly diffused AI are such that the industry statement on mitigating AI risk requests a level of prioritization comparable to the work done in mitigating risks from other global, societal-scale risks ((Center for AI Safety, 2023). Humans have faced global crises stemming from scalable risks in the past. Historically, we have often been under-prepared, such that these crises catch us off-guard and have far-reaching, devastating effects on the global populace. A classic, relatively recent case in point is the world’s preparedness and response to the Covid-19 pandemic (COVID-19 National Preparedness Collaborators, 2022). We also love to claim that we learn from history to better prepare for future risks. However,  we may as well be on the infamous path we took regarding other global risks, when it comes to AI. If powerful AI were to be diffused tomorrow, or in the next few months, or even a couple of years from now, the world would not be ready. This is evident in several of the things happening currently.

There is an asymmetry in information dissemination, such that the bulk of the world’s peoples are starkly ignorant of the capabilities of even today’s frontier models, or of how far AI models and agents have come in capabilities on both short-term and long-horizon tasks (METR, 2026). Juxtaposed against this deep dearth of information, which I sometimes refer to as people simply just being aware of broader, low-level capabilities such as AI content creation and AI image generation, is the deeply aware but small and enclosed community, which I have described in the past as the cyclical AI bubble, and which Andrej Karpathy seemingly also referenced within his broader thesis on the growing gaps in the understanding and perception of AI capability. I would argue that Karpathy’s growing gaps thesis is applicable to AI risks too, such that the majority of the people, and even more so, a good percentage of people who are aware of the understanding and perception of AI capability, are unaware of the extent to which AI risks extend.

On the other hand, there is also the notion of people being misinformed about AI risks. This is evident in the most popular rhetoric about AI risks among the broader populace: “AI will take my job and yours,” “We will be in the permanent underclass,” and the like. Do not get me wrong, economic disempowerment (Kulveit et al., 2025) and labor replacement (Goldman Sachs Research, 2025) are indeed valid AI risks. However, these risks are being platformed and represented in ways that signal doomsday and fatality to the populace, all the more so to the younger and emerging population.

It is true that safety organizations, watchdogs, and think tanks — such as METR, Andon Labs, Apollo Research, and Redwood Research — are doing incredible work to disseminate information regarding AI risks, AI safety, and AI capabilities. Some frontier labs are also doing good work regarding these. However, the fact remains that no one has their fingers on the pulse as much as the frontier AI labs. And, given what the world’s readiness for insanely capable, powerful, and broadly diffused AI is, I would argue that these frontier labs are not doing nearly enough regarding AI risk communication. Better put, this prototreatise builds its argument around the thesis that AI labs (using selected frontier labs as the sample) overrepresent AI capabilities in their public communication, when compared to risk communication. It is not rocket science that these AI labs are incentivized to speak strategically and intensely about AI capabilities. The structural asymmetry between AI capabilities and AI risk communication could be attributed to the benefits of AI capabilities communication, such as lobbying policies, fundraising, signaling value, government backing, IPOing, attracting more enterprise clients, and attracting more users. Essentially, AI labs speaking about AI capabilities — and especially the increasing capabilities of their own models — is a strategic no-brainer. However, since life runs on economics and resources are finite (regardless of how much AI labs have raised) resources, the seeming opportunity cost of AI labs’ communication about AI capabilities is a lack of strategic, broadly diffused, public-readiness risk communication, and its attendant outcome of our world being ill-prepared and unready for powerful, insanely capable, and broadly diffused AI.

Sampling written communications from Open AI, Anthropic, and Google DeepMind across X (Twitter) and the companies’ official websites, this prototreatise examines the state of risk communication regarding AI, and how this compares to the state of AI capabilities communication. Through the synthesized data, I also strengthen the argument that AI risk communication is structurally underrepresented, and that this underrepresentation could be devastating for humanity, if not systematically addressed.

Defining Terms and Concepts

This prototreatise makes use of terms and concepts, some of which are measured or at least evaluated in my data analysis. As such, it is useful to start off by defining and clarifying what I mean by powerful AI, AI capabilities communication, risk communication, and public-readiness risk communication.

Powerful AI within the context of this prototreatise is defined as AI that is insanely capable of executing on nearly all, if not all, the tasks and jobs in the economy, has perpetual knowledge, can continually learn, is contextually and temporally aware, can recursively self-improve, can work for longer horizons matching human workers, and is autonomous and not needing human supervisors or direction. I do not think the current setup — the training architecture, user interface, training data, training priorities, compute, memory, and broader resources — of current LLM models can get us there. However, that is an argument for another piece. Essentially, to then describe broadly diffused powerful AI means that this AI, as described above, is available within the broader population, occasioned by economic, technological, or mixed shifts that have made this possible.

I also need to clarify that AI capabilities communication, as used in this work, refers to strategic communication whose primary function is to foreground what AI systems can do, what they enable users, enterprises, or institutions to do, how their performance is improving, where they can be deployed, or how AI capabilities are advancing. AI capabilities communication could take any format or be further classified into other categories (reports, announcements, research reports, whitepapers, or what have you), however, the most important thing is that this type of communication advances the goal of communicating increasing AI capabilities. As such, communications around model launches, product or feature announcements, benchmark or performance claims, demos, coding, science, agent, enterprise, or productivity claims, multimodal generation claims, robotics or tool-use claims, deployment reports, infrastructure, and ecosystem announcements that paint AI capability as more usable, scalable, deployable, interoperable, or normalized, will fall squarely within AI capabilities communication. This is not necessarily an established subtype of strategic communication. However, it is increasingly seen in communication outputs within the AI industry.

