Skip to main content
AI Jobs·Report

ANALYSISReported by AI Jobs Report

Lots of safety talk, just 40 safety jobs in our tracker

Commentators flagged pauses, surveillance and hype; employer job feeds show continued expansion in engineering and data, with minimal growth in evaluation and safety roles.

Read the original at AI Jobs Report
9 min read2 viewsBy AI Jobs Report

Commentators flagged pauses, surveillance and hype; employer job feeds show continued expansion in engineering and data, with minimal growth in evaluation and safety roles.

A split-screen week: pause, surveillance, product tweaks, and contradictions

Zvi Mowshowitz surveyed a “quiet aftermath” and declared that “OpenAI is attempting to turn its ship around.” Gary Marcus warned that “This certainly raises privacy concerns.” Simon Willison spotted a concrete product change: “ChatGPT search now uses the site:operator at scale.” Emily M. Bender and Alex Hanna aimed their show at sweeping claims of AI transcendence. And Ethan Mollick summed up the mood with: “Its going to be an era of contradictions.”

Across these posts, the shared thread is a tension between calls for caution and criticisms of hype on one side, and evidence of continued, practical product iteration on the other. Our job-listings tracker provides a way to test which side is showing up in hiring.

What employers are actually hiring for

Our tracker counts 4,137 open AI roles across 174 employers, taken from their own job feeds. By role family, the demand is dominated by hands-on build and data work:

  • Modelling and engineering: 1,619 open, 376 opened and 106 closed in 30 days, across 143 employers
  • Data: 800 open, 180 opened and 19 closed in 30 days, across 118 employers
  • Infrastructure: 388 open, 63 opened and 18 closed in 30 days, across 90 employers
  • Research: 224 open, 32 opened and 4 closed in 30 days, across 49 employers
  • Product and design: 99 open, 20 opened and 2 closed in 30 days, across 50 employers
  • Evaluation and safety: 40 open, 8 opened and 2 closed in 30 days, across 17 employers

Openings outnumber closings in every family over the past 30 days. The largest concentrations of open roles are at firms operationalizing AI inside enterprises and platforms:

  • Accenture: 603 open, 206 opened in 30 days
  • Amgen: 176 open, 16 opened in 30 days
  • Capital One: 176 open, 28 opened in 30 days
  • OpenAI: 148 open, 62 opened in 30 days
  • Anthropic: 113 open, 32 opened in 30 days
  • Waymo: 101 open, 15 opened in 30 days
  • PwC: 93 open, 17 opened in 30 days
  • Databricks: 88 open, 19 opened in 30 days

The signal is clear: employers are adding engineering, data and infrastructure capacity far more than they are adding evaluation and safety roles.

Claims of pauses and slowdowns vs. ongoing expansion

Mowshowitz’s read is that “OpenAI is attempting to turn its ship around” after recent incidents, with “pauses to development while new safeguards are put in place and problems are diagnosed.” He also notes that “Anthropic revenue continues to climb as they prepare for their IPO, although growth has slowed somewhat recently,” and that “we still await the post-mortem of the HuggingFace attack.”

Our hiring data does not show a broad freeze at either lab. OpenAI lists 148 open roles and 62 opened in the past 30 days. Anthropic lists 113 open and 32 opened in the past 30 days. Without past baselines we cannot confirm Mowshowitz’s claim of slowed growth in Anthropic’s revenue, but their current openings do not suggest a hiring standstill. Across the market, openings exceeding closings in every category indicates continued expansion, not a pause.

Safety talk, tiny safety hiring

Marcus argues the industry’s finances have a “circular” feel and writes, “I am not saying that the generative AI industry is literally check-kiting.” He adds: “And it’s not clear that all the debts that are being issued can actually be paid off.” In a separate post he warns, “OpenAI is becoming a surveillance company,” adding: “I STRONGLY recommend you not consent. I certainly won’t.”

If firms were shifting aggressively into safety and oversight in response to these risks, we would expect to see that reflected in postings. Instead, evaluation and safety account for 40 open roles, with 8 opened and 2 closed in 30 days, across just 17 employers. That is the smallest category by a wide margin. The hiring picture today does not support the idea of a market pivot toward safety headcount. The emphasis, as measured by open roles, remains on building and deploying systems.

