Skip to main content
AI Jobs·Report

ANALYSISReported by AI Jobs Report

Safety alarms grow; only 39 safety roles in 3,624 AI openings

This week’s posts flag rising alignment risk and rapid adoption, while our tracker shows hiring remains concentrated in engineering and data, with few evaluation and safety roles.

Read the original at AI Jobs Report
10 min read8 viewsBy AI Jobs Report

This week’s posts flag rising alignment risk and rapid adoption, while our tracker shows hiring remains concentrated in engineering and data, with few evaluation and safety roles.

The week’s split-screen: louder risk, faster uptake, builder-heavy hiring

A cluster of posts in the past two days highlight two forces pulling in opposite directions. On one side, writers emphasize mounting alignment risks and overreach in claims about AI’s impact on medicine and society. On the other, they point to rapid diffusion and new technical frontiers. Our jobs tracker sides mostly with the builders: of 3,624 open AI roles we count across 142 employers, only 39 are in evaluation and safety, while engineering and data dominate.

Zvi Mowshowitz calls Anthropic’s new transparency a positive surprise even as he flags alarming details: "I am grateful that Anthropic is producing periodic Risk Reports." He adds, "The other revelation is the existence of the world’s likely best model, ‘Model 2.’"

Ethan Mollick sees adoption accelerating in unusual places and risk management moving up the agenda. He writes, "AI diffusion in political campaigns is quite high (and it looks like Anthropic’s fight with the Pentagon may have been good for its uptake among politicians, through it is impossible to establish a definitive cause for its rapid rise)" and, on safety, "If alignment issues are becoming big enough in their new models that OpenAI is willing to commit 20% of research inference compute to chain-of-thought monitoring, that suggests that alignment issues are becoming a serious concern." He also notes, "Some early evidence that AI may actually be accelerating discoveries in the areas where you would expect to see AI acceleration happening first" and points to cognitive capabilities: "The fact that LLMs have theory of mind at all is a huge deal (and was controversial when discovered in GPT-4), but you can still see the limitations when models need to consider more than one audience - like how they struggle to separate end user & creator needs/perspectives when coding or writing."

Azeem Azhar puts numbers to the market: "Since our report in June, revenues have continued to grow, with this July sitting three times higher year-over-year. The annualized run-rate is now over $210 billion." He is also expanding his own effort to study the economy of AI: "Exponential View is appointing its first Research Fellow." The role, he writes, will turn questions into "testable economic mechanisms, assumptions and back-of-the-envelope estimates."

Jack Clark focuses on capabilities measurement: "The new frontier for analyzing AI systems is understanding how good they are at inferring the unwritten rules of their environment…", highlighting DiG-bench’s hidden-rule games as a probe of intuition and discovery.

Nathan Lambert argues the locus of open development is drifting from open weights toward fully open recipes: "The open-source language model – i.e. only models that come with a full training recipe, data, code, etc. – is a closer analogue to the open-source operating system." That has operational implications: the recipe, he says, is "a resource intensive process that any company can pick up, modify, and press “run” on to produce a new set of model weights."

Simon Willison amplifies reporting on the scramble for training data. He notes, "Online forum discussions between Amazon workers confirmed that VGT3 destructively scans large volumes of books."

Gary Marcus pushes back on hype and diagnoses strategic drift. On Google, he writes, "Google is clearly on the back foot in AI; it should, by rights, be dominating AI." On healthcare, he challenges Dario Amodei’s timeline as naive, opening with, "Perfect example of egregious CEO Says journalism" and adding, "And I wish it were true. But his timeline is so naive as to be absurd, both with respect to how medical science works and why it is challenging." He also says public sentiment is cooling: "Brief update to last’s week’s note re: the growing backlash."

What our tracker sees in the hiring market

Our tracker counts 3,624 open AI roles across 142 employers, taken from their own job feeds. The mix is weighted to building and operating systems rather than policing them:

  • Modelling and engineering: 1,382 open, 1,579 opened and 197 closed in 30 days, across 113 employers.
  • Data: 730 open, 800 opened and 70 closed in 30 days, across 97 employers.
  • Infrastructure: 328 open, 381 opened and 53 closed in 30 days, across 70 employers.
  • Research: 201 open, 223 opened and 22 closed in 30 days, across 44 employers.
  • Product and design: 89 open, 101 opened and 12 closed in 30 days, across 47 employers.
  • Evaluation and safety: 39 open, 46 opened and 7 closed in 30 days, across 15 employers.