On the other hand, risk communication can be broadly defined as “the exchange of information among interested parties about the nature, magnitude, significance, or control of a risk (Covello, 1992, p. 359). The conceptualization of risk communication within a broader sense recognizes the interactivity of the communication process, wherein individuals, groups, and/or institutions exchange opinions and information regarding risk (National Research Council, 1989) . This also frames risk communication within broader risk governance and public engagement practices. This conceptualization of risk communication is in the third phase within Leiss (1996) documentation of risk communication’s evolution. The first phase was information-heavy and largely focused on the quantitative expression of risk estimates. It favored the mechanical and logical, perhaps on the assumption that everything else was held ceteris paribus, and that the risk estimates would be the core factor for public decision-making. The second phase started to shift away from this technically overwhelming approach towards the elements of successful communication (such as the source’s credibility, message clarity, the clarity of the message, and a focus on the needs and perceived realities of the audience). While this was a welcome shift, it also heavily prioritized elements that could sometimes be nothing but extraneous to the public’s decision-making process. The third phase, in understanding that risk communication is laden with trust issues (and rightly so), relies on the social context presumption that “despite the controversial nature of many risk management issues, there are forces at  work also that favor consensus building, meaningful stakeholder interaction, and acceptance  of reasonable government regulatory framework (Leiss, 1996).”  Essentially, the third phase of risk communication is one where risk communication is not merely persuasive for internal strategic purposes, but one where risk communication is more calibrated towards a shared, overarching goal. This approach has fueled risk communication during global-scale crises such as the Covid-19 pandemic — wherein the overarching, shared goal is to ensure humans are safe and protected from the virus, such that sub-goals (such as companies’ internal strategic ambitions) were subsumed under this shared goal that dictated the direction for communication.

Nonetheless, Reynolds and Seeger (2005) describe risk communication as an effort whose goal concerns “the probabilities of some harm and associated methods for reducing the likelihood of such harm becoming reality (p. 48).” Therefore, risk messages embody the scientific and technical understanding of risk factors and the social and cultural beliefs regarding the said risk, thus giving a holistic view of what the risk is, how it is being manifested, people’s perceptions of the risk, and further, the publics’ likelihood of responding positively or negatively to the risk messages, especially if an element of self-efficacy or public action is necessary for reducing the likelihood of the risk becoming reality (Reynolds & Seeger, 2005). Furthermore, institutions and organizations often deploy risk messages as a way to address the cultural or social factors influencing the perceptions of the said risk, by translating, operationalizing, and then diffusing the technical understanding of the risk and its factors into behaviors, through persuasive techniques . 

Understandably, a company’s outputs can communicate risk while serving other strategic goals. Boholm (2019) noted that institutional risk communication is not always public-serving, as some organizations may use risk communication for other strategic goals such as informing the public, calming and reassuring their audiences, legitimizing their brand, managing trust, or even preserving brand authority. Furthermore, instrumental approaches to risk communication may be structured around securing acceptance of an institution’s preferred risk-management or technology-policy outcome, rather than opening the underlying decision to substantive public deliberation (Wardman, 2008).

The risk communication strategy to be adopted and the effectiveness of risk communication generally depend heavily on the nature of the risk in question. Schweizer et al. (2022) distinguish between two natures of risks: conventional risks and systemic risks. Since conventional risks have a linear path, they can often be effectively addressed through interventions that target their cause-and-effect chain. However, systemic risks are high-complexity, transboundary risks with nonlinear cause-and-effect patterns and probabilistic relationships. As such, addressing systemic risks effectively requires strategic, multi-level, interdisciplinary interventions that sometimes include behavioral, environmental, societal, and even economic sub-interventions. Schweizer et al. (2022) also note that these systemic risks are often volatile with several tipping points, and they are further complicated by the fact that they hardly receive enough public attention. Catastrophic risks are a prime example of this. The Covid-19 pandemic was devastating as a systemic, slow-developing, but seemingly inevitable risk — or at least, the general possibility of a pandemic was foreseeable, even though its timing, origin, and particular manifestation were not (Gertler & the 80,000 Hours team, 2026).

If these systemic risks often fly under the public radar, are volatile, seemingly Hydra-headed, and dangerously devastating, should they become full-blown crises, then, it is necessary for our world to adequately prepare for them. This is the category within which this work classifies the all-encompassing term of AI risk and its embedded risks, such as concentration of power, government takeover and global wars, among others. Furthermore, this prototreatise argues that a lack of risk communication uniquely calibrated towards effectively preparing the public for the myriad of risks that powerful, broadly diffused AI poses is an AI risk in itself, and a high-priority one at that.

Thus, I conceptualize public-readiness risk communication as an interactive process involving risk messages and risk communication efforts that are primarily geared towards providing the different stakeholders and publics with relevant information for informed decision-making, thereby helping them build comprehension, AI literacy skills, resistance to misinformation, disinformation, and malinformation, and preparedness/efficacy for the probability of AI risks becoming reality. This conceptualization of public-readiness risk communication leans heavily towards the peculiarities of systemic and global-scale risks, adapting the pragmatic, skill-building, and self-efficacy orientation of risk models often applicable to health risks and crises such as the CERC, PADM, and IDEA models (see Lindell & Perry, 2012; Reynolds & Seeger, 2005; Sellnow et al., 2017). However, a point of deviation from these otherwise fitting models is that AI risks cannot be really classified as acute risks. They present as slower onset risks that build over time and compound. Thus, while humanity can greatly benefit from AI risk communication that optimizes for public readiness, better awareness of the risks, and adequate workforce and human resource development and preparedness, the difference in the nature and type of risk these communication models and my conceptualization address is not to be trivialized.

Approach

This work sought to answer the question: is risk communication — specifically risk communication as conceptualized within this work (public-readiness-oriented risk communication) — structurally underrepresented in the communication outputs of frontier labs, especially when compared to capabilities communication? To answer that question, I followed a directed qualitative content analysis approach. Hsieh and Shannon (2005) classify directed content analysis as an approach where prior research or theory informs the analysis process. In line with the processes Hsieh and Shannon (2005) highlighted, I first identified concepts that define the scope of my work (capabilities communication, AI risk communication, public-readiness risk communication) as my initial coding categories. Then, these conceptualized terms served as the a priori categories for the analysis (Hsieh & Shannon, 2005). Following this, I developed operational definitions for these categories, which then enabled me to develop a unified codebook.

Data Collection

The data collection focused on written communication artifacts as the unit of analysis. Since there are several AI labs and I do not have the resources to analyze communication from the full AI lab population, I made the methodological choice of sampling three frontier AI labs: OpenAI, Anthropic, and Google DeepMind. People have different criteria for classifying AI labs. In this work, I recognize frontier AI labs based on Longterm Wiki’s classification as a first-level criterion, while also prioritizing labs that have had robust communication across the two sites of analysis (official website and X) within the analyzed period.

