Commentators pushed safety, regulation and agents, but our tracker shows most AI hiring remains in modelling, data and infrastructure, with few roles in evaluation and safety.
A safety turn in commentary, a deployment push in practice
Across the past two days, a common thread is concern about AI risks and a call to act. Noah Smith says "Surprisingly, the issue hasn’t fallen victim to partisan polarization yet." Emily M. Bender highlights a concrete policy stance with Timnit Gebru: "It is high time for us to regulate usage of systems sold as “AI”, not because they are magical machines, but because they are error-prone products that shouldn’t be used in high stakes scenarios." Gary Marcus frames law enforcement as overdue, quoting Zephyr Teachout’s bottom line that "There’s plenty of evidence akin to car crashes."
Alongside that, warnings about agents intensified. Simon Willison quoted Matthew Green: "Put these pieces together and you have the two halves of a worm: a payload that hijacks the agent, and an agent that will carry the payload to the next agent."
Yet the same 48 hours were also about stronger models and faster deployment. Zvi Mowshowitz’s verdict on Anthropic’s latest is blunt: "The new hotness remains Claude Opus 5.5. This model rocks. It has made me considerably more productive and made my day more pleasant. It should raise your ambitions." Ethan Mollick, revising his earlier expectations about orchestrated agent teams, writes: "Nope." He adds: "I fell prey to The Bitter Lesson, the hard truth, learned over and over again, that things that we thought required elaborate human rules and thinking can be solved with the brute force of better machine learning systems and more AI." On capacity in a concrete job family, Mollick also shares: "We find that on medium-length, well-defined accounting tasks, frontier AI models are now faster and more accurate than junior accountants, even the best one in our study."
Zvi also reports on high-level coordination: "We left with a White House agreement that is nonzero Actual Progress rather than a step backwards." He says the accord "called upon them to complete the Quest for Embedded Evaluators."
So the argument in short: more risk awareness and regulatory appetite on one side, more capable models and agents on the other. Here is how our hiring tracker cuts through that.
Our tracker counts 4,325 open AI roles across 180 employers, taken from their own job feeds. Our headline jobs-created indicator is calibrated to the World Economic Forum’s Future of Jobs Report 2025 (11M AI roles created, 9M displaced, by 2030).
Does hiring back up the safety turn?
Not yet. Evaluation and safety roles sit at 48 open positions across 20 employers. Over the last 30 days, just 13 of those roles opened and 2 closed. That is a small corner of the market compared to modelling and engineering, which has 1,713 open roles, with 1,043 opened and 819 closed in 30 days across 138 employers. Data roles stand at 763 open, infrastructure at 474 open, and research at 235 open.
This is a plain mismatch with the volume and urgency of the safety rhetoric. Smith argues risk has gone mainstream; Bender and Gebru argue for limits in high stakes settings; Teachout says enforcement should proceed where there is "evidence akin to car crashes." Our data does not show a surge in safety hiring to match those priorities. If anything, safety appears underweighted relative to deployment.
Zvi’s account of a White House process that is "nonzero Actual Progress" and a push for the "Quest for Embedded Evaluators" would, if it were translating into staffing, show up as growth in evaluation and safety openings. We do not see that in the last 30 days.
Are frontier models shifting work and productivity?
Here our data is more aligned with what Zvi and Mollick report. Zvi says "This model rocks. It has made me considerably more productive," and Mollick argues that newer systems reduce the need for elaborate prompting and agent choreography: "planning steps have much less value."
The hiring signal tilts toward build and rollout, not pure research. Modelling and engineering dominates with 1,713 open roles. Data roles are 763 open. Infrastructure is 474 open. Research, by contrast, is 235 open. Product and design is 102 open. That mix suggests organizations are investing most in integrating, scaling, and maintaining systems, which is consistent with a deployment moment in which improved models are taken to production and embedded in workflows.
Top employers reinforce that. Accenture lists 665 open AI roles with 626 opened in 30 days, and PwC lists 101 open with 66 opened. Capital One has 216 open with 268 opened in 30 days. These are the firms that operationalize AI across many business functions. That pattern supports Mollick’s claim that practical, end-to-end application is advancing. It does not confirm his specific accounting result, and our tracker does not isolate accounting teams, but the concentration of openings in consulting and financial services is consistent with automation and augmentation spreading through white-collar workflows.