Watermarks are “free,” and the job market seems to agree

In a separate essay, Mowshowitz argues that the technical foundations for watermarking AI text are available and cheap to apply: “Scott Aaronson, while working at OpenAI, largely solved AI text watermarking together with Hendrik Kirchner.” He describes the method and claims, “This has no practical impact on outputs. Humans cannot tell the difference, at all.” He emphasizes the cost: “The marginal cost of doing this is very close to zero.” And he notes the policy and product context: “The European Union Code of Practice, signed by the major Western AI labs, requires future AI models to use such watermarks.” Plus, “Google implemented this, including for Gemini 3.7 Flash, and they have been rolling out this feature since 2024.”

Our numbers line up with that framing. If the watermarking burden is near-zero marginal cost, you would not expect waves of new hiring to implement it. The small footprint of evaluation and safety roles is consistent with “free and good” controls being layered into existing teams rather than driving a staffing spike.

Product tweaks and the work of making AI useful

Willison’s observation that “ChatGPT search now uses the site:operator at scale” points to product iteration aimed at more targeted retrieval. His developer notes elsewhere reinforce the ground-level reality of AI’s ecosystem. On the Bun runtime he writes, “Today saw the long awaited release of Bun 1.4, the first stable version since the infamous Rust rewrite a few months ago,” and highlights a long list of changes, including: “And it rewrites Bun from Zig to Rust.” Shipping glue code is a job unto itself: “Fresh installs of LLM stopped working the other day because the OpenAI Python library dropped its usage of httpx,” he reports, before explaining a quick pin to restore functionality.

Our tracker shows where that work is funded. Modelling and engineering has 1,619 open roles across 143 employers, data has 800 across 118, and infrastructure 388 across 90. Product and design shows 99 open across 50. The balance tilts toward building and maintaining systems that make models useful in production, which is exactly the environment in which changes like site: fanout and runtime rewrites matter.

Hype, skepticism, and what employers prioritize

Bender and Hanna are explicit about their stance toward grandiose narratives. Bender writes: “Mystery AI Hype Theater 3000, episode 83, in which @alexhanna.bsky.social and I wade through recent claims of the singularity, “rogue AI”, and possible machine consciousness. You know, the usual bullshit.” Mowshowitz’s own table of contents captures the debate with adjacent headings: “Language Models Offer Mundane Utility. The token rich get richer.” followed by “Language Models Don’t Offer Mundane Utility. Still can’t discriminate.”

Measured against postings, the center of gravity is with mundane utility. Research roles total 224 open across 49 employers, a fraction of the engineering and data openings. That does not mean foundational research is unimportant, only that the hiring signal right now is stronger for application, integration, and operations than for moonshot science.

Mollick anticipates the social paradox of adoption: “Its going to be an era of contradictions. Polls will show everyone hates AI overall but also everyone will secretly use AI all the time for lots of stuff.” Our counts align with the adoption half of that sentence. Accenture alone lists 603 open roles, with 206 opened in the past 30 days, and large adopters like Capital One (176 open) and Databricks (88 open) are active too. Those are not the footprints of a market retreating from AI.

Mollick also notes a product limitation: “Even when LLMs write well, the lack of variety in style is crippling.” He adds, “Prompting & temperature only gets you so far. Real variation is needed (and under-researched).” Employers may well aim at that problem, but the current postings do not isolate a surge in roles labeled for stylistic diversity or content variation. As with safety, any push here appears embedded within broader engineering and research teams rather than as a standalone hiring wave.

The bottom line

This week’s commentary spans caution about surveillance, skepticism of hype, practical product updates, and a call for better outputs. Our employer-sourced data shows a market that is still hiring to build: modelling and engineering, data and infrastructure dwarf safety and research by open-role counts, and openings exceed closings across every family in the past month. Where posts predict a pivot to safeguards or a broader pause, our tracker does not yet see that pivot in headcount. Where posts describe low-cost controls like text watermarking, the absence of a hiring spike in safety roles supports the claim that such controls are being absorbed by existing teams.

The signal from the job market is simple. Companies are still staffing the unglamorous work of making AI useful at scale, even as the arguments about what AI should do next get louder.

What we read

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

Polls will show everyone ha](https://bsky.app/profile/emollick.bsky.social/post/3mtms457ksk2c), I asked GPT-5.6 Sol to create the most Claude-y possible parody image , Even when LLMs write well, the lack of variety in style is crippling.

More on this

More from ANALYSIS