The top hiring employers we track are a mix of services, finance, biotech, and core AI labs:

  • Accenture: 596 open, 20 opened in 30 days.
  • Capital One: 179 open, 8 opened in 30 days.
  • Amgen: 173 open, 5 opened in 30 days.
  • OpenAI: 148 open, 57 opened in 30 days.
  • Waymo: 105 open, 14 opened in 30 days.
  • Anthropic: 103 open, 20 opened in 30 days.
  • Databricks: 89 open, 18 opened in 30 days.
  • PwC: 86 open, 6 opened in 30 days.

Our headline jobs-created indicator is counted from job listings and calibrated to the World Economic Forum's Future of Jobs Report 2025 (11M AI roles created, 9M displaced, by 2030).

Where commentary aligns with the labor signals

  • Growth and diffusion. Azhar’s point on revenue expansion is consistent with the continued appetite for talent. Across core build functions, openings far outnumber closures over the past 30 days. That is most obvious in modelling and engineering, where 1,579 roles opened against 197 closed, and in data, where 800 opened against 70 closed. Mollick’s observations about accelerating discoveries and broader uptake are in tune with this net-adding posture.

  • Open recipes need data and infra. Lambert’s claim that open-source recipes are resource intensive is reflected in demand for the enabling functions. Data roles stand at 730 open across 97 employers, and infrastructure at 328 across 70. The scale of data hiring also rhymes with Willison’s amplified reporting about large scale scanning for training data. We cannot speak to specific practices, but the hiring footprint indicates that data acquisition, preparation, and management remain central activities.

  • Labs still growing. Zvi’s focus on Anthropic’s disclosures lands alongside active hiring at frontier labs. OpenAI lists 148 open roles, with 57 opened in the past 30 days. Anthropic lists 103 open, with 20 opened in 30 days. Clark’s emphasis on new benchmarks for discovery fits with 201 open research roles across 44 employers.

Where the posts outrun or contradict our data

  • Alignment is a headline concern but a sliver of jobs. Mollick says alignment issues are becoming “a serious concern.” Zvi describes Anthropic’s report as containing "some of it rather alarming." Our tracker shows only 39 open evaluation and safety roles across 15 employers. That is a small fraction of the 3,624 openings, especially next to 1,382 in modelling and engineering. If risk is rising, employers are not yet matching that with large scale safety hiring.

  • Political campaign diffusion is not visible here. Mollick reports that campaign uptake is high. Our tracker aggregates roles from 142 employers and the top hiring list skews to services, finance, biotech, autonomous systems, and AI labs. We do not see a signal we can attribute to political campaign hiring in these figures.

  • Healthcare hype meets steady pharma hiring. Marcus rejects the idea that it "will actually be possible to cure most human disease in ~5-10 years" and calls the timeline naive. Our data do not speak to outcomes. They do show that industry investment continues: Amgen lists 173 open AI roles. That evidences demand, not clinical impact, so it does not resolve the disagreement Marcus raises about timelines.

  • Google’s standing is not confirmed by hiring here. Marcus writes, "Google is clearly on the back foot in AI." Google is not among our current top employers by open roles. We do not infer more than that from this data, but it does not contradict his broader claim of relative slippage.

  • Premium model spend plateaus are not yet reflected in role mix. Azhar reports that usage of a top model has flattened as a share of tokens and spend. Our role mix suggests firms are still prioritizing core build and data capabilities over product and design, with 89 open product and design roles across 47 employers. That could be consistent with cost optimization and internal tooling focus, but we cannot directly connect model usage patterns to hiring from these figures alone.

What this means for job seekers and policymakers

  • Builders dominate. If you are a modeller, software engineer, or data specialist, demand remains broad based across 113 to 97 employers in those families. Infrastructure is also active across 70 employers.

  • Safety hiring is thin. If the commentary is right that alignment risk is mounting, we would expect evaluation and safety openings to grow from the current 39 across 15 employers. We will watch that number.

  • Labs and biotech are hot spots. OpenAI and Anthropic remain in the top tier of hiring labs. Biotech’s presence via Amgen indicates continued sector investment even amid skepticism about near term medical breakthroughs.

  • Research is steady but not exploding. With 201 open research roles across 44 employers, there is healthy demand, which aligns with Clark’s focus on new measurement frontiers and Zvi’s interest in risk reports. But safety-specific research still looks niche in our data.

  • The economy narrative needs grounding in labor demand. Azhar’s $210 billion run-rate framing and Mollick’s diffusion claims are consistent with net role creation in our tracker. The window into what companies actually prioritize is clearer here: staffing builders over safety and product polish.

We will continue to test the week’s claims against movements in our counts. The most valuable divergence today is between escalating alignment rhetoric and the still small share of evaluation and safety roles. If that gap closes, we will see it first in the job feeds.

What we read

Every quote above is taken verbatim from one of these posts.

More on this

More from ANALYSIS