Agents, worms, and who is staffing for it
Willison’s post highlights Matthew Green’s warning about agent-to-agent propagation. Zvi flags the arrival of "Dots, your always-on AI agent based on Astra," signaling more ambient, persistent agents in users’ hands. If firms were reacting to those risks at scale, we would expect to see more safety and evaluation roles. Instead, evaluation and safety remains at 48 open roles, with limited recent opening activity. Infrastructure has 474 open roles, which could include reliability and some security-related work, but our taxonomy does not let us reclassify those as safety. On the evidence we have, hiring is not yet catching up to the agent risk narrative.
Whose languages and whose priorities?
Alex Hanna writes, "Indigenous languages are often hidden to LLMs, for good reason," and describes a conversation about "Indigenous data sovereignty" and "why it’s worth being skeptical when the “AI" industry says it cares about preserving Indigenous languages." Bender notes interviewing Keoni Mahelona: "What a thrill to get to interview Keoni Mahelona of Te Hiku Media for Mystery AI Hype Theater 3000!"
Our tracker does not break out roles for community language work or data sovereignty. The top employers over the period are Accenture, Capital One, OpenAI, Amgen, Anthropic, Waymo, PwC, and Databricks. That does not contradict Hanna and Bender, but it offers no evidence that hiring is pivoting toward the priorities they highlight. On the contrary, the largest hiring engines are consultancies, finance, labs, and platform companies.
Policy talk versus policy jobs
Marcus opens his interview by calling it "timely," and Teachout’s claim that "I joined the current AI debate because I saw way too many people treating repeat evidence of lawbreaking as wholly unreachable by current law" points to enforcement, not just discussion. Zvi, for his part, warns that "They Are Going To Try This Moratorium Insanity Again During the Lame Duck Session."
Our dataset does not include a legal or policy category, and we do not see evaluation and safety expanding. That means we cannot confirm any move from policy talk to policy staffing inside companies. The balance of investment, as reflected in open roles, remains tilted toward technical buildout.
What this means for where AI jobs are growing
Two conclusions stand out when we test this week’s arguments against our numbers:
-
Safety is center stage in commentary, but a niche in hiring. Out of 4,325 open roles, only 48 are in evaluation and safety. Over the past 30 days, 13 opened and 2 closed. That lags far behind modelling and engineering, data, and infrastructure. The call to "complete the Quest for Embedded Evaluators" is not yet material in posted jobs.
-
Productivity and deployment claims are reflected in who is hiring. Zvi’s personal productivity gains and Mollick’s view that agents now need less handholding align with a market that is staffing model engineers, data teams, and infrastructure at scale, and with consultancies and banks sitting among the top employers.
There are also places where our data cannot arbitrate. Dwarkesh Patel’s historical analogy to conquest is thought provoking: "We often forget how wild human history is." and "People talk about AI takeover as if it would be some crazy, unprecedented sci-fi outcome." Our tracker measures employer demand for AI roles; it does not address geopolitical analogies or public risk perception.
Finally, the hiring pulse remains hot at a handful of large integrators and labs. Accenture’s 665 open roles and 626 opens in 30 days, Capital One’s 216 open and 268 opened, OpenAI’s 180 open and 48 opened, and Anthropic’s 137 open and 40 opened point to rapid iteration in services and platforms. For readers tracking where AI jobs are actually created, that is the momentum to watch, even as the debate shifts toward regulation, safety, and social priorities. The commentary is clear and, in many cases, urgent. The hiring markets, for now, are clearer still.
What we read
Every quote above is taken verbatim from one of these posts.
- Zvi Mowshowitz: AI #188: Gemini Dot Argon, A ‘Morally Binding’ White House Accord on AI Safety
- Noah Smith: Roundup #89: It isn’t X, it’s Y
- Ethan Mollick: The Dot and the Swarm, 'We find that on medium-length, well-defined accounting tasks, frontie
- Dwarkesh Patel: Si Sheppard – How did a few hundred Spanish soldiers topple two empire
- Gary Marcus: Can companies like OpenAI keep getting away with what they are doing?
- Emily M. Bender: What a thrill to get to interview Keoni Mahelona of Te Hiku Media for , There's a long discussion going on in my mentions about whether or not, [New from @timnitgebru.blacksky.app and me:
'It is high time for us to](https://bsky.app/profile/emilymbender.bsky.social/post/3mwsxbeju2kxj)
- Alex Hanna: Indigenous languages are often hidden to LLMs, for good reason. We spe
- Simon Willison: Quoting Matthew Green