Agentic AI, in this case Hermes Agent and OpenAI Codex, were useful for the collection process. For X, I tasked Hermes Agent (with Codex as its harness) with collating written communication from the official accounts of the sampled labs: @AnthropicAI, @GoogleDeepMind, @OpenAI, and @OpenAINewsroom. OpenAI Newsroom had to be included in the accounts for analysis, since OpenAI handles the scope of communication this prototreatise is interested in via two distinct channels. 

The X collation covered posts published between December 1, 2025, and May 31, 2026. A standalone post was treated as one communication artifact. A root post and its connected self-replies were also treated as one artifact. Since the data collation was conducted through API calls to X’s servers, there were instances where the initial source inventory bundled independent X posts into a single record. In such instances, the bundled posts were separated and treated as distinct analytic artifacts. These bundles were identified through a direct audit of the reply relationships among the status URLs contained in each source record. Posts were retained as one artifact only where they formed a connected root-post and self-reply chain. Posts that appeared as independent roots or belonged to separate reply chains were separated and treated as distinct artifacts. Differences in subject matter were not, by themselves, used to divide an artifact. The resulting analytic corpus contained 326 X artifacts: 96 from Anthropic, 87 from Google DeepMind, and 143 from OpenAI’s two accounts.

For each X artifact, the collected fields included the lab, account, publication month, root-post URL, URLs for all posts contained in the artifact, the directly visible written content, source format, number of posts in the artifact, and a description of the communication. The directly visible written content incorporated the root post, connected self-replies, and visible link-card text. Source format, thread length, media, and outbound link information were retained as metadata but were not subsequently used as measures of prominence.

The collected X metrics were root-post views, aggregate views across the artifact, likes, reposts, quotes, and comments. The collation process ran between June 22 and June 29, 2026. Metrics were captured in batches on June 28 and June 29, with June 29, 2026, serving as the final metrics date. Therefore, the figures represent metric snapshots at the time of collection. Furthermore, the website analysis ran until August 4, 2026. Therefore, the current elevation criterion for website prominence reflects data for this date.

For the website corpus, I tasked Codex with compiling the official publications issued by the three labs during the same December 1, 2025, to May 31, 2026 period. Publications were collected from the official OpenAI, Anthropic, and Google DeepMind websites using their publication surfaces, official sitemaps, page metadata, and additional official publication registers. The collection covered dated, substantive publications appearing on surfaces such as News, Research, Safety, Engineering, Product, Global Affairs, model cards, and system cards.

For each website publication, the collected fields included the lab, publication date, official section, artifact type, title, official URL, official summary, and available publication text or page metadata. Duplicate URLs were removed, publication dates were checked against the study period, and the collected URLs were reconciled with the review register. The website review register contained 280 pages. Of these, 277 were dated publications falling between December 1, 2025, and May 31, 2026, and were included in the analytic corpus: 89 from Anthropic, 63 from Google DeepMind, and 125 from OpenAI. Three undated OpenAI reference or hub pages were retained for review but excluded from the volume analysis because they could not be assigned to the study period.

The website and X corpora were retained as separate datasets. A publication appearing on a lab’s website and subsequently promoted on X remained one website artifact and one X artifact because the two records represented communication on different sites of analysis. 

Coding Scheme

In this work, I examine the presence (or otherwise) of structural asymmetry between capabilities and risk communication in the AI sector using the communicative goal of an artifact, the overall volume (of output) per communicative goal, and the overall prominence (of output) per communicative goal as evaluative criteria. Since the argument concerns the structural underrepresentation of public-readiness risk communication, volume provided the most direct measure of how much of the sampled communication was devoted to public-readiness risk communication, relative to capabilities communication. Prominence added a second dimension by showing the relative salience of each communication category. It therefore allowed me to determine whether the observable differences in output volume across all categories were reinforced or further complicated by the observable prominence of each category, across both channels, and per lab. A category could, for example, be small in volume but comparatively prominent, or substantial in volume but comparatively less prominent.

Communicative Goal

The principal coding decision concerned the artifact’s predominant communicative goal. Predominance was determined from the title or opening statement, the proportion and specificity of the message devoted to each possible goal, its conclusion or call to action, its apparent intended audience, and what the communication ultimately asked that audience to know, value, judge, adopt, or do. Where a given artifact had more than one communicative goal, the dominant or overarching communicative goal was adopted. For instance, if an artifact promoting OpenAI’s strategic collaboration with Nvidia to ramp up ChatGPT capabilities had a paragraph spotlighting recent risks, the piece would be considered primarily AI capabilities communication. 

Thus, I conducted the coding process in two stages. The first stage assigned a provisional communicative goal code to each artifact and identified cases requiring further review. Following the refinement of the operational definitions of the work’s concepts, the two corpora were re-read against a unified codebook. The final record for each artifact contained its communicative goal, a confidence assessment, a coding rationale and, where applicable, its public-readiness classification and prominence level. Differences between the provisional and final decisions were retained in the adjudication records.

The coding did not rely on the occurrence of particular words or on the institutional section in which an artifact appeared. Therefore, the presence of safety language did not automatically make an artifact risk communication, just as the presence of capability evidence did not automatically make it capabilities communication. Three predominant goal categories were then applied: AI capabilities communication, AI risk communication, and Other communication.

An artifact was coded as AI capabilities communication when its primary function was to foreground what an AI system could do, enable, improve, automate, scale, deploy, interoperate with, or make more usable. This included model and product launches, demonstrations, benchmark and performance claims, coding, science, agent and productivity claims, deployment reports, infrastructure announcements, ecosystem developments, and accounts of AI adoption or normalization. Safety or risk material contained within these artifacts did not alter the code, where it primarily reassured audiences that the capability being presented was safe, aligned, trustworthy, or ready for use.

The AI risk communication category was assigned to artifacts whose primary function concerned the nature, magnitude, consequences, distribution, governance, mitigation, or control of risks created or amplified by AI. This included communication about model behavior and alignment, misuse, AI-enabled cyber or biological harm, misinformation and manipulation, labor and economic disruption, concentration of power, autonomy and loss-of-control pathways, and broader social or institutional consequences. Ordinary software incidents, account-security notices, service disruptions, legal or corporate statements, and conventional product support were not coded as AI risk unless the artifact itself established a substantive connection to risks created or amplified by AI.

Other communication covered artifacts whose predominant goal was neither the communication of AI capability nor the communication of AI risk. This included corporate appointments, awards, events, grants, philanthropy, ordinary organizational affairs, conventional operational notices, and other institutional communication falling outside the principal comparison.

The work then extended into sub-categories within the risk communication category. The presence or absence of a public-readiness focus (as explained in the terms section) was assessed only after an artifact had been coded as AI risk communication. Thus, an AI-risk artifact qualified as public-readiness risk communication when it met two conditions. First, it had to concern a risk associated with powerful or broadly diffused AI, or a present-day AI risk that informed that broader trajectory. Second, its primary function had to involve preparing or equipping the public to understand, judge, anticipate, or respond to that risk.

The readiness function could involve building comprehension, AI literacy, resistance to misinformation, disinformation or malinformation, anticipatory preparedness, protective or civic efficacy, workforce or economic preparedness, governance capacity, or meaningful participation in decisions concerning AI risk. It did not have to contain a direct instruction, but it had to provide an intelligible public-facing layer that equipped the public for comprehension, judgment, preparedness, or response (see image below for examples).

The public availability of risk information did not by itself make an artifact public-readiness communication. Technical safety research, expert-facing evaluations, institutional frameworks, policy signaling, organizational mitigation claims, and institutional reassurance were coded as non-readiness AI risk when they lacked the required public-facing readiness function. Likewise, an actionable instruction did not by itself establish public readiness. A conventional software update, password reset, product recall, service notice, or isolated security remediation did not qualify without a substantive connection to AI-caused or AI-amplified risk.

Essentially, artifacts initially classified as AI risk communication were then further classified as either public-readiness AI risk communication or non-readiness AI risk communication. The four final display categories were therefore AI capabilities communication, non-readiness AI risk communication, public-readiness AI risk communication, and Other communication.

Establishing each artifact’s communicative goal was the primary coding task required to then execute on the volume dimension. Once the communicative goal enabled me to take the volume count for each category and then calculate the volume percentages as needed, the next task concerned coding prominence. Since observable data on the labs’ websites and X differ in the data properties available and what can be reliably measured, I operationalized prominence differently for both channels.

Nonetheless, the important factor in thinking about prominence was to extend the concept of structural representation of any given category beyond its volume count or volume percentage. I will discuss the operationalization of prominence across both channels subsequently.

Editorial Prominence on the Website

First, I defined website editorial prominence as the observable emphasis that labs’ websites gave to a given publication, relative to the likelihood that website visitors would see that publication. Since this emphasis cannot be easily quantified due to the differences in organizational culture and how they affect website architecture, the dynamic state of content types, the iterative, frequently changing nature of go-to-market strategies, and incessantly changing brand priorities at labs (which is not uncommon in the tech industry, and the AI sub-industry more specifically), I examined website editorial prominence from three separate sub-dimensions. These sub-dimensions are: historical elevation, current elevation, and structural discoverability. The sub-dimensions were not recalculated into a composite score because they represent different forms of editorial emphasis and are not necessarily equivalent. Each artifact’s official website section and content type were descriptive metadata, although neither field the publication’s communicative category or automatically established its editorial prominence. Furthermore, website editorial prominence should be strictly interpreted as organization-controlled opportunities for the labs’ communications, as represented by each artifact; it is neither a proxy for page visits, page events, and other website analytics data, nor is it an avenue to infer the audience response or broader communication effects of any of the analyzed artifacts.

In analyzing, the primary function of historical elevation within the prominence dimension was to answer the question: was the artifact observed on the homepage or a major editorial page in the available captures surrounding its publication? This question is a derivative of the prevailing rule that each lab’s homepage is its definitive pinnacle of editorial placement and, therefore, content prominence (as it is for most, if not all, brands). It also then considered pages that are directly accessible from the homepage or persistent global navigation; serve a broad, organization-recognized publication function; are maintained as a principal discovery surface; and/or do not require a narrow topical filter to reach them as major editorial pages.

Using the recovered evidence for each artifact’s publication, Codex’s analysis then classified each artifact into one of the following: homepage-direct placement, major-page placement, deeper or non-major page placement, or no positive elevation was recovered. Insufficient evidence was a flag for cases in which the available captures did not permit reliable classification. Historical absence from the recovered evidence was therefore not treated automatically as proof that a publication had never been elevated.

On the other hand, current elevation sought to answer this question: on the measurement date, was the publication displayed on the homepage or a major editorial page? This also adopted the prevailing rule and major-page rules earlier established for the historical elevation sub-dimension. Codex’s instructions bound the data analysis to the links observed on live homepages or major pages only. Also, a publication currently observed on the homepage or a major page necessarily also satisfies the condition of having appeared there at least once, that is, the historical elevation criterion.

Finally, structural discoverability concerned the shortest observed route from the homepage to the publication. Path 1 indicated that the publication was linked directly from the homepage. Path 2 indicated that at least one intervening page or section had to be reached before the publication. Path 3+ indicated that three or more page transitions or other necessary discovery actions were required to eventually reach the publication. Opening a navigation menu or scrolling did not count as an additional transition. However, necessary page filter (such as was observed with OpenAI’s research page), pagination control (as was observed with most of Google DeepMind’s pages), or “Load more” (as was observed with Anthropic’s pages) did did count. Where no route was recovered from the examined website architecture, the publication was recorded as accessible only through its direct URL or through search. Of course, this designation does not account for all the plausible routes that may lead towards the page; essentially, a publication or artifact being designated as a direct URL/search does not necessarily mean it is an orphan page.

Finally, there was the consideration made for the nature of prominence. This was only applied to publications that were currently observed on the homepage or a major page. Within this subset, Codex was instructed to record whether a given currently featured artifact received hero or lead treatment, prominent feature treatment, ordinary visible listing treatment, or exposure only after an additional action.

Furthermore, I retained each publication’s official section and content type as metadata, although neither determined its communicative category nor automatically established its editorial prominence. Essentially, website editorial prominence represented organization-controlled opportunities for encountering a publication, thus serving as the cleanest way to measure prominence, in the absence of internal data on page visits, bounce rate, average time spent on the page, page events, or other communication effects.

Prominence and Interaction Measures on X

On X, I evaluated prominence using the platform’s observable visibility and interaction metrics. Root-post views showed the visibility of the post that initiated an artifact, while aggregate views represented the combined views received by all the posts within an artifact. Likes, reposts, quotes, and comments captured different forms of interaction with the artifact. I also calculated total engagement by adding likes, reposts, quotes, and comments; engagement rate by dividing total engagement by aggregate views; and amplification rate by dividing reposts and quotes by aggregate views. These metrics were read alongside one another in examining the prominence received by each communicative category. They were not combined into a single prominence score.

However, to place individual artifacts into comparable High, Medium, and Low prominence tiers, I relied on root-post views only. Thus, I compared the root-post views received by each artifact with those received by the other artifacts published by the same account. Artifacts within the top 25% were classified as having high prominence. Those within the middle 50% were classified as having medium prominence, while those within the bottom 25% were classified as having low prominence. For OpenAI, posts from @OpenAI and @OpenAINewsroom were first assessed within their respective accounts before being combined for the lab-level analysis. Thus, an artifact’s account was only used to determine the other artifacts with which it should be compared; being published by a particular account did not itself make the artifact prominent.

Where an artifact was a thread, its prominence tier was determined from the views received by its root post. Its aggregate views and interaction metrics were still included in the broader prominence analysis. Therefore, the prominence tier captured an artifact’s relative visibility within its account, while the remaining metrics provided additional measures of the visibility and interaction it received.

Results

This prototreatise set out with an argument that there is a structural asymmetry between communication about AI risk and communication about AI capabilities, considering that frontier AI labs are the best positioned to communicate with the public regarding both. This section now turns to whether, and to what extent, the collected data support that argument. The findings are organized around the following questions:
What did the selected labs communicate during the period of analysis?

• How did this communication vary by lab and channel?
• How much of the risk communication was public-readiness risk communication?
• How prominent was each type of communication?
• What did the qualifying public-readiness communication look like?
• What do the findings say about the argument tabled at the outset?
  

1. Volume

1. 1 Selected frontier AI labs’ communication during the period of analysis

Website publications by all three labs within the period of analysis totaled 277, while X posts totaled 326. There is a relative closeness in output volume across both channels, especially given that website content will, on average, be longer and more in-depth than X posts. Following the communicative goal criterion, I classified artifacts into the following categories: AI capabilities communication, non-public-readiness risk communication, public-readiness risk communication, and Other. On the labs’ website, capabilities communication accounted for 60.3% (167) of the total publications; non-readiness risk communication accounted for 27.1% (75); public-readiness risk communication accounted for 7.6% (21); and Other accounted for 5.1% (14). Public-readiness risk communication is structurally underrepresented across these three sampled labs when output volume is used as the evaluative criterion (see the image below), such that even when we account for public-readiness risk communication as a share of the total risk communication published on these labs’ websites within the given period, it only accounts for 21.9% (see appendix section).

The analysis revealed that 63.2% (206) X artifacts across all four accounts, and within the period of analysis, fall within the capabilities communication category; 12.0% (39) artifacts fall within the non-readiness risk communication category; 5.2% (17) fall within the public-readiness risk communication category; and 19.6% (64) fall within Other. Similar to the observation with website publications, public-readiness risk communication is severely underrepresented across the sampled labs, when output volume is the evaluative criterion, such that even within a broader risk communication classification, public-readiness risk communication only accounts for 30.4% of all outputs (see appendix section).

When artifacts from both sites of analysis are pooled together (totaling 603 artifacts), AI capabilities communication still accounts for a sizeable percentage of output volume, at 61.9% (373), non-readiness AI risk communication accounts for 18.9% (114), public-readiness AI risk communication accounts for 6.3% (38), while other accounts for 12.9% (78).

Regardless of the foregrounding in this piece, an argument can be made against conceptualizing public-readiness risk communication as the focal point of evaluation. In that case, risk communication would be examined in its entirety. If that were the case, AI capabilities communication would account for 60.3% (167) of the website publications, while AI risk communication would account for 34.7% (96) of the outputs. Similarly, AI capabilities communication would account for 63.2% (206) of X artifacts, while AI risk communication would only account for 17.2% (56). See the image below for further breakdown. Essentially, the argument that AI capabilities communication is overrepresented in AI labs’ communication compared to AI risk communication is validated when we examine the volume of communication output across two of the most widely used communication channels in the AI industry/sector.

1.2: Capabilities Communication vs. Public-Readiness Risk Communication: A Lab-level Analysis

Given the general findings revealing the underrepresentation of public-readiness risk communication (or any form of risk communication at all) in frontier AI labs’ output volume, it is important to examine communication behavior at the individual level, which further helps us understand the severity of the problem.

Anthropic

Anthropic published 89 website communications content within the period of analysis, 58.4% (52) of which fall under capabilities communication, with 28.1% (25) being non-readiness risk communication, 9% (8) being public-readiness risk communication, and 4.5% (4) being classified as Other. On the other hand, Anthropic’s X account had 96 artifacts within the analysis period, of which 37.5% (36) fall under capabilities communication, 25% (24) are non-readiness risk communication, 10.4% (10) are public-readiness risk communication, and 27.1% (26) are classified as Other. While Anthropic’s website publications align with the broadly noted asymmetry between capabilities and risk communication, X artifacts seem to have a relatively close representation between capabilities communication and risk communication (given the total risk communication from Anthropic’s X artifacts is at 35.4%, boosted heavily by non-readiness risk communication).

Essentially, while Anthropic makes considerable efforts to publish more risk communication content on X, comparable to its communication about Claude’s capabilities, the larger share of this risk communication is still not content designed to improve or contribute to public readiness and preparedness for insanely powerful and broadly diffused AI.

Google DeepMind

Google DeepMind published 63 website communications content within the period of analysis, 68.3% (43) of which fall under capabilities communication, with 28.6% (18) being non-readiness risk communication, 3.2% (2) being public-readiness risk communication, and nothing classified as Other. The website publications from Google DeepMind are both in line with the asymmetry findings from the general communication mix and the website findings from Anthropic. Similarly, Google DeepMind’s X account had 87 artifacts within the analysis period, of which a staggering 83.9% (73) fall under capabilities communication, while 3.4% (3) are non-readiness risk communication, 3.4% (3) are public-readiness risk communication, and 9.2% (8) are classified as Other. Google DeepMind’s stance has been overwhelmingly focused on strategically communicating AI’s increasing capabilities.

OpenAI

OpenAI has the highest number of artifacts published within the analysis period and across both channels. The company published 125 website communications content within the period of analysis, 57.6% (72) of which fall under capabilities communication, with 25.6% (32) being non-readiness risk communication, 8.8% (11) being public-readiness risk communication, and 8.0% (10) being classified as Other. On the other hand, OpenAI’s two X accounts had 143 artifacts within the analysis period, 105 of which were posted by the main handle @OpenAI, and 38 of which were posted by @OpenAINewsroom. Of the 143 artifacts, 67.8% (97) are 8.4% (12) are non-readiness risk communication, 2.8% (4) are public-readiness risk communication, and 21.0% (30) are classified as Other. Similar to Google DeepMind, OpenAI’s X artifacts were overwhelmingly about capabilities communication, at the detriment of both public-readiness and non-readiness risk communication. Still, the level of asymmetry varied by account.

OpenAI, which is the official account (and which had more artifacts for the given period), had a staggering share of its output within the capabilities communication classification, at 74.3%, compared to the 5.7% of non-readiness risk communication published and the 2.9% of public-readiness risk communication published. On the other hand, 50% of the X artifacts from the OpenAINewsroom account fall under the capabilities classification, with 15.8% being non-readiness risk communication, and 2.6% being public-readiness risk communication.

Furthermore, an angle of findings worth considering in discussing the results at the lab-by-lab level is the structural underrepresentation of public-readiness AI risk communication. Although the output volume for AI risk communication seemed comparatively level with AI capabilities communication and other forms of communication for some accounts (Anthropic’s X and website communication, and OpenAI’s website communication, especially come to mind), a comparative analysis of public-readiness risk communication and blanket, general AI risk communication shows the paucity of public-readiness outputs (see image below).

However, this structural underrepresentation of public-readiness AI risk communication is not uniform across all labs, from channel to channel. Essentially, the frontier labs’ investment in content type via a channel could signal priorities that are not generalizable to the labs’ overall communications, or more so, priorities that could be misread for what they are.

Anthropic’s AI capabilities communication shifted from a higher level of output volume on its website (58.4%) to a much lower level (37.5%). This also has the biggest downward shift in AI capabilities communication among the sampled labs. Similarly, Anthropic’s risk communication saw a downward turn from 37.1% on the website to 35.4% on X, although this is not nearly as huge a drop as capabilities communication. Non-readiness risk communication likewise saw a slight downward turn from 28.1% to 25%. However, public-readiness AI risk communication saw a slight upward turn, 9% to 10.4%. So, the data shows that Anthropic’s AI capabilities and AI risks communication on their X account took a smaller share of their total output volume, compared to the share of output volume these same categories took in their website publications.. However, the significant downward turn in the AI capabilities output could make it so that if the public interacts with Anthropic’s communications more on X than any other platform, the public perception about Anthropic’s emphasis on risk communication (or even public-readiness AI risk communication specifically), could be higher than it is (see Image 12 below).

For Google DeepMind, the case is starkly different. The lab’s AI capabilities communication saw an upward turn from an already high 68.3% on the website to 83.9% on X, also marking the highest percentage-point increase in AI capabilities communication among the sampled labs. On the other hand, Google DeepMind’s AI risk communication saw a downward shift in output volume from website to X (31.7% to 6.9%), caused mainly by a similar shift in non-readiness risk communication (28.6% to 3.4%) — also the biggest shift in AI risk communication of all sampled labs — even though there was a slight upward shift in public-readiness communication (3.2% to 3.4%). All figures are illustrated in Image 13 below.

Drawing a parallel to the Anthropic data, if the public mainly interacts with Google DeepMind via its X account, the big upward shift in capabilities communication and downward shift in AI risk communication could shape public perceptions towards thinking Google DeepMind’s AI risk communication is non-existent, a perception that would be misleading, seeing as Google DeepMind has a comparable percentage of risk communication with Anthropic, when the website is considered.

OpenAI’s case is quite different from the other two labs. Although the lab’s AI capabilities communication saw a considerable upward shift (57.6% to 67.8%), other categories saw varying downward shifts: AI risk generally saw a two-figure percentage-point drop, from 34.4% on the website to 11.2% on X; non-readiness AI risk communication saw a shift from 25.6% to 8.4%, while public-readiness AI risk communication saw a shift from 8.8% to 2.8%. Furthermore, while OpenAI does not generally have the lowest level of public-readiness risk communication across all channels, it has the lowest level of public-readiness risk communication on X (see Image 14).

Essentially, Anthropic’s capabilities communication output volume on X does not match its website’s output volume, whereas the reverse is true for the other labs. On the other hand, OpenAI’s public-readiness risk communication output volume on X does not match that on its website, whereas the reverse was true for the other labs.

Having examined the communication output volume from all these dimensions, it is also important to draw a parallel between the dichotomies of this prototreatise’s argument. When public-readiness risk communication is compared with AI capabilities communication, the structural underrepresentation discussed so far becomes starkly evident (see the image below). The magnitude of the asymmetry varied considerably; however, there was no lab-channel combination where public-readiness risk communication exceeded capabilities communication in output volume. Furthermore, the smallest gap, indicating the smallest volume-based underrepresentation, appeared in Anthropic’s X communication, at 27.1 percentage points, while the largest gap appeared in Google DeepMind’s X communication, at 80.5 percentage points.

2. Prominence

Overrepresentation and underrepresentation are concepts that cannot be tied solely to output volume. For instance, it is possible to have published five sponsored/promoted and high-performing public-readiness X posts, whose combined reach and influence on the conversation pulses could be equivalent to 20 AI capabilities posts. In that case, one cannot straightforwardly claim an underrepresentation of public-readiness risk communication. As such, I also analyzed the prominence of artifacts published by these three frontier AI labs across both sites of analysis, as copiously explained earlier.

2.1 Website Editorial Prominence

These findings do not assign cardinal figures or scores to the sub-dimensions, since editorial prominence was operationalized such that its sub-dimensions are not necessarily equivalent. Furthermore, the results will be presented both at the broader level and also at the individual, lab-based level.

2.1.1 Historical Placement

The data analysis revealed that across all three labs, 43.1% (72) of AI capabilities communication artifacts were directly placed on the homepage, one way or the other, within their publication window, while 47.3% (79) were placed on a major page, and 9.6% (16) had no recorded placement in the captures, in wich case we cannot reliably determine if they were placed on the homepage or a major page during the publication window or not.

On the other hand, only 12% (9) of risk communication artifacts were directly placed on the homepage within their publication window, although 74.7% (56) were placed on a major page, while 13.3% (10) had no recorded placement in the captures. Similar to AI capabilities communication, no figures were recorded for placement on a deeper or non-major page.

For public-readiness risk communication, 23.8% (5) of all artifacts were placed on the homepage within their publication window, while a larger 66.7% (14) were placed on a major page, and 9.5% (2) had no recorded placement captures.

Furthermore, Other had 42.9% (6) of artifacts with a homepage placement within their publication window, while 50% (7) were placed on a major page, and 14.3% (2) had no recorded placement captures. Worthy of note is that none of the four categories had deeper or non-major-page placements for any of their artifacts, as observed within each artifact’s publication window.

The data is interesting on different counts. First, public-readiness risk artifacts have a comparable percentage share of category total representation, relative to AI capabilities communication (23.8% vs. 43.1%, 66.7% vs. 47.3%); this is also the case for non-readiness risk communication (12% vs. 43.1%, 74.7% vs. 47.3%), and especially for major-page placement. Thus, if we side with those who would consider risk communication as risk communication — without the public-readiness risk communication sub-concept — these percentage shares from the (total) risk communication category will even be more comparable to AI capabilities communication (see image below). Since historical placement could be further extended as an anecdotal prediction of future placements, one can essentially say that at any given time, a better percentage of all risk communication artifacts created within a given period across all three labs go on a major website page, compared to the percentage of all capabilities communication artifacts created within the same period, ergo, no structural underrepresentation of one category (or any of its sub-categories).

However, that does not provide a full understanding of the issue, particularly the interrelatedness of volume and prominence in gauging structural underrepresentation or overrepresentation. This historical placement data still reinforces the position that AI capabilities communication receives much more representation in AI labs’ communication efforts, and therefore, public-readiness risk communication is underrepresented. We can look at it this way: comparable or even lesser percentage shares of total AI capabilities communication received homepage or major-page placement within their observed publication window, relative to other categories like non-readiness AI risk communication and public-readiness AI risk communication, yet, given the structural overrepresentation of AI capabilities communication through output volume, it is such that in absolute values, the homepage is 14.4x more likely to have AI capabilities communication artifacts than it is to have public-readiness AI risk communication artifact, while major pages are 5.64x more likely to have AI capabilities communication than public-readiness AI risk communication. Even when all forms of risk communication are considered, AI capabilities communication artifacts are still 5.14x more likely to be on the homepage than AI risk communication artifacts, while major pages are 1.13x more likely to have AI capabilities communication artifacts than AI risk communication artifacts. Note that all of these are examined still within the defined historical placement; that is, when the metadata of these artifacts were examined, which ones received the aforementioned placements within seven days of publication? And if we maintain the anecdotal extension applied earlier, we can say that at any given time, and at the pinnacle of website editorial prominence, AI capabilities communication is structurally more represented than AI risk communication (or, specifically, public-readiness AI risk communication). The image below illustrates this further.

Furthermore, it is important to examine this historical placement at the lab level. In so doing, we can see that lab-specific arguments can be made regarding the structural underrepresentation (or otherwise) of AI risk communication. Figures stated in this category are within the context of the full dataset, wherein although the results only report homepage and major-page placement, the data accounts for other contexts, such as a lack of positive capture within the examined seven days, and deeper or non-major pages.

Anthropic

For Anthropic, 11.5% (6) of all AI capabilities communication artifacts were historically placed on the homepage, while an overwhelming 86.5% (45) were historically placed on major pages. The figures are quite different for non-readiness AI risk communication, with no historical homepage placements and an overwhelming 72% (18) of placements on major pages. Similarly, no public-readiness AI risk communication content had historical homepage placements, although that can be attributed to 100% (8) of the pieces rather making it to major pages (see image below).

Looking at the figures and already accounting for disproportionate output volume (wherein Anthropic published AI capabilities content at 2.08x the number of non-readiness AI risk content and 6.5x the number of public-readiness risk content), it can be said that Anthropic’s website artifacts received fairly comparable homepage and especially major-page placements within the first seven days of publication, regardless of their category.

Google DeepMind

For Google DeepMind, the bulk of AI capabilities communication artifacts, at 67.4% (29), were historically placed on the homepage, while 32.6% (14) were historically placed on major pages. The figures are quite different for non-readiness AI risk communication, with just 11.1% (2) historically placed on the homepage, and an overwhelming 88.9% (16) historically placed on major pages. On the other hand, public-readiness risk communication artifacts were all historically represented on the homepage, even though there were only two (see image below).

Looking at the figures and already accounting for disproportionate output volume (wherein Anthropic published AI capabilities content at 2.08x the number of non-readiness AI risk content and 6.5x the number of public-readiness risk content), it can be said that Anthropic’s website artifacts received fairly comparable homepage and especially major-page placements within the first seven days of publication, regardless of their category.

References

1. Aaron Gertler, & 80,000 Hours team. (2026, 05 15). Preventing Catastrophic Pandemics. 80,000 Hours. Retrieved 2026-07-31. https://80000hours.org/problem-profiles/preventing-catastrophic-pandemics/
2. Amodei, D. (2024, October). Machines of loving grace: How AI could transform the world for the better. https://darioamodei.com/essay/machines-of-loving-grace
3. Ananya Gairola. (2026, 02 21). Demis Hassabis Predicts AGI Will Have 10x The Impact Of The Industrial Revolution — And It Will Happen In A Decade, Not A Century. Yahoo Finance. https://finance.yahoo.com/news/demis-hassabis-predicts-agi-10x-143113425.html
4. Anthony M. Barrett, & Seth D. Baum. (2017). A model of pathways to artificial superintelligence catastrophe for risk and decision analysis. Journal of Experimental & Theoretical Artificial Intelligence, 29(2), 397–414. https://doi.org/10.1080/0952813x.2016.1186228
5. Barbara Reynolds, & Matthew W. Seeger. (2005, 01). Crisis and Emergency Risk Communication as an Integrative Model. Journal of Health Communication, 10(1), 43–55. https://doi.org/10.1080/10810730590904571
6. Center for AI Safety. (2023, 05 30). Statement on AI Risk. Israeli Association for Ethics in Artificial Intelligence. Retrieved 2026-07-30. https://www.iaeai.org/tpost/xa4htcma61-statement-on-ai-risk
7. COVID-19 National Preparedness Collaborators. (2022, 04 16). Pandemic preparedness and COVID-19: An exploratory analysis of infection and fatality rates, and contextual factors associated with preparedness in 177 countries, from Jan 1, 2020, to Sept 30, 2021. The Lancet, 399(10334), 1489–1512. https://doi.org/10.1016/s0140-6736(22)00172-6
8. Deanna D. Sellnow, Derek R. Lane, Timothy L. Sellnow, & Robert S. Littlefield. (2017, 10 20). The IDEA Model as a Best Practice for Effective Instructional Risk and Crisis Communication. Communication Studies, 68(5), 552–567. https://doi.org/10.1080/10510974.2017.1375535
9. Demis Hassabis. (2026, 07 14). A Framework for Frontier AI and the Dawning of a New Age. X. Retrieved 2026-07-30. https://x.com/demishassabis/article/2076957440109625718
10. Goldman Sachs Research. (2025, 08 13). How Will AI Affect the Global Workforce? Goldman Sachs. Retrieved 2026-07-31. https://www.goldmansachs.com/insights/articles/how-will-ai-affect-the-global-workforce
11. Hsiu-Fang Hsieh, & Sarah E. Shannon. (2005, 11). Three Approaches to Qualitative Content Analysis. Qualitative Health Research, 15(9), 1277–1288. https://doi.org/10.1177/1049732305276687
12. Jake Angelo. (2026, 05 16). Microsoft AI chief gives it 18 months—for all white-collar work to be automated by AI. Fortune. https://fortune.com/article/why-microsoft-ai-chief-mustafa-suleyman-predicts-ai-automation-18-months/
13. Jamie K. Wardman. (2008, 12). The Constitution of Risk Communication in Advanced Liberal Societies. Risk Analysis, 28(6), 1619–1637. https://doi.org/10.1111/j.1539-6924.2008.01108.x
14. Jan Kulveit, Raymond Douglas, Nora Ammann, Deger Turan, David Krueger, & David Duvenaud. (2025, 01 28). Gradual Disempowerment: Systemic Existential Risks from Incremental AI Development [Preprint]. arXiv. https://doi.org/10.48550/arxiv.2501.16946
15. METR. (2026, 05 08). Task-Completion Time Horizons of Frontier AI Models. Retrieved 2026-07-30. https://metr.org/time-horizons/
16. Michael K. Lindell, & Ronald W. Perry. (2012, 04). The Protective Action Decision Model: Theoretical Modifications and Additional Evidence. Risk Analysis, 32(4), 616–632. https://doi.org/10.1111/j.1539-6924.2011.01647.x
17. National Research Council. (1989). Improving Risk Communication. National Academies Press. https://doi.org/10.17226/1189
18. Pia-Johanna Schweizer, Robert Goble, & Ortwin Renn. (2022, 07). Social Perception of Systemic Risks. Risk Analysis, 42(7), 1455–1471. https://doi.org/10.1111/risa.13831
19. Prarthana Prakash. (2023, 04 17). Alphabet CEO Sundar Pichai says that A.I. could be ‘more profound’ than both fire and electricity—but he’s been saying the same thing for years. Fortune. https://fortune.com/2023/04/17/sundar-pichai-a-i-more-profound-than-fire-electricity/
20. Sam Altman. (2025, 06 10). The Gentle Singularity. Sam Altman's Blog. Retrieved 2026-07-30. https://blog.samaltman.com/the-gentle-singularity
21. Scott McLean, Gemma J. M. Read, Jason Thompson, Chris Baber, Neville A. Stanton, & Paul M. Salmon. (2023). The risks associated with Artificial General Intelligence: A systematic review. Journal of Experimental & Theoretical Artificial Intelligence, 35(5), 649–663. https://doi.org/10.1080/0952813x.2021.1964003
22. Thomas Larsen, Romeo Dean, Brendan Halstead, Eli Lifland, Ryan Greenblatt, & Daniel Kokotajlo. (2026, 07 09). AI 2040: Plan A. AI Futures Project. https://ai-2040.com/
23. Vincent T. Covello. (1992, 01). Risk Communication: An Emerging Area of Health Communication Research. Annals of the International Communication Association, 15(1), 359–373. https://doi.org/10.1080/23808985.1992.11678816
24. William Leiss. (1996, 05). Three Phases in the Evolution of Risk Communication Practice. The ANNALS of the American Academy of Political and Social Science, 545(1), 85–94. https://doi.org/10.1177/0002716296545001009
25. Åsa Boholm. (2019, 08). Risk Communication as Government Agency Organizational Practice. Risk Analysis, 39(8), 1695–1707. https://doi.org/10.1111/risa